Ok I find it funny that people compare models and are like, opus 4.7 is SOTA and is much better etc, but I have used glm 5.1 (I assume this comes form them training on both opus and codex) for things opus couldn't do and have seen it make better code, haven't tried the qwen max series but I have seen the local 122b model do smarter more correct things based on docs than opus so yes benchmarks are one thing but reality is what the modes actually do and you should learn and have the knowledge of the real strengths that models posses. It is a tool in the end you shouldn't be saying a hammer is better then a wrench even tho both would be able to drive a nail in a piece of wood.
The only reason I'm stuck with Claude and Chatgpt is because of their tool calling. They do have some pretty useful features like skills etc. I've tried using qwen and deepseek but they can't even output documents. How are you guys handling documents and excels with these tools? I'd love to switch tbh.
> I've tried using qwen and deepseek but they can't even output documents
What agent harness did you use? Usually, "write_file", "shell_exec" or similar is two of the first tools you add to an agent harness, after read_file/list_files. If it doesn't have those tools, unsure if you could even call it a agent harness in the first place.
Sorry for the confusion, I was actually talking about their Web based chat. Since most of my work is governance and docs, I just use their Web chats and they just refuse to output proper documents like Claude or Chatgpt do.
Aha... Well, I let Codex (Claude Code would work too) manage/troubleshoot .xlsx files too, seems to handle it just fine (it tends to un-archive them and browse the resulting XML files without issues), seen it do similar stuff for .app and .docx files too so maybe give that a try with other harnesses/models too, they might get it :)
I gave it a try a few weeks ago tbh, I'll give it another shot tho. I mainly use their Web chats since that's easier to use and previously, qwen, deepseek, kimi, all were unable to output proper docx files or use skills.
I've been using qwen-code (the software, not to be confused with Qwen Code the service or Qwen Coder the model) which is a fork of gemini-cli and the tool use with Qwen models at least has been great.
You can use both codex and Claude CLI with local models. I used codex with Gemma4 and it did pretty well. I did get one weird session where the model got confused and couldn't decide which tools actually existed in its inventory, but usually it could use tools just fine.
Qwen3-Coder produced much better rust code (that utilized rust's x86-64 vectorized extensions) a few months ago than Claude Opus or Google Gemini could. I was calling it from harnesses such as the Zed editor and trae CLI.
Codex is pretty good at Rust with x86 and arm intrinsics too, it replaced a bunch of hand written C/assembly code I was using. I will try Qwen and Kimi on this kind of task too.
The models test roughly equal on benchmarks, with generally small differences in their scores. So, it’s reasonable to choose the model based on other criteria. In my case, I’d switch to any vendor that had a decent plugin for JetBrains.
Every time I try to build something with it, the output is worse than other models I use (Gemini, Claude), it takes longer to reach an answer and plenty of times it gets stuck in a loop.
I've been running Opus and GLM side-by side for a couple weeks now, and I've been impressed with GLM. I will absolutely agree that it's slow, but if you let it cook, it can be really impressive and absolutely on the level of Opus. Keep in mind, I don't really use AI to build entire services, I'm mostly using it to make small changes or help me find bugs, so the slowness doesn't bother me. Maybe if I set it to make a whole web app and it took 2 days, that would be different.
The big kicker for GLM for me is I can use it in Pi, or whatever harness I like. Even if it was _slightly_ below Opus, and even though it's slower, I prefer it. Maybe Mythos will change everything, but who knows.
Opus is about 7 times more expensive than GLM with API pricing. And since you can only use the Opus subscription plan in CC, you're essentially locked into API pricing for Pi and any other harness.
So you're either paying $1000's for Opus in Pi, or $30/month for GLM in Pi. If the results are mostly equivalent that's an easy choice for most of us.
Perhaps I'm being extremely daft: If the API is 7 times more expensive, then why is it $1000 vs $30? Or is there a GLM subscription one can use with Pi? Certainly not available in my (arguably outdated) Pi.
I'm not the OP, but it's the latter. I'm currently using the "Lite" GLM subscription with OpenCode, for example. I'm not using it very heavily, but I haven't come close to hitting the limits, whereas I burned through my weekly limits with Claude very regularly.
Or tell pi to add support for the coding plan directly. That gave me GLM-5.1 support in no time along with support for showing the remaining quota, etc, too.
It also compresses the context at around 100k tokens.
I have used GLM 4.7, 5 and 5.1 now for about 3 month via OpenCode harness and I don't remember it every being stuck in a loop.
You have to keep it below ~100 000 token, else it gets funny in the head.
I only use it for hobby projects though. Paid 3 EUR per month, that is not longer available though :( Not sure what I will choose end of month. Maybe OpenCode Go.
IDK about GLM but GPT 5.4 Extra High has been great when I've used it in the VS Code Copilot extension, I see no actual reason Opus should consume 3x more quota than it the way it does
I tried GLM and Qwen last week for a day. And some issues it could solve, while some, on surface relatively easy, task it just could not solve after a few tries, that Opus oneshotted this morning with the same prompt. It’s a single example ofcourse, but I really wanted to give it a fair try. All it had to do was create a sortable list in Magento admin. But on the other hand, GLM did oneshot a phpstorm plugin
Benchmarking is grossly misleading. Claude’s subscription with Code would not score this high on the benchmarks because how they lobotomized agentic coding.
Maybe a bit misleading. I have used in in two places.
One Is for local opencode coding and config of stuff the other is for agent-browser use and for both it did better (opus 4.6) for the thing I was testing atm. The problem with opus at the moment I tired it was overthinking and moving itself sometimes I the wrong direction (not that qwen does overthink sometimes). However sometimes less is more - maybe turning thinking down on opus would have helped me. Some people said that it is better to turn it of entirely when you start to impmenent code as it already knows what it needs to do it doesn't need more distraction.
Another example is my ghostty config I learned from queen that is has theme support - opus would always just make the theme in the main file
I tried GLM5.1 last week after reading about it here. It was slow as molasses for routine tasks and I had to switch back to Claude.
It also ran out of 5H credit limit faster than Claude.
If you view the "thinking" traces you can see why; it will go back and forth on potential solutions, writing full implementations in the thinking block then debating them, constantly circling back to points it raised earlier, and starting every other paragraph with "Actually…" or "But wait!"
I was watching Qwen3.6-35B-A3B (locally) doing the same dance yesterday. It eventually finished and had a reasonable answer, but it sure went back and forth on a bunch of things I had explicitly said not to do before coming to a conclusion. At least said conclusion was not any of the things I'd said not to do.
That is essentially what the reasoning reinforcement training does. It is getting the model to say things that are more likely to result in the correct final answer. Everything it does in between doesn't necessarily need to be valid argument to produce the answer. You can think of it as filling the context with whatever is needed to make the right answer come out next. Valid arguments obviously help. but so might expressions of incorrect things that are not obviously untrue to the model until it sees them written out. The What's The Magic Word paper shows how far that could go. If the policy model managed to learn enough magic words it would be theoretically possible to end up with an LLM that spouts utter gibberish until delivering the correct answer seemingly out of the blue.
That's pretty cool, thanks for the extra context! (pardon the... not even pun I guess)
Also, thanks for pointing me at that specific paper; I spend a lot more of my life closer to classical control theory than ML theory so it's always neat to see the intersection of them. My unsubstantiated hypothesis is that controls & ML are going to start getting looked at more holistically, and not in the way I normal see it (which is "why worry about classical control theory, just solve the problem with RL"). Control theory is largely about steering dynamic systems along stable trajectories through state space... which is largely what iterative "fill in the next word" LLM models are doing. The intersection, I hope, will be interesting and add significant efficiency.
> I've used Claude for many months now. Since February I see a stark decline in the work I do with it.
I find myself repeating the following pattern: I use an AI model to assist me with work, and after some time, I notice the quality doesn't justify the time investment. I decide to try a similar task with another provider. I try a few more tests, then decide to switch over for full time work, and it feels like it's awesome and doing a good job. A few months later, it feels like the model got worse.
I wonder about this. I see two obvious possibilities (if we ignore bias):
1. The models are purposefully nerfed, before the release of the next model, similar to how Apple allegedly nerfed their older phones when the next model was out.
2. You are relying more and more on the models and are using your talent less and less. What you are observing is the ratio of your vs. the model’s work leaning more and more to the model’s. When a new model is released, it produces better quality code then before, so the work improves with it, but your talent keeps deteriorating at a constant rate.
I definitely find your last point is true for me. The more work I am doing with AI the more I am expecting it to do, similar to how you can expect more over time from a junior you are delegating to and training. However the model isn't learning or improving the same way, so your trust is quickly broken.
As you note, the developer's input is still driving the model quite a bit so if the developer is contributing less and less as they trust more, the results would get worse.
> However the model isn't learning or improving the same way, so your trust is quickly broken.
One other failure mode that I've seen in my own work while I've been learning: the things that you put into AGENTS.md/CLAUDE.md/local "memories" can improve performance or degrade performance, depending on the instructions. And unless you're actively quantitatively reviewing and considering when performance is improving or degrading, you probably won't pick up that two sentences that you added to CLAUDE.md two weeks ago are why things seem to have suddenly gotten worse.
> similar to how you can expect more over time from a junior you are delegating to and training
That's the really interesting bit. Both Claude and Codex have learned some of my preferences by me explicitly saying things like "Do not use emojis to indicate task completion in our plan files, stick to ASCII text only". But when you accidentally "teach" them something that has a negative impact on performance, they're not very likely to push back, unlike a junior engineer who will either ignore your dumb instruction or hopefully bring it up.
> As you note, the developer's input is still driving the model quite a bit so if the developer is contributing less and less as they trust more, the results would get worse.
That is definitely a thing too. There have been a few times that I have "let my guard down" so to speak and haven't deeply considered the implications of every commit. Usually this hasn't been a big deal, but there have been a few really ugly architectural decisions that have made it through the gate and had to get cleaned up later. It's largely complacency, like you point out, as well as burnout trying to keep up with reviewing and really contemplating/grokking the large volume of code output that's possible with these tools.
Your version of the last point is a bit softer I think — parent was putting it down to “loss of talent” but yours captures the gaps vs natural human interaction patterns which seems more likely, especially on such short timescales.
I confusingly say both. First I say that the ratio of work coming from the model is increasing, and when I am clarifying I say “your talent keeps deteriorating”. You correctly point out these are distinct, and maybe this distinction is important, although I personally don‘t think so. The resulting code would be the same either way.
Personally I can see the case for both interpretation to be true at the same time, and maybe that is precisely why I confused them so eagerly in my initial post.
I don’t think the providers intentionally nerf the models to make the new one look better. It’s a matter of them being stingy with infrastructure, either by choice to increase profit and/or sheer lack of resources to keep n+1 models deployed in parallel without deprecating older ones when a new one is released.
I’d prefer providers to simply deprecate stuff faster, but then that would break other people’s existing workflows.
Point 2 is so true, I definitely find myself spending more time reading code vs writing it. LLMs can teach you a lot, but it's never the same as actually sitting down and doing it yourself.
I think it might have to do with how models work, and fundamental limits with them (yes, they're stochastic parrots, yes they confabulate).
Newer (past two years?) models have improved "in detail" - or as pragmatic tools - but they still don't deserve the anthropomorphism we subject them to because they appear to communicate like us (and therefore appear to think and reason, like us).
But the "holes" are painted over in contemporary models - via training, system prompts and various clever (useful!) techniques.
But I think this leads us to have great difficulty spotting the weak spots in a new, or slightly different model - but as we get to know each particular tool - each model - we get better at spotting the holes on that model.
Maybe it's poorly chosen variable names. A tendency to write plausible looking, plausibly named, e2e tests that turns out to not quite test what they appear to test at first glance. Maybe there's missing locking of resources, use of transactions, in sequencial code that appear sound - but end up storing invalid data when one or several steps fail...
In happy cases current LLMs function like well-intentioned junior coders enthusiasticly delivering features and fixing bugs.
But in the other cases, they are like patholically lying sociopaths telling you anything you want to hear, just so you keep paying them money.
When you catch them lying, it feels a bit like a betrayal. But the parrot is just tapping the bell, so you'll keep feeding it peanuts.
I agree - the problem is it’s hard to see how people who say they’re using it effectively actually are using it, what they’re outputting, and making any sort of comparison on quality or maintainability or coherence.
In the same way, it’s hard to see how people who say they’re struggling are actually using it.
There’s truth somewhere in between “it’s the answer to everything” and “skill issue”. We know it’s overhyped. We know that it’s still useful to some extent, in many domains.
What is it that is dogma free? If one goes hardcore pyrrhonism, doubting that there is anything currently doubting as this statement is processed somehow, that is perfectly sound.
At some point the is a need to have faith in some stable enough ground to be able to walk onto.
All people think dogmatically. The only difference is what the ontological commitments and methaphysical foundations are. Take out God and people will fit politics, sports teams, tools, whatever in there. Its inescapable.
All people think dogmatically, but religion does not prevent people from acting dogmatically in politics, sports, etc. It just doesn't. It never did.
Under normal circumstances I'd consider this a nit and decline to pick it, but the number of evangelists out there arguing the equivalent of "cure your alcohol addiction with crystal meth!" is too damn high.
I'd encourage you to check it out for yourself. It's certainly possible to be a dogmatic Buddhist, but one of the foundational beliefs of Buddhism is that the type of dogmatic attachment you're describing is avoidable. It's not easy, but that's why you meditate.
The Western Zen? In my experience it is downgraded from being a religion to being a system of practice which relieves it of the broader Mahayana cosmology. But I would suggest the dogma is less obvious but still there, often just somewhere else, such as in its own limitations, or in a philosophical container at a higher level such as scientism.
Ah and there is the dogma -- the otherness of the enlightened.
The binaries still functionally exist. I see a lot of value in reflective practices. At the same time it seems unlikely to me that the point of existing is to not trouble your mind.
There's a saying in Zen: if you meet the buddha on the road, kill him. The point being, the very exaltation of enlightenment is an impediment.
If Buddhism can be said to have a goal, it is to reduce suffering (including your own), so troubling your own mind is indeed something it can help with. The point of existence would be something interesting to meditate on. If you discover it, let us all know!
I wonder to what degree it depends on how easy you find coding in general. I find for the early steps genAI is great to get the ball rolling, but rapidly it becomes more work to explain what it did wrong and how to fix it (and repeat until it does so) than to just fix the code myself.
The way to develop in this space seems to be to give away free stuff, get your name out there, then make everything proprietary. I hope they still continue releasing open weights. The day no one releases open weights is a sad day for humanity. Normal people won’t own their own compute if that ever happens.
I think that's an overgeneralization. We've seen all the American models be closed and proprietary from the start. Meanwhile the non-American (especially the Chinese ones) have been open since the start. In fact they often go the opposite direction. Many Chinese models started off proprietary and then were later opened up (like many of the larger Qwen models)
> We've seen all the American models be closed and proprietary from the start.
Most*.
OpenAI, contrary to popular belief, actually used to believe in open research and (more or less) open models. GPT1 and GPT2 both were model+code releases (although GPT2 was a "staged" release), GPT3 ended up API-only.
Also the Chinese models aren't following a typical American SaaS playbook which relies on free/cheap proprietary software for early growth. They are not just publishing their weights but also their code and often even publishing papers in Open Access journals to explicitly highlight what methods and advancements were made to accomplish their results
I think they're in a win-win situation. Big AI companies would love to see local computing die in favour of the cloud because they are well aware the moment an open model that can run on non ludicrous consumer hardware appears, they're screwed. In this situation Nvidia, AMD and the like would be the only ones profiting from it - even though I'm not convinced they'd prefer going back to fighting for B2C while B2B Is so much simpler for them
If you want to run AI models at scale and with reasonably quick response, there's not many alternatives to datacenter hardware. Consumer hardware is great for repurposing existing "free" compute (including gaming PCs, pro workstations etc. at the higher end) and for basic insurance against rug pulls from the big AI vendors, but increased scale will probably still bring very real benefits.
Currently, yes. But I don't find it hard to imagine that in a while we could get reasonably light open models with a level of reasoning similar to current opus, for instance. In such a scenario how many people would opt to pay for a way more expensive cloud subscription? Especially since lots of people are already not that interested in paying for frontier models nowadays where it makes sense. Unless keep on getting a constant, never ending stream of improvements we're basically bound to get to a point where unless you really need it you are ok with the basic, cheaper local alternative you don't have to pay for monthly.
I think average users are already okay with the reasoning level they'd get with current open models. But the big AI firms have pivoted their frontier models towards the enterprise: coding and research, as opposed to general chat. And scale is quite important for these uses, ordinary pro hardware is not enough.
This is really just a question of product design meeting the technology.
Today, lots of integer compute happens on local devices for some purposes, and in the cloud for others.
Same is already true for matmul, lots of FLOPS being spent locally on photo and video processing, speech to text, …
No obvious reason you wouldn’t want to specialize LLM tasks similarly, especially as long-running agents increasingly take over from chatbots as the dominant interaction architecture.
At a consistent amount of usage, datacenters are at least an order of magnitude more hardware efficient. I'm sure Nvidia and AMD would be fine fighting for B2C if it meant volume would be 10+x.
Now, given they can't satisfy current volume, they are forced to settle for just having crazy margins.
The problem with B2C is that you need to have leverage of some kind (more demanding applications, planned obsolescence, ...) in order to get people to keep on buying your product. The average consumer may simply consider themselves satisfied with their old product they already own and only replace it when it breaks down. On the contrary, with the cloud you can keep people hooked on getting the latest product whether they need it or not, and get artificial demand from datacentres and such.
I think businesses running datacenters are much less likely to frivolously buy the latest GPUs with no functional incentive than general consumers are...
This is obviously a strategic move at a national level. Keep publishing competing free models to erode the moat western companies could have with their proprietary models. As long as the narrative serves China there will be no turn to proprietary models.
Always has been, it’s literally saas; the slight difference is that the lowest tier subscriptions at the frontier labs are basically free trials nowadays, too
I'm a little more optimistic than that. I suspect that the open-weight models we already have are going to be enough to support incremental development of new ones, using reasonably-accessible levels of compute.
The idea that every new foundation model needs to be pretrained from scratch, using warehouses of GPUs to crunch the same 50 terabytes of data from the same original dumps of Common Crawl and various Russian pirate sites, is hard to justify on an intuitive basis. I think the hard work has already been done. We just don't know how to leverage it properly yet.
None of that is true, at least in theory. You can trivially change layer size simply by adding extra columns initialized as 0, effectively embedding your smaller network in a larger network. You can add layers in a similar way, and in fact LLMs are surprisingly robust to having layers added and removed - you can sometimes actually improve performance simply by duplicating some middle layers[0]. Tokenization is probably the hardest but all the layers between the first and last just encode embeddings; it's probably not impossible to retrain those while preserving the middle parts.
You took a simple path, embedding smaller into larger. What if you need to reduce number of layers and/or width of hidden layers? How will you embed larger into smaller? As for the "addition of same layers" - would the process of "layers to add" selection be considered training?
What if you still have to obtain the best result possible for given coefficient/tokenization budget?
I think that my comment express general case, while yours provide some exceptions.
And yet the KL divergence after changing all that stuff remains remarkably similar between different models, regardless of the specific hyperparameters and block diagrams employed at pretraining time. Some choices are better, some worse, but they all succeed at the game of next-token prediction to a similar extent.
To me, that suggests that transformer pretraining creates some underlying structure or geometry that hasn't yet been fully appreciated, and that may be more reusable than people think.
Ultimately, I also doubt that the model weights are going to turn out to be all that important. Not compared to the toolchains as a whole.
That "underappreciated underlying structure or geometry" can be just an artifact of the same tokenization used with different models.
Tokenization breaks up collocations and creates new ones that are not always present in the original text as it was. Most probably, the first byte pair found by simple byte pair encoding algorithm in enwik9 will be two spaces next to each other. Is this a true collocation? BPE thinks so. Humans may disagree.
What does concern me here is that it is very hard to ablate tokenization artifacts.
I do not think it's common crawl anymore, its common crawl++ using paid human experts to generate and verify new content, weather its code or research.
I believe US is building this off the cost difference from other countries using companies like scale, outlier etc, while china has the internal population to do this
The Chinese state wants the world using their models.
People think that Chinese AI labs are just super cool bros that love sharing for free.
The don't understand it's just a state sponsored venture meant to further entrench China in global supply and logistics. China's VCs are Chinese banks and a sprinkle of "private" money. Private in quotes because technically it still belongs to the state anyway.
China doesn't have companies and government like the US. It just has government, and a thin veil of "company" that readily fool westerners.
Also many of these Chinese companies aren't just opening their weights. They are open sourcing their code AND publishing detailed research papers alongside them to reveal how they accomplished what they accomplished.
That's very different from an American SaaS model which relies of free but proprietary software for early growth
I'm not sure how local AI models are meant to "entrench China in global supply and logistics". The two areas have nothing to do with one another. You can easily run a Chinese open model on all-American hardware.
They are building a pipeline, and the goal is to get people in the door.
If you forever stand at the entrance eating the free samples, that's fine, they don't care. Other people are going through the door and you are still consuming what they feed you. Doesn't mean it's going to be bad or evil, but they are staking their territory of control.
Oh for sure, they're getting a whole lot of Chinese people and other non-Westerners through the door already - mostly, the people who are being ignored or even blocked outright by the big Western labs. That's territory we purposely abandoned, and they're going to control it by default.
Like with nuclear technology, it's not healthy for only one country to dominate AI. The cat is already out of the bag and many countries now have the ability to train and run models. Silicon Valley has bootstrapped this space. But it should be noted that they are using AI talent from all over the world and it was sort of inevitable that this technology would get around. Lots of Chinese, Indian, Russian, and Europeans are involved.
As for what comes next, it's probably going to be a bit of a race for who can do the most useful and valuable things the cheapest. If OpenAI and Anthropic don't make it, the technology will survive them. If they do, they'll be competing on quality and cost.
As for state sponsorship, a lot of things are state sponsored. Including in the US. Silicon Valley has a rich history that is rooted in massive government funding programs. There's a great documentary out there the secret history of Silicon Valley on this. Not to mention all the "cheap" gas that is currently powering data centers of course comes on the back of a long history of public funding being channeled into the oil and gas industry.
>As for state sponsorship, a lot of things are state sponsored.
You can make any comparison you want if you use adjectives rather than values. I can say that cars use a massive amount of water (all those radiators!) to try and downplay agricultural water usage. But its blatantly disingenuous.
SV is overwhelmingly private (actual constitutional private) money. To the point that you should disregard people saying otherwise, just like you would the people saying cars use massive amounts of water.
So an OPEN model that I can run on my own fucking hardware will entrench China in global supply and logistics how?
Contrary: How will the closed, proprietary models from Anthropic, "Open"AI and Co. lead us all to freedom? Freedom of what exactly? Freedom of my money?
At some point this "anti-communism" bullshit propaganda has to stop. And that moment was decades ago!
The US examples you just gave happened decades (and in some cases hundreds) of years ago. The difference is that it's happening in China right now, and nobody cares.
i think as the pricing has gone up on the Chinese models it has made them less appealing, and with the introduction of Gemma-4 not many are at the pareto frontier (also in my experience, not just the stats): https://arena.ai/leaderboard/text/overall?viewBy=plot
The irony of this announcement is in the name: Max-Preview is proprietary, cloud-only. The Qwen models that actually matter — the ones running on real hardware people own — are the open weights series. I run the 32B and 72B variants locally on dual A4000s. The gap between those and the hosted Max is real, but it's shrinking with every release. The interesting question isn't how Max compares to Opus. It's how long until the open-weight tier makes the cloud tier irrelevant for most workloads.
Everybody's out here chasing SOTA, meanwhile I'm getting all my coding done with MiniMax M2.5 in multiple parallel sessions for $10/month and never running into limits.
For serious work, the difference between spending $10/month and $100/month is not even worth considering for most professional developers. There are exceptions like students and people in very low income countries, but I’m always confused by developers with in careers where six figure salaries are normal who are going cheap on tools.
I find even the SOTA models to be far away from trustworthy for anything beyond throwaway tasks. Supervising a less-than-SOTA model to save $10 to $100 per month is not attractive to me in the least.
I have been experimenting with self hosted models for smaller throwaway tasks a lot. It’s fun, but I’m not going to waste my time with it for the real work.
You need to supervise the model anyway, because you want that code to be long-term maintainable and defect free, and AI is nowhere near strong enough to guarantee that anytime soon. Using the latest Opus for literally everything is just a huge waste of effort.
Yes, but I find supervision much easier and faster with a strong model. It makes fewer dumb mistakes that I have to catch and correct, and it’ll follow my instructions more reliably.
$100 / month will get you rate limited to much to rely on with the Claude plans. People still report getting rate limited on the $200 / plan.
Also not everyone wants to use Claude Code, so if they're paying API pricing it's more likely thousands of dollars a month. If you can get the same results by spending a fraction of that, why wouldn't you?
For actually serious work, it's a stark difference if your proprietary and security relevant code is sent abroad to a foreign, possibly future hostile country, or is sent to some data center around the corner. It doesn't even need to be defence related.
AFAIK all these companies have SOTA or near-SOTA models available under enterprise licenses. AI companies are not interested in your secret sauce, they are trying to capture the SDLC wholesale.
I’m not sure what you are implying by “enterprise license”, but if you think it provides any meaningful protection against malicious US government actors, you really need to read and internalize the US CLOUD Act.
On a related note, I really need to try some local models (probably starting with qwen), since, at least in 2026, the Chinese models are way better at protecting democracy and free speech than the US models.
That doesn't address the concern. Google isn't interested in violating 1st and 4th amendment rights of people who criticize the government... but they do anyway (or more correctly assist the government in doing so).
If an American company, let's say a company that writes software for power stations, would use the services of a French or Chinese AI company under such enterprise licenses, how long would you think it would take until someone, in Congress e.g., would interfere?
What if they learned that half of the American small and medium sized companies would have started pouring all their business information into such a service?
While Qwen advertises large context windows, in practice the effectiveness of long-context usage seems to depend heavily on its context caching behavior. According to the official documentation, Qwen provides both implicit and explicit context caching, but these come with constraints such as short TTL (around a few minutes), prefix-based matching, and minimum token thresholds.
Because of these constraints, especially in workflows like coding agents where context grows over time, cache reuse may not scale as effectively as expected. As a result, even though the per-token price looks low, the effective cost in long sessions can feel higher due to reduced cache hit rates and repeated computation.
That said, in certain areas such as security-related tasks, I’ve personally had cases where Qwen performed better than Opus.
In my personal experience, Qwen tends to perform much better than Opus on shorter units like individual methods or functions. However, when looking at the overall coding experience, I found it works better as a function-level generator rather than as an autonomous, end-to-end coding assistant like Claude.
TBF, it's certainly best practice, advised by the model providers themselves, to cut sessions short and start new ones.
Anthropic's "Best Practices" doc[0] for Claude Code states, "A clean session with a better prompt almost always outperforms a long session with accumulated corrections."
I've been using glm5.1 for pretty much all my coding work, but Claude is too expensive for me. Haven't tried qwen yet though. China's coding models are now very cost-effective.
With them comparing to Opus 4.5, I find it hard to take some of these in good faith. Opus 4.7 is new, so I don't expect that, but Opus 4.6 has been out for quite some time.
The thing is, Opus 4.5 is where the model reached Good Enough, at least for a wide variety of problems I use LLMs for. Before that, I almost never thought it was a more productive use of my time to use AI for development tasks, because it would always hallucinate something that would waste a bunch of my time. It just wasn't a good trade.
But, if for some reason everything stopped at Opus 4.5 level and we never got a better model (and 4.6/4.7 are better, if only marginally so and mostly expanding the kind of work it can do rather than making it better at making web apps), we could still do a lot of real work real fast with Opus 4.5, and software development would never go back to everyone handwriting most of the code.
A model as good as Opus 4.5 (or slightly better according to the mostly easily gamed benchmarks) at a 10th the price is probably a worthwhile proposition for a lot of people. $100 a month, or more, to get Opus 4.7 is well worth it for a western developer...the time the lower-end models waste is far more expensive than the cost of using the most expensive models. For the foreseeable future, I'll keep paying a premium for the models that waste less of my time and produce better results with less prodding.
But, also, it's wild how fast things move. Open models you can run on relatively modest hardware are competitive with frontier models of two years ago. I mean, you can run Qwen 3.6 MoE 35B A3B or the larger Gemma 4 models on normal hardware, like a beefy Macbook or a Strix Halo or any recentish 24GB/32GB GPU...not much more expensive than the average developer laptop of pre-AI times. And, it can write code. It can write decent prose (Qwen is maybe better at code, Gemma definitely has better prose), they can use tools, they have a big enough context window for real work. They aren't as good as Opus 4.5, yet.
Anyway, I use several models at this point, for security and code reviews, even if Claude Code with Opus is still obviously the best option for most software development tasks. I'll give Qwen a try, too. I like their small models, which punch well above their weight, I'll probably like the big one, too.
If money is no object, then nothing else is worth considering if it isn't Codex 5.4/Opus 4.7/SOTA. But for many to most people, value Vs. relative quality are huge levers.
Even many people on a Claude subscription aren't choosing or able to choose Opus 4.7 because of those cost/usage pressures. Often using Sonnet or an older opus, because of the value Vs. quality curve.
Unfortunately, like with the release of Qwen3.6-Plus, this model also isn’t released for local use. From the linked article: “Qwen3.6-Max-Preview is the hosted proprietary model available via Alibaba Cloud Model Studio”
Cost may or may not be a factor in my choice of model, but knowing the capabilities and knowing they will remain consistent, reliable, and available over time is always a dominant consideration. Lately, Anthropic in particular has not been great at that.
anecdotally the quality of output isn't significantly different, the speed seems to be what you're really paying for, and since the alternative is free I'll stick to local.
When Sonnet 4.6 was released, I switchmed my default from Opus to Sonnet because it was about en par with Opus 4.5. While 4.6 and 4.7 are "better", the leap is too small for most tasks for me to need it, and so reducing cost is now a valid reason to stay at that level.
If even cheaper models start reaching that level (GLM 5.1 is also close enough that I'm using it at lot), that's a big deal, and a totally valid reason to compare against Opus 4.5
Yeah Claude Haiku (don't remember the version) did it first, they claimed it was because "it's smarter now" (it's still dumb). Then OpenAI did it with GPT-5 and Google did the same with Gemini Flash and now every new model version is at least twice as expensive than the one before that.
Considering the propaganda value in controlling the inputs to the machine that answers peoples questions, I rather expect them to be subsidized forever.
Consider the propaganda value of a centrally-controlled apparatus like the iPhone, and then reflect on the 100%+ profit margins that product has enjoyed for the past decade.
Their Plus series have existed since Qwen chat was available , as far as I remember. I can at least remember trying out their Plus model early last year.
Nowadays, I'm working on a realtime path tracer where you need proper understanding of microfacet reflection models, PDFs, (multiple) importance sampling, ReSTIR, etc.. Saying that mine is a somewhat specific use case.
And I use Claude, Gemini, GLM, Qwen to double check my math, my code and to get practical information to make my path tracer more efficient. Claude and Gemini failed me more than a couple of times with wrong, misleading and unnecessary information but on the other hand Qwen always gave me proper, practical and correct information. I’ve almost stopped using Claude and Gemini to not to waste my time anymore.
Claude code may shine developing web applications, backends and simple games but it's definitely not for me. And this is the story of my specific use case.
I have said similar things about someone experiencing similar things while writing some OpenGL code (some raytracing etc) that these models have very little understanding and aren't good at anything beyond basic CRUD web apps.
In my own experience, even with web app of medium scale (think Odoo kind of ERP), they are next to useless in understanding and modling domain correctly with very detailed written specs fed in (whole directory with index.md and sub sections and more detailed sections/chapters in separate markdown files with pointers in index.md) and I am not talking open weight models here - I am talking SOTA Claude Opus 4.6 and Gemini 3.1 Pro etc.
But that narrative isn't popular. I see the parallels here with the Crypto and NFT era. That was surely the future and at least my firm pays me in cypto whereas NFTs are used for rewarding bonusess.
a semester ago i was taking a machine learning exam in uni and the exam tasked us with creating a neural network using only numerical libraries (no pytorch ecc). I'm sure that there are a huge lot of examples looking all the same, but given that we were just students without a lot of prior experience we probably deviated from what it had in its training data, with more naive or weird solutions. Asking gemini 3 to refactor things or in very narrow things to help was ok, but it was quite bad at getting the general context, and spotting bugs, so much that a few times it was easier to grab the book and get the original formula right
otoh, we spotted a wrong formula regarding learning rate on wikipedia and it is now correct :) without gemini and just our intuition of "mhh this formula doesn't seem right", that definitely inflated our ego
What size of Qwen is that, though? The largest sizes are admittedly difficult to run locally (though this is an issue of current capability wrt. inference engines, not just raw hardware).
for Anthropic and OpenAI there is a very real danger that people invest serious time finding the strengths of alternative models, esp Chinese/open models that can to some degree be run locally as well
it puts a massive backstop at the margins they can possibly extract from users
How "social" does Quen feel? The way I am using LLMs for coding makes this actually the most important aspect by now. Claude 4.6 felt like a nice knowledgeable coworker who shared his thinking while solving problems. Claude 4.7 is the difficult anti-social guy who jumps ahead instead of actually answering your questions and does not like to talk to people in general. How are Qwen's social skills?
This is not my experience at all, Qwen3.6-Plus spits out multiple paragraphs of text for the prompts I give. It wasn't like this before. Now I have to explicitly tell it not to yap so much and keep it short, concise and direct.
Is this going to be an open weights model or not? The post doesn’t make it clear. It seems the weights are not available today, but maybe that’s because it’s in preview?
I find it odd that none of OpenAI models was used in comparison, but used Z GLM 5.1. Is Z (GLM 5.1) really that good? It is crushing Opus 4.5 in these benchmarks, if that is true, I would have expected to read many articles on HN on how people flocked CC and Codex to use it.
GLM 5.1 is pretty good, probably the best non-US agentic coding model currently available. But both GLM 5.0 and 5.1 have had issues with availability and performance that makes them frustrating to use. Recently GLM 5.1 was also outputting garbage thinking traces for me, but that appears to be fixed now.
In fact it is appreciated that Qwen is comparing to a peer. I myself and several eng I know are trying GLM. It's legit. Definitely not the same as Codex or Opus, but cheaper and "good enough". I basically ask GLM to solve a program, walk away 10-15 minutes, and the problem is solved.
cheaper is quite subjective, I just went to their pricing page [0] and cost saving compared to performance does not sell it well (again, personal opinion).
CC has a limited capacity for Opus, but fairly good for Sonnet. For Codex, never had issues about hitting my limits and I'm only a pro user.
Yes. GLM 5.1 is that good. I don't think it is as good as Claude was in January or February of this year, but it is similar to how Claude runs now, perhaps better because I feel like it's performance is more consistent.
GLM 5.1 is the first model I've found good enough to spring for a subscription for other than Claude and Codex.
It's not crushing Opus 4.5 in real-life use for me, but it's close enough to be near interchangeable with Sonnet for me for a lot of tasks, though some of the "savings" are eaten up by seemingly using more tokens for similar complexity tasks (I don't have enough data yet, but I've pushed ~500m tokens through it so far.
I'm using GLM 5.1 for the last two weeks as a cheaper alternative to Sonnet, and it's great - probably somewhere between Sonnet and Opus. It's pretty slow though.
GLM-5 is good, like really good. Especially if you take pricing into consideration. I paid 7$ for 3 months. And I get more usage than CC.
They have difficulty supplying their users with capacity, but in an email they pointed out that they are aware of it. During peak hours, I experience degraded performance. But I am on their lowest tier subscription, so I understand if my demand is not prioritized during those hours.
I've been using it through OpenCode Go and it does seem decent in my limited experience. I haven't done anything which I could directly compare to Opus yet though.
I did give it one task which was more complex and I was quite impressed by. I had a local setup with Tiltdev, K3S and a pnpm monorepo which was failing to run the web application dev server; GLM correctly figured out that it was a container image build cache issue after inspecting the containers etc and corrected the Tiltfile and build setup.
Most HN commenters seem to be a step behind the latest developments, and sometimes miss them entirely (Kimi K2.5 is one example). Not surprising as most people don't want to put in the effort to sift through the bullshit on Twitter to figure out the latest opinions. Many people here will still prefer the output of Opus 4.5/4.6/4.7, nowadays this mostly comes down to the aesthetic choices Anthropic has made.
Not just aesthetics though, from time to time I implement the same feature with CC and Codex just to compare results, and I yet to find Codex making better decisions or even the completeness of the feature.
For more complicated stuff, like queries or data comparison, Codex seems always behind for me.
its an SKU from OpenAI's perspective, broader goal and vision is (was) different. Look at the Claude and GLM, both were 95% committed to dev tooling: best coding models, coding harness, even their cowork is built on top of claude code
I'm not sure how this makes sense when Claude models aren't even coding specific: Haiku, Sonnet, Opus are the exact same models you'd use for chat or (with the recent Mythos) bleeding edge research.
But they detect it under the hood and apply a similar "variant", as API results are not the same than on Claude Code (that was documented before by someone).
I tried it asked to write it an SVG with a cat holding a guitar it wrote a pic of my gradma's look alike taking a poop. Seems alibaba has it on the spot! Lolz try it for your selves for remarkable svg's and png's!
Ok I find it funny that people compare models and are like, opus 4.7 is SOTA and is much better etc, but I have used glm 5.1 (I assume this comes form them training on both opus and codex) for things opus couldn't do and have seen it make better code, haven't tried the qwen max series but I have seen the local 122b model do smarter more correct things based on docs than opus so yes benchmarks are one thing but reality is what the modes actually do and you should learn and have the knowledge of the real strengths that models posses. It is a tool in the end you shouldn't be saying a hammer is better then a wrench even tho both would be able to drive a nail in a piece of wood.
The only reason I'm stuck with Claude and Chatgpt is because of their tool calling. They do have some pretty useful features like skills etc. I've tried using qwen and deepseek but they can't even output documents. How are you guys handling documents and excels with these tools? I'd love to switch tbh.
> I've tried using qwen and deepseek but they can't even output documents
What agent harness did you use? Usually, "write_file", "shell_exec" or similar is two of the first tools you add to an agent harness, after read_file/list_files. If it doesn't have those tools, unsure if you could even call it a agent harness in the first place.
Sorry for the confusion, I was actually talking about their Web based chat. Since most of my work is governance and docs, I just use their Web chats and they just refuse to output proper documents like Claude or Chatgpt do.
Aha... Well, I let Codex (Claude Code would work too) manage/troubleshoot .xlsx files too, seems to handle it just fine (it tends to un-archive them and browse the resulting XML files without issues), seen it do similar stuff for .app and .docx files too so maybe give that a try with other harnesses/models too, they might get it :)
You're not giving an AI command line access to your work computer? How do you expect to keep up? /s
You give it command line access in a VM...
You mean a VM like the one that contains a 0day that can escape the sandbox that gets found every year at pwn2own?
When was the last time you used Qwen models? Their 3.5 and 3.6 models are excellent with tool calling.
I gave it a try a few weeks ago tbh, I'll give it another shot tho. I mainly use their Web chats since that's easier to use and previously, qwen, deepseek, kimi, all were unable to output proper docx files or use skills.
Try loading the models up in a coding harness like Claude Code. There's a few docx skills listed on Vercel's skill index.
https://skills.sh/tfriedel/claude-office-skills/docx
You can use GLM-5.1 with claude code directly, I use ccs, GLM-5.1 setup as plan, but goes via API key.
I've been using qwen-code (the software, not to be confused with Qwen Code the service or Qwen Coder the model) which is a fork of gemini-cli and the tool use with Qwen models at least has been great.
You can use both codex and Claude CLI with local models. I used codex with Gemma4 and it did pretty well. I did get one weird session where the model got confused and couldn't decide which tools actually existed in its inventory, but usually it could use tools just fine.
Qwen3-Coder produced much better rust code (that utilized rust's x86-64 vectorized extensions) a few months ago than Claude Opus or Google Gemini could. I was calling it from harnesses such as the Zed editor and trae CLI.
I was very impressed.
Codex is pretty good at Rust with x86 and arm intrinsics too, it replaced a bunch of hand written C/assembly code I was using. I will try Qwen and Kimi on this kind of task too.
The models test roughly equal on benchmarks, with generally small differences in their scores. So, it’s reasonable to choose the model based on other criteria. In my case, I’d switch to any vendor that had a decent plugin for JetBrains.
Consider that SWE benchmarking is mainly done with python code. It tells something
I wonder why glm is viewed so positively.
Every time I try to build something with it, the output is worse than other models I use (Gemini, Claude), it takes longer to reach an answer and plenty of times it gets stuck in a loop.
I've been running Opus and GLM side-by side for a couple weeks now, and I've been impressed with GLM. I will absolutely agree that it's slow, but if you let it cook, it can be really impressive and absolutely on the level of Opus. Keep in mind, I don't really use AI to build entire services, I'm mostly using it to make small changes or help me find bugs, so the slowness doesn't bother me. Maybe if I set it to make a whole web app and it took 2 days, that would be different.
The big kicker for GLM for me is I can use it in Pi, or whatever harness I like. Even if it was _slightly_ below Opus, and even though it's slower, I prefer it. Maybe Mythos will change everything, but who knows.
> The big kicker for GLM for me is I can use it in Pi, or whatever harness I like.
Yes, but... isn't the same true for Opus and all the other models too?
Opus is about 7 times more expensive than GLM with API pricing. And since you can only use the Opus subscription plan in CC, you're essentially locked into API pricing for Pi and any other harness.
So you're either paying $1000's for Opus in Pi, or $30/month for GLM in Pi. If the results are mostly equivalent that's an easy choice for most of us.
Perhaps I'm being extremely daft: If the API is 7 times more expensive, then why is it $1000 vs $30? Or is there a GLM subscription one can use with Pi? Certainly not available in my (arguably outdated) Pi.
I'm not the OP, but it's the latter. I'm currently using the "Lite" GLM subscription with OpenCode, for example. I'm not using it very heavily, but I haven't come close to hitting the limits, whereas I burned through my weekly limits with Claude very regularly.
You can use GLM’s coding plan in Pi, just use the anthropic API instead of the OpenAI compatible one they give.
Or tell pi to add support for the coding plan directly. That gave me GLM-5.1 support in no time along with support for showing the remaining quota, etc, too.
It also compresses the context at around 100k tokens.
In case anyone is interested: https://github.com/sebastian/pi-extensions/tree/main/.pi/ext...
I have used GLM 4.7, 5 and 5.1 now for about 3 month via OpenCode harness and I don't remember it every being stuck in a loop.
You have to keep it below ~100 000 token, else it gets funny in the head.
I only use it for hobby projects though. Paid 3 EUR per month, that is not longer available though :( Not sure what I will choose end of month. Maybe OpenCode Go.
I think it offers a very good tradeoff of cost vs competency
4.7 is better, but its also wildly expensive
IDK about GLM but GPT 5.4 Extra High has been great when I've used it in the VS Code Copilot extension, I see no actual reason Opus should consume 3x more quota than it the way it does
You're probably just holding it wrong.
I tried GLM and Qwen last week for a day. And some issues it could solve, while some, on surface relatively easy, task it just could not solve after a few tries, that Opus oneshotted this morning with the same prompt. It’s a single example ofcourse, but I really wanted to give it a fair try. All it had to do was create a sortable list in Magento admin. But on the other hand, GLM did oneshot a phpstorm plugin
Do you use Opus through the API or with subscription? Did you use OpenCode or Code?
Opus trough Claude Code, the Chinese models trough OpenCode Go, which seems like a great package to test them out.
If you showed me code from GLM 5.1, Opus 4.6, and Kimi K2.6, my ranking for best model would be highly random.
Benchmarking is grossly misleading. Claude’s subscription with Code would not score this high on the benchmarks because how they lobotomized agentic coding.
>but I have seen the local 122b model do smarter more correct things based on docs than opus
Could you please share more about this
Maybe a bit misleading. I have used in in two places.
One Is for local opencode coding and config of stuff the other is for agent-browser use and for both it did better (opus 4.6) for the thing I was testing atm. The problem with opus at the moment I tired it was overthinking and moving itself sometimes I the wrong direction (not that qwen does overthink sometimes). However sometimes less is more - maybe turning thinking down on opus would have helped me. Some people said that it is better to turn it of entirely when you start to impmenent code as it already knows what it needs to do it doesn't need more distraction.
Another example is my ghostty config I learned from queen that is has theme support - opus would always just make the theme in the main file
I tried GLM5.1 last week after reading about it here. It was slow as molasses for routine tasks and I had to switch back to Claude. It also ran out of 5H credit limit faster than Claude.
If you view the "thinking" traces you can see why; it will go back and forth on potential solutions, writing full implementations in the thinking block then debating them, constantly circling back to points it raised earlier, and starting every other paragraph with "Actually…" or "But wait!"
I see this with Opus too.
Indeed. And that’s with Anthropic hiding reading traces unlike these other comparisons.
> "Actually…" or "But wait!"
You’re absolutely right!
Jokes apart, I did notice GLM doing these back and forth loops.
I was watching Qwen3.6-35B-A3B (locally) doing the same dance yesterday. It eventually finished and had a reasonable answer, but it sure went back and forth on a bunch of things I had explicitly said not to do before coming to a conclusion. At least said conclusion was not any of the things I'd said not to do.
That is essentially what the reasoning reinforcement training does. It is getting the model to say things that are more likely to result in the correct final answer. Everything it does in between doesn't necessarily need to be valid argument to produce the answer. You can think of it as filling the context with whatever is needed to make the right answer come out next. Valid arguments obviously help. but so might expressions of incorrect things that are not obviously untrue to the model until it sees them written out. The What's The Magic Word paper shows how far that could go. If the policy model managed to learn enough magic words it would be theoretically possible to end up with an LLM that spouts utter gibberish until delivering the correct answer seemingly out of the blue.
That's pretty cool, thanks for the extra context! (pardon the... not even pun I guess)
Also, thanks for pointing me at that specific paper; I spend a lot more of my life closer to classical control theory than ML theory so it's always neat to see the intersection of them. My unsubstantiated hypothesis is that controls & ML are going to start getting looked at more holistically, and not in the way I normal see it (which is "why worry about classical control theory, just solve the problem with RL"). Control theory is largely about steering dynamic systems along stable trajectories through state space... which is largely what iterative "fill in the next word" LLM models are doing. The intersection, I hope, will be interesting and add significant efficiency.
Z.ai’s cloud offering is poor, try it with a different provider.
Many people averted religion (which I can get behind with), but have never removed the dogmatic thinking that lay at its root.
As so many things these days: It's a cult.
I've used Claude for many months now. Since February I see a stark decline in the work I do with it.
I've also tried to use it for GPU programming where it absolutely sucks at, with Sonnet, Opus 4.5 and 4.6
But if you share that sentiment, it's always a "You're just holding it wrong" or "The next model will surely solve this"
For me it's just a tool, so I shrug.
> I've used Claude for many months now. Since February I see a stark decline in the work I do with it.
I find myself repeating the following pattern: I use an AI model to assist me with work, and after some time, I notice the quality doesn't justify the time investment. I decide to try a similar task with another provider. I try a few more tests, then decide to switch over for full time work, and it feels like it's awesome and doing a good job. A few months later, it feels like the model got worse.
I wonder about this. I see two obvious possibilities (if we ignore bias):
1. The models are purposefully nerfed, before the release of the next model, similar to how Apple allegedly nerfed their older phones when the next model was out.
2. You are relying more and more on the models and are using your talent less and less. What you are observing is the ratio of your vs. the model’s work leaning more and more to the model’s. When a new model is released, it produces better quality code then before, so the work improves with it, but your talent keeps deteriorating at a constant rate.
I definitely find your last point is true for me. The more work I am doing with AI the more I am expecting it to do, similar to how you can expect more over time from a junior you are delegating to and training. However the model isn't learning or improving the same way, so your trust is quickly broken.
As you note, the developer's input is still driving the model quite a bit so if the developer is contributing less and less as they trust more, the results would get worse.
> However the model isn't learning or improving the same way, so your trust is quickly broken.
One other failure mode that I've seen in my own work while I've been learning: the things that you put into AGENTS.md/CLAUDE.md/local "memories" can improve performance or degrade performance, depending on the instructions. And unless you're actively quantitatively reviewing and considering when performance is improving or degrading, you probably won't pick up that two sentences that you added to CLAUDE.md two weeks ago are why things seem to have suddenly gotten worse.
> similar to how you can expect more over time from a junior you are delegating to and training
That's the really interesting bit. Both Claude and Codex have learned some of my preferences by me explicitly saying things like "Do not use emojis to indicate task completion in our plan files, stick to ASCII text only". But when you accidentally "teach" them something that has a negative impact on performance, they're not very likely to push back, unlike a junior engineer who will either ignore your dumb instruction or hopefully bring it up.
> As you note, the developer's input is still driving the model quite a bit so if the developer is contributing less and less as they trust more, the results would get worse.
That is definitely a thing too. There have been a few times that I have "let my guard down" so to speak and haven't deeply considered the implications of every commit. Usually this hasn't been a big deal, but there have been a few really ugly architectural decisions that have made it through the gate and had to get cleaned up later. It's largely complacency, like you point out, as well as burnout trying to keep up with reviewing and really contemplating/grokking the large volume of code output that's possible with these tools.
Your version of the last point is a bit softer I think — parent was putting it down to “loss of talent” but yours captures the gaps vs natural human interaction patterns which seems more likely, especially on such short timescales.
I confusingly say both. First I say that the ratio of work coming from the model is increasing, and when I am clarifying I say “your talent keeps deteriorating”. You correctly point out these are distinct, and maybe this distinction is important, although I personally don‘t think so. The resulting code would be the same either way.
Personally I can see the case for both interpretation to be true at the same time, and maybe that is precisely why I confused them so eagerly in my initial post.
I don’t think the providers intentionally nerf the models to make the new one look better. It’s a matter of them being stingy with infrastructure, either by choice to increase profit and/or sheer lack of resources to keep n+1 models deployed in parallel without deprecating older ones when a new one is released.
I’d prefer providers to simply deprecate stuff faster, but then that would break other people’s existing workflows.
Point 2 is so true, I definitely find myself spending more time reading code vs writing it. LLMs can teach you a lot, but it's never the same as actually sitting down and doing it yourself.
I think it might have to do with how models work, and fundamental limits with them (yes, they're stochastic parrots, yes they confabulate).
Newer (past two years?) models have improved "in detail" - or as pragmatic tools - but they still don't deserve the anthropomorphism we subject them to because they appear to communicate like us (and therefore appear to think and reason, like us).
But the "holes" are painted over in contemporary models - via training, system prompts and various clever (useful!) techniques.
But I think this leads us to have great difficulty spotting the weak spots in a new, or slightly different model - but as we get to know each particular tool - each model - we get better at spotting the holes on that model.
Maybe it's poorly chosen variable names. A tendency to write plausible looking, plausibly named, e2e tests that turns out to not quite test what they appear to test at first glance. Maybe there's missing locking of resources, use of transactions, in sequencial code that appear sound - but end up storing invalid data when one or several steps fail...
In happy cases current LLMs function like well-intentioned junior coders enthusiasticly delivering features and fixing bugs.
But in the other cases, they are like patholically lying sociopaths telling you anything you want to hear, just so you keep paying them money.
When you catch them lying, it feels a bit like a betrayal. But the parrot is just tapping the bell, so you'll keep feeding it peanuts.
I agree - the problem is it’s hard to see how people who say they’re using it effectively actually are using it, what they’re outputting, and making any sort of comparison on quality or maintainability or coherence.
In the same way, it’s hard to see how people who say they’re struggling are actually using it.
There’s truth somewhere in between “it’s the answer to everything” and “skill issue”. We know it’s overhyped. We know that it’s still useful to some extent, in many domains.
Well summarized.
We're also seeing that the people up top are using this to cull the herd.
What is it that is dogma free? If one goes hardcore pyrrhonism, doubting that there is anything currently doubting as this statement is processed somehow, that is perfectly sound.
At some point the is a need to have faith in some stable enough ground to be able to walk onto.
Who controls that need for you?
All people think dogmatically. The only difference is what the ontological commitments and methaphysical foundations are. Take out God and people will fit politics, sports teams, tools, whatever in there. Its inescapable.
All people think dogmatically, but religion does not prevent people from acting dogmatically in politics, sports, etc. It just doesn't. It never did.
Under normal circumstances I'd consider this a nit and decline to pick it, but the number of evangelists out there arguing the equivalent of "cure your alcohol addiction with crystal meth!" is too damn high.
Allow me to introduce you to Buddhism
Elaborate. Buddhism is going to have the same epistemological issues as anything, since its a human consciousness issue.
> since its a human consciousness issue
I'd encourage you to check it out for yourself. It's certainly possible to be a dogmatic Buddhist, but one of the foundational beliefs of Buddhism is that the type of dogmatic attachment you're describing is avoidable. It's not easy, but that's why you meditate.
https://en.wikipedia.org/wiki/Prat%C4%ABtyasamutp%C4%81da
Which one?
Zen
The Western Zen? In my experience it is downgraded from being a religion to being a system of practice which relieves it of the broader Mahayana cosmology. But I would suggest the dogma is less obvious but still there, often just somewhere else, such as in its own limitations, or in a philosophical container at a higher level such as scientism.
All Zen is about releasing those attachments. Granted it's pretty hard, because if you succeed, you're enlightened.
East, West, Religion, Practice… From a Zen perspective, you're just troubling your mind with binaries and conflict.
Ah and there is the dogma -- the otherness of the enlightened.
The binaries still functionally exist. I see a lot of value in reflective practices. At the same time it seems unlikely to me that the point of existing is to not trouble your mind.
There's a saying in Zen: if you meet the buddha on the road, kill him. The point being, the very exaltation of enlightenment is an impediment.
If Buddhism can be said to have a goal, it is to reduce suffering (including your own), so troubling your own mind is indeed something it can help with. The point of existence would be something interesting to meditate on. If you discover it, let us all know!
Dogmatism is a spectrum and for too many people it's on the animal side of the scale.
I wonder to what degree it depends on how easy you find coding in general. I find for the early steps genAI is great to get the ball rolling, but rapidly it becomes more work to explain what it did wrong and how to fix it (and repeat until it does so) than to just fix the code myself.
The way to develop in this space seems to be to give away free stuff, get your name out there, then make everything proprietary. I hope they still continue releasing open weights. The day no one releases open weights is a sad day for humanity. Normal people won’t own their own compute if that ever happens.
I think that's an overgeneralization. We've seen all the American models be closed and proprietary from the start. Meanwhile the non-American (especially the Chinese ones) have been open since the start. In fact they often go the opposite direction. Many Chinese models started off proprietary and then were later opened up (like many of the larger Qwen models)
> We've seen all the American models be closed and proprietary from the start
What about Gemma and Llama and gpt-oss, not to mention lots of smaller/specialized models from Nvidia and others?
I would never argue that China isn't ahead in the open weights game, of course, but it's not like it's "all" American models by any stretch.
gpt-oss is good but I haven't heard anything about an update. It seems like one and done, to shut up people complaining about non-Open AI
> We've seen all the American models be closed and proprietary from the start.
Most*.
OpenAI, contrary to popular belief, actually used to believe in open research and (more or less) open models. GPT1 and GPT2 both were model+code releases (although GPT2 was a "staged" release), GPT3 ended up API-only.
OpenAI has released their GPT-OSS series more recently.
Recently, more like 20 years ago in LLM-years.
It's a good model though, would be nice with a refresh.
That's fair but those days seem so long gone now.
Also the Chinese models aren't following a typical American SaaS playbook which relies on free/cheap proprietary software for early growth. They are not just publishing their weights but also their code and often even publishing papers in Open Access journals to explicitly highlight what methods and advancements were made to accomplish their results
The Nvidia Nemotron models are recent, and of course the Gemma 4 series from Google.
Any idea why they do that?
gasp Science!
I think it is in the interest of chip makers to make sure we all get local models
I think they're in a win-win situation. Big AI companies would love to see local computing die in favour of the cloud because they are well aware the moment an open model that can run on non ludicrous consumer hardware appears, they're screwed. In this situation Nvidia, AMD and the like would be the only ones profiting from it - even though I'm not convinced they'd prefer going back to fighting for B2C while B2B Is so much simpler for them
If you want to run AI models at scale and with reasonably quick response, there's not many alternatives to datacenter hardware. Consumer hardware is great for repurposing existing "free" compute (including gaming PCs, pro workstations etc. at the higher end) and for basic insurance against rug pulls from the big AI vendors, but increased scale will probably still bring very real benefits.
Currently, yes. But I don't find it hard to imagine that in a while we could get reasonably light open models with a level of reasoning similar to current opus, for instance. In such a scenario how many people would opt to pay for a way more expensive cloud subscription? Especially since lots of people are already not that interested in paying for frontier models nowadays where it makes sense. Unless keep on getting a constant, never ending stream of improvements we're basically bound to get to a point where unless you really need it you are ok with the basic, cheaper local alternative you don't have to pay for monthly.
I think average users are already okay with the reasoning level they'd get with current open models. But the big AI firms have pivoted their frontier models towards the enterprise: coding and research, as opposed to general chat. And scale is quite important for these uses, ordinary pro hardware is not enough.
This is really just a question of product design meeting the technology.
Today, lots of integer compute happens on local devices for some purposes, and in the cloud for others.
Same is already true for matmul, lots of FLOPS being spent locally on photo and video processing, speech to text, …
No obvious reason you wouldn’t want to specialize LLM tasks similarly, especially as long-running agents increasingly take over from chatbots as the dominant interaction architecture.
At a consistent amount of usage, datacenters are at least an order of magnitude more hardware efficient. I'm sure Nvidia and AMD would be fine fighting for B2C if it meant volume would be 10+x.
Now, given they can't satisfy current volume, they are forced to settle for just having crazy margins.
The problem with B2C is that you need to have leverage of some kind (more demanding applications, planned obsolescence, ...) in order to get people to keep on buying your product. The average consumer may simply consider themselves satisfied with their old product they already own and only replace it when it breaks down. On the contrary, with the cloud you can keep people hooked on getting the latest product whether they need it or not, and get artificial demand from datacentres and such.
I think businesses running datacenters are much less likely to frivolously buy the latest GPUs with no functional incentive than general consumers are...
Definitely. Many big hardware firms are directly supporting HuggingFace for this very reason.
True, chip companies have the opposite mindset, Nvidia is making their own open weights I believe
This is obviously a strategic move at a national level. Keep publishing competing free models to erode the moat western companies could have with their proprietary models. As long as the narrative serves China there will be no turn to proprietary models.
Always has been, it’s literally saas; the slight difference is that the lowest tier subscriptions at the frontier labs are basically free trials nowadays, too
Its the new freeware model!
I'm a little more optimistic than that. I suspect that the open-weight models we already have are going to be enough to support incremental development of new ones, using reasonably-accessible levels of compute.
The idea that every new foundation model needs to be pretrained from scratch, using warehouses of GPUs to crunch the same 50 terabytes of data from the same original dumps of Common Crawl and various Russian pirate sites, is hard to justify on an intuitive basis. I think the hard work has already been done. We just don't know how to leverage it properly yet.
Change layer size and you have to retrain. Change number of layers and you have to retrain. Change tokenization and you have to retrain.
None of that is true, at least in theory. You can trivially change layer size simply by adding extra columns initialized as 0, effectively embedding your smaller network in a larger network. You can add layers in a similar way, and in fact LLMs are surprisingly robust to having layers added and removed - you can sometimes actually improve performance simply by duplicating some middle layers[0]. Tokenization is probably the hardest but all the layers between the first and last just encode embeddings; it's probably not impossible to retrain those while preserving the middle parts.
[0] https://news.ycombinator.com/item?id=47431671 https://news.ycombinator.com/item?id=47322887
You took a simple path, embedding smaller into larger. What if you need to reduce number of layers and/or width of hidden layers? How will you embed larger into smaller? As for the "addition of same layers" - would the process of "layers to add" selection be considered training?
What if you still have to obtain the best result possible for given coefficient/tokenization budget?
I think that my comment express general case, while yours provide some exceptions.
there is evidence it is useful in some cases, but obviously no evidence it is enough if you chase to beat SOTA.
Hopefully we will find a way to make it so that making minor changes don't require a full retrain. Training how to train, as a concept, comes to mind.
And yet the KL divergence after changing all that stuff remains remarkably similar between different models, regardless of the specific hyperparameters and block diagrams employed at pretraining time. Some choices are better, some worse, but they all succeed at the game of next-token prediction to a similar extent.
To me, that suggests that transformer pretraining creates some underlying structure or geometry that hasn't yet been fully appreciated, and that may be more reusable than people think.
Ultimately, I also doubt that the model weights are going to turn out to be all that important. Not compared to the toolchains as a whole.
That "underappreciated underlying structure or geometry" can be just an artifact of the same tokenization used with different models.
Tokenization breaks up collocations and creates new ones that are not always present in the original text as it was. Most probably, the first byte pair found by simple byte pair encoding algorithm in enwik9 will be two spaces next to each other. Is this a true collocation? BPE thinks so. Humans may disagree.
What does concern me here is that it is very hard to ablate tokenization artifacts.
I do not think it's common crawl anymore, its common crawl++ using paid human experts to generate and verify new content, weather its code or research.
I believe US is building this off the cost difference from other countries using companies like scale, outlier etc, while china has the internal population to do this
Any reason for them to do this other than altruism? I don’t think this can be regulated.
Bake ads into them.
The Chinese state wants the world using their models.
People think that Chinese AI labs are just super cool bros that love sharing for free.
The don't understand it's just a state sponsored venture meant to further entrench China in global supply and logistics. China's VCs are Chinese banks and a sprinkle of "private" money. Private in quotes because technically it still belongs to the state anyway.
China doesn't have companies and government like the US. It just has government, and a thin veil of "company" that readily fool westerners.
As opposed to the US, which just has companies and a thin veil of “government”.
Also many of these Chinese companies aren't just opening their weights. They are open sourcing their code AND publishing detailed research papers alongside them to reveal how they accomplished what they accomplished.
That's very different from an American SaaS model which relies of free but proprietary software for early growth
I'm not sure how local AI models are meant to "entrench China in global supply and logistics". The two areas have nothing to do with one another. You can easily run a Chinese open model on all-American hardware.
They are building a pipeline, and the goal is to get people in the door.
If you forever stand at the entrance eating the free samples, that's fine, they don't care. Other people are going through the door and you are still consuming what they feed you. Doesn't mean it's going to be bad or evil, but they are staking their territory of control.
Oh for sure, they're getting a whole lot of Chinese people and other non-Westerners through the door already - mostly, the people who are being ignored or even blocked outright by the big Western labs. That's territory we purposely abandoned, and they're going to control it by default.
Like with nuclear technology, it's not healthy for only one country to dominate AI. The cat is already out of the bag and many countries now have the ability to train and run models. Silicon Valley has bootstrapped this space. But it should be noted that they are using AI talent from all over the world and it was sort of inevitable that this technology would get around. Lots of Chinese, Indian, Russian, and Europeans are involved.
As for what comes next, it's probably going to be a bit of a race for who can do the most useful and valuable things the cheapest. If OpenAI and Anthropic don't make it, the technology will survive them. If they do, they'll be competing on quality and cost.
As for state sponsorship, a lot of things are state sponsored. Including in the US. Silicon Valley has a rich history that is rooted in massive government funding programs. There's a great documentary out there the secret history of Silicon Valley on this. Not to mention all the "cheap" gas that is currently powering data centers of course comes on the back of a long history of public funding being channeled into the oil and gas industry.
>As for state sponsorship, a lot of things are state sponsored.
You can make any comparison you want if you use adjectives rather than values. I can say that cars use a massive amount of water (all those radiators!) to try and downplay agricultural water usage. But its blatantly disingenuous.
SV is overwhelmingly private (actual constitutional private) money. To the point that you should disregard people saying otherwise, just like you would the people saying cars use massive amounts of water.
So an OPEN model that I can run on my own fucking hardware will entrench China in global supply and logistics how?
Contrary: How will the closed, proprietary models from Anthropic, "Open"AI and Co. lead us all to freedom? Freedom of what exactly? Freedom of my money?
At some point this "anti-communism" bullshit propaganda has to stop. And that moment was decades ago!
Anything that isn't explicitly to the benefit of US interests must be against them /s
So what?
I still prefer that over US total dominance.
Let them fight it out.
Yeah, a lot of people are still living within the paradigm of tribalism: my team good, other team bad.
But the events of the past decade or so have clearly demonstrated that there are no "good" actors.
I personally couldn't care less who wins in the China vs US AI competition, both sides have a long list of pros and cons.
I'd get a bit informed about what exactly Chinese dominance entails. Ask a few Uyghurs, Cantonese Hong Kongers, or even Tibetans.
Then decide ...
Ask a few Native Americans about dominance.
Or maybe families of African descent.
Or maybe families of Japanese Americans who lived in the US during WWII.
Or maybe people of Latin descent living in the US today.
The US examples you just gave happened decades (and in some cases hundreds) of years ago. The difference is that it's happening in China right now, and nobody cares.
You really don't see the difference?
The US is the biggest threat to the world right now, and is actively supporting a genocide in Palestine as well as war crimes in Lebanon.
I'm perfectly happy to let the chinese get a piece of the pie and fight the US, no matter how bad they are right now.
Well, isn't this what the US and really any other power in the world has always done, since forever?
Why is it sad? These things are useles all around, along with the people who overuse them.
It would be a great day for humanity if people would stopping glazing text autocomplete as revolutionary.
Kimi K2.6 also released today. I think it's fair to compare the two models.
Qwen appears to be much more expensive:
- Qwen: $1.3 in / $7.8 out
- Kimi: $0.95 in / $4 out
--
The announcement posts only share two overlapping benchmark results. Qwen appears to score slightly lower on SWE-Bench Pro and Terminal-Bench 2.0.
Qwen:
- Teminal-Bench 2.0: 65.4
- SWE-Bench Pro: 57.3
Kimi:
- Terminal-Bench 2.0: 66.8
- SWE-Bench Pro: 58.6
--
Different models have different strong suits, and benchmarks don't cover everything. But from a numbers perspective, Kimi looks much more appealing.
i think as the pricing has gone up on the Chinese models it has made them less appealing, and with the introduction of Gemma-4 not many are at the pareto frontier (also in my experience, not just the stats): https://arena.ai/leaderboard/text/overall?viewBy=plot
The irony of this announcement is in the name: Max-Preview is proprietary, cloud-only. The Qwen models that actually matter — the ones running on real hardware people own — are the open weights series. I run the 32B and 72B variants locally on dual A4000s. The gap between those and the hosted Max is real, but it's shrinking with every release. The interesting question isn't how Max compares to Opus. It's how long until the open-weight tier makes the cloud tier irrelevant for most workloads.
Everybody's out here chasing SOTA, meanwhile I'm getting all my coding done with MiniMax M2.5 in multiple parallel sessions for $10/month and never running into limits.
For serious work, the difference between spending $10/month and $100/month is not even worth considering for most professional developers. There are exceptions like students and people in very low income countries, but I’m always confused by developers with in careers where six figure salaries are normal who are going cheap on tools.
I find even the SOTA models to be far away from trustworthy for anything beyond throwaway tasks. Supervising a less-than-SOTA model to save $10 to $100 per month is not attractive to me in the least.
I have been experimenting with self hosted models for smaller throwaway tasks a lot. It’s fun, but I’m not going to waste my time with it for the real work.
You need to supervise the model anyway, because you want that code to be long-term maintainable and defect free, and AI is nowhere near strong enough to guarantee that anytime soon. Using the latest Opus for literally everything is just a huge waste of effort.
Yes, but I find supervision much easier and faster with a strong model. It makes fewer dumb mistakes that I have to catch and correct, and it’ll follow my instructions more reliably.
Waste of effort... of Opus? If "Opus effort" is cheaper, than dev hours managing yourself more dumb/effective model, what is the point?
rich people dont concern themselves with the cost of tokens.
It is not even rich. If you earn more than $30k it is worth your employer spending $3k on AI tools.
$100 / month will get you rate limited to much to rely on with the Claude plans. People still report getting rate limited on the $200 / plan.
Also not everyone wants to use Claude Code, so if they're paying API pricing it's more likely thousands of dollars a month. If you can get the same results by spending a fraction of that, why wouldn't you?
For actually serious work, it's a stark difference if your proprietary and security relevant code is sent abroad to a foreign, possibly future hostile country, or is sent to some data center around the corner. It doesn't even need to be defence related.
AFAIK all these companies have SOTA or near-SOTA models available under enterprise licenses. AI companies are not interested in your secret sauce, they are trying to capture the SDLC wholesale.
I’m not sure what you are implying by “enterprise license”, but if you think it provides any meaningful protection against malicious US government actors, you really need to read and internalize the US CLOUD Act.
On a related note, I really need to try some local models (probably starting with qwen), since, at least in 2026, the Chinese models are way better at protecting democracy and free speech than the US models.
That doesn't address the concern. Google isn't interested in violating 1st and 4th amendment rights of people who criticize the government... but they do anyway (or more correctly assist the government in doing so).
If an American company, let's say a company that writes software for power stations, would use the services of a French or Chinese AI company under such enterprise licenses, how long would you think it would take until someone, in Congress e.g., would interfere?
What if they learned that half of the American small and medium sized companies would have started pouring all their business information into such a service?
Who are you paying $10/month? OpenRouter?
https://platform.minimax.io/docs/guides/pricing-token-plan
How do you use this? Do you use opencode or another frontend?
https://www.alibabacloud.com/help/en/model-studio/context-ca... I’ve also been testing models like Opus, Codex, and Qwen, and Qwen is strong in many coding tasks. However, my main concern is how it behaves in long-running sessions.
While Qwen advertises large context windows, in practice the effectiveness of long-context usage seems to depend heavily on its context caching behavior. According to the official documentation, Qwen provides both implicit and explicit context caching, but these come with constraints such as short TTL (around a few minutes), prefix-based matching, and minimum token thresholds.
Because of these constraints, especially in workflows like coding agents where context grows over time, cache reuse may not scale as effectively as expected. As a result, even though the per-token price looks low, the effective cost in long sessions can feel higher due to reduced cache hit rates and repeated computation.
That said, in certain areas such as security-related tasks, I’ve personally had cases where Qwen performed better than Opus.
In my personal experience, Qwen tends to perform much better than Opus on shorter units like individual methods or functions. However, when looking at the overall coding experience, I found it works better as a function-level generator rather than as an autonomous, end-to-end coding assistant like Claude.
TBF, it's certainly best practice, advised by the model providers themselves, to cut sessions short and start new ones.
Anthropic's "Best Practices" doc[0] for Claude Code states, "A clean session with a better prompt almost always outperforms a long session with accumulated corrections."
[0] https://code.claude.com/docs/en/best-practices
Unless stuff changed since I last checked, context caching just reduces cost / latency. It does not change what tokens are emitted.
I've been using glm5.1 for pretty much all my coding work, but Claude is too expensive for me. Haven't tried qwen yet though. China's coding models are now very cost-effective.
With them comparing to Opus 4.5, I find it hard to take some of these in good faith. Opus 4.7 is new, so I don't expect that, but Opus 4.6 has been out for quite some time.
The thing is, Opus 4.5 is where the model reached Good Enough, at least for a wide variety of problems I use LLMs for. Before that, I almost never thought it was a more productive use of my time to use AI for development tasks, because it would always hallucinate something that would waste a bunch of my time. It just wasn't a good trade.
But, if for some reason everything stopped at Opus 4.5 level and we never got a better model (and 4.6/4.7 are better, if only marginally so and mostly expanding the kind of work it can do rather than making it better at making web apps), we could still do a lot of real work real fast with Opus 4.5, and software development would never go back to everyone handwriting most of the code.
A model as good as Opus 4.5 (or slightly better according to the mostly easily gamed benchmarks) at a 10th the price is probably a worthwhile proposition for a lot of people. $100 a month, or more, to get Opus 4.7 is well worth it for a western developer...the time the lower-end models waste is far more expensive than the cost of using the most expensive models. For the foreseeable future, I'll keep paying a premium for the models that waste less of my time and produce better results with less prodding.
But, also, it's wild how fast things move. Open models you can run on relatively modest hardware are competitive with frontier models of two years ago. I mean, you can run Qwen 3.6 MoE 35B A3B or the larger Gemma 4 models on normal hardware, like a beefy Macbook or a Strix Halo or any recentish 24GB/32GB GPU...not much more expensive than the average developer laptop of pre-AI times. And, it can write code. It can write decent prose (Qwen is maybe better at code, Gemma definitely has better prose), they can use tools, they have a big enough context window for real work. They aren't as good as Opus 4.5, yet.
Anyway, I use several models at this point, for security and code reviews, even if Claude Code with Opus is still obviously the best option for most software development tasks. I'll give Qwen a try, too. I like their small models, which punch well above their weight, I'll probably like the big one, too.
If money is no object, then nothing else is worth considering if it isn't Codex 5.4/Opus 4.7/SOTA. But for many to most people, value Vs. relative quality are huge levers.
Even many people on a Claude subscription aren't choosing or able to choose Opus 4.7 because of those cost/usage pressures. Often using Sonnet or an older opus, because of the value Vs. quality curve.
Also us weirdos with local model uses. But your point stands.
Unfortunately, like with the release of Qwen3.6-Plus, this model also isn’t released for local use. From the linked article: “Qwen3.6-Max-Preview is the hosted proprietary model available via Alibaba Cloud Model Studio”
The Max series was never available for local use, though. So this is expected.
Sure, not plus or max. I just use their lesser moe ones locally (that would never come close to massive sota models) all the time.
Cost may or may not be a factor in my choice of model, but knowing the capabilities and knowing they will remain consistent, reliable, and available over time is always a dominant consideration. Lately, Anthropic in particular has not been great at that.
anecdotally the quality of output isn't significantly different, the speed seems to be what you're really paying for, and since the alternative is free I'll stick to local.
What are the best models to run locally?
Codex 5.4 is not out?
Codex subscription is very generous at pro tiers
Opus 4.6 performance has been so wildly inconsistent over the past couple of months, why waste the tokens?
When Sonnet 4.6 was released, I switchmed my default from Opus to Sonnet because it was about en par with Opus 4.5. While 4.6 and 4.7 are "better", the leap is too small for most tasks for me to need it, and so reducing cost is now a valid reason to stay at that level.
If even cheaper models start reaching that level (GLM 5.1 is also close enough that I'm using it at lot), that's a big deal, and a totally valid reason to compare against Opus 4.5
Wow I couldn't disagree more.
For me, Opus 4.5 and 4.6 feel so different compared to sonnet.
Maybe I'm lazy or something but sonnet is much worse in my experience at inferring intent correctly if I've left any ambiguity.
That effect is super compounding.
You compare with what's most comparable.
In any case a benchmark provided by the provider is always biased, they will pick the frameworks where their model fares well. Omit the others.
Independent benchmarks are the go to.
Opus 4.6 was released in February. It can take quite some time to run all these benchmarks properly
Quite some time is a little over 2 months. I understand this is actually true right now, but it’s still a bit hard to accept.
Comparing it with Opus 4.6 is difficult, since Anthropic may ban accounts and accuse users of state-sponsored hacking.
I think its only been like 10 weeks. I meant that's forever in AI time, but not a long time in normie people time.
Notice the pattern that Chinese providers are now:
1. Keeping models closed source.
2. Jacking up pricing. A lot. Sometimes up to 100% increase.
Huh yeah, that's truly a unique trait these Chinese companies don't share with companies in other countries.
No it is not, but they had a unique positioning around open-source and the parent commenter means that they are losing it.
> Jacking up pricing. A lot. Sometimes up to 100% increase.
How is that different from American?
Are you talking about GLM 5.1, DeepSeek V3.2 or Kimi K2.6 (released one hour ago!)?
Oh wait, it doesn't apply to those…
Z.ai's Coding Plan with GLM 5.1 (Max) did more than double in price. It was $80 two weeks ago, and now it's $160.
Coding plans are subsidised crap anyways, the real price win is the API pricing which is not.
Yet.
US companies hate that trick?!
you mean: invented
Yeah Claude Haiku (don't remember the version) did it first, they claimed it was because "it's smarter now" (it's still dumb). Then OpenAI did it with GPT-5 and Google did the same with Gemini Flash and now every new model version is at least twice as expensive than the one before that.
Well, they can't subsidize forever. And, it is kinda expected?
Considering the propaganda value in controlling the inputs to the machine that answers peoples questions, I rather expect them to be subsidized forever.
Consider the propaganda value of a centrally-controlled apparatus like the iPhone, and then reflect on the 100%+ profit margins that product has enjoyed for the past decade.
what only Oracle can do it?
Yeah, its almost like the casinos started rigging the game after they got all the addicts hooked. Who saw that coming???
If you overuse LLMs or get excited about them at all, you're ngmi and a complete idiot.
The fun thing is, you can be aware of the entire range of Qwen models that are available for local running, but not at all about their cloud models.
I knew of all the 3.5’s and the one 3.6, but only now heard about the Plus.
Their Plus series have existed since Qwen chat was available , as far as I remember. I can at least remember trying out their Plus model early last year.
Nowadays, I'm working on a realtime path tracer where you need proper understanding of microfacet reflection models, PDFs, (multiple) importance sampling, ReSTIR, etc.. Saying that mine is a somewhat specific use case.
And I use Claude, Gemini, GLM, Qwen to double check my math, my code and to get practical information to make my path tracer more efficient. Claude and Gemini failed me more than a couple of times with wrong, misleading and unnecessary information but on the other hand Qwen always gave me proper, practical and correct information. I’ve almost stopped using Claude and Gemini to not to waste my time anymore.
Claude code may shine developing web applications, backends and simple games but it's definitely not for me. And this is the story of my specific use case.
I have said similar things about someone experiencing similar things while writing some OpenGL code (some raytracing etc) that these models have very little understanding and aren't good at anything beyond basic CRUD web apps.
In my own experience, even with web app of medium scale (think Odoo kind of ERP), they are next to useless in understanding and modling domain correctly with very detailed written specs fed in (whole directory with index.md and sub sections and more detailed sections/chapters in separate markdown files with pointers in index.md) and I am not talking open weight models here - I am talking SOTA Claude Opus 4.6 and Gemini 3.1 Pro etc.
But that narrative isn't popular. I see the parallels here with the Crypto and NFT era. That was surely the future and at least my firm pays me in cypto whereas NFTs are used for rewarding bonusess.
Someone exactly said it better here[0] already.
[0]. https://news.ycombinator.com/item?id=47817982
a semester ago i was taking a machine learning exam in uni and the exam tasked us with creating a neural network using only numerical libraries (no pytorch ecc). I'm sure that there are a huge lot of examples looking all the same, but given that we were just students without a lot of prior experience we probably deviated from what it had in its training data, with more naive or weird solutions. Asking gemini 3 to refactor things or in very narrow things to help was ok, but it was quite bad at getting the general context, and spotting bugs, so much that a few times it was easier to grab the book and get the original formula right
otoh, we spotted a wrong formula regarding learning rate on wikipedia and it is now correct :) without gemini and just our intuition of "mhh this formula doesn't seem right", that definitely inflated our ego
What size of Qwen is that, though? The largest sizes are admittedly difficult to run locally (though this is an issue of current capability wrt. inference engines, not just raw hardware).
I'm directly using https://chat.qwen.ai (Qwen3.6-Plus) and planning to switch to Qwen Code with subscription.
for Anthropic and OpenAI there is a very real danger that people invest serious time finding the strengths of alternative models, esp Chinese/open models that can to some degree be run locally as well
it puts a massive backstop at the margins they can possibly extract from users
You may be interested in "radiance cascades"
What do you use instead of the Claude code client app?
How "social" does Quen feel? The way I am using LLMs for coding makes this actually the most important aspect by now. Claude 4.6 felt like a nice knowledgeable coworker who shared his thinking while solving problems. Claude 4.7 is the difficult anti-social guy who jumps ahead instead of actually answering your questions and does not like to talk to people in general. How are Qwen's social skills?
Qwen feels like wise Chinese philosopher. Talks in very short elegant sentences, but does very solid work.
> Talks in very short elegant sentences
This is not my experience at all, Qwen3.6-Plus spits out multiple paragraphs of text for the prompts I give. It wasn't like this before. Now I have to explicitly tell it not to yap so much and keep it short, concise and direct.
Is this going to be an open weights model or not? The post doesn’t make it clear. It seems the weights are not available today, but maybe that’s because it’s in preview?
The Max series has never been open.
why do people continue to benchmark their sota models against older models.
I find it odd that none of OpenAI models was used in comparison, but used Z GLM 5.1. Is Z (GLM 5.1) really that good? It is crushing Opus 4.5 in these benchmarks, if that is true, I would have expected to read many articles on HN on how people flocked CC and Codex to use it.
GLM 5.1 is pretty good, probably the best non-US agentic coding model currently available. But both GLM 5.0 and 5.1 have had issues with availability and performance that makes them frustrating to use. Recently GLM 5.1 was also outputting garbage thinking traces for me, but that appears to be fixed now.
Use them via DeepInfra instead of z.ai. No reliability issues.
https://deepinfra.com/zai-org/GLM-5.1
Looks like fp4 quantization now though? Last week was showing fp8. Hm..
Deepinfra's implementation of it is not correct. Thinking is not preserved, and they're not responding to my submitted issue about it.
I also regularly experience Deepinfra slow to an absolute crawl - I've actually gotten more consistent performance from Z.ai.
I really liked Deepinfra but something doesn't seem right over there at the moment.
Damn. Yeah, that sucks. I did play with it earlier again and it did seem to slow down.
It's frankly a bummer that there's not seemingly a better serving option for GLM 5.1 than z.AI, who seems to have reliability and cost issues.
In fact it is appreciated that Qwen is comparing to a peer. I myself and several eng I know are trying GLM. It's legit. Definitely not the same as Codex or Opus, but cheaper and "good enough". I basically ask GLM to solve a program, walk away 10-15 minutes, and the problem is solved.
cheaper is quite subjective, I just went to their pricing page [0] and cost saving compared to performance does not sell it well (again, personal opinion).
CC has a limited capacity for Opus, but fairly good for Sonnet. For Codex, never had issues about hitting my limits and I'm only a pro user.
https://z.ai/subscribe
Yes. GLM 5.1 is that good. I don't think it is as good as Claude was in January or February of this year, but it is similar to how Claude runs now, perhaps better because I feel like it's performance is more consistent.
GLM 5.1 is the first model I've found good enough to spring for a subscription for other than Claude and Codex.
It's not crushing Opus 4.5 in real-life use for me, but it's close enough to be near interchangeable with Sonnet for me for a lot of tasks, though some of the "savings" are eaten up by seemingly using more tokens for similar complexity tasks (I don't have enough data yet, but I've pushed ~500m tokens through it so far.
I'm using GLM 5.1 for the last two weeks as a cheaper alternative to Sonnet, and it's great - probably somewhere between Sonnet and Opus. It's pretty slow though.
This is what kills it for me… The long thinking blocks can make a simple task take 30 minutes.
If you only look at open models, GLM 5.1 is the best performance you can get on on the Pareto distribution
https://arena.ai/leaderboard/text?viewBy=plot&license=open-s...
GLM-5 is good, like really good. Especially if you take pricing into consideration. I paid 7$ for 3 months. And I get more usage than CC.
They have difficulty supplying their users with capacity, but in an email they pointed out that they are aware of it. During peak hours, I experience degraded performance. But I am on their lowest tier subscription, so I understand if my demand is not prioritized during those hours.
Where are you getting 3 months for $7?
They had a Christmas deal that ended January 31.
I've been using it through OpenCode Go and it does seem decent in my limited experience. I haven't done anything which I could directly compare to Opus yet though.
I did give it one task which was more complex and I was quite impressed by. I had a local setup with Tiltdev, K3S and a pnpm monorepo which was failing to run the web application dev server; GLM correctly figured out that it was a container image build cache issue after inspecting the containers etc and corrected the Tiltfile and build setup.
Most HN commenters seem to be a step behind the latest developments, and sometimes miss them entirely (Kimi K2.5 is one example). Not surprising as most people don't want to put in the effort to sift through the bullshit on Twitter to figure out the latest opinions. Many people here will still prefer the output of Opus 4.5/4.6/4.7, nowadays this mostly comes down to the aesthetic choices Anthropic has made.
Not just aesthetics though, from time to time I implement the same feature with CC and Codex just to compare results, and I yet to find Codex making better decisions or even the completeness of the feature.
For more complicated stuff, like queries or data comparison, Codex seems always behind for me.
maybe they decided OpenAI has different market, hence comparing only with companies who are focusing in dev tooling: Claude, GLM
Haven’t you heard about Codex?
its an SKU from OpenAI's perspective, broader goal and vision is (was) different. Look at the Claude and GLM, both were 95% committed to dev tooling: best coding models, coding harness, even their cowork is built on top of claude code
I'm not sure how this makes sense when Claude models aren't even coding specific: Haiku, Sonnet, Opus are the exact same models you'd use for chat or (with the recent Mythos) bleeding edge research.
Anthropic models and training data is optimized for coding use cases, this is the difference.
OpenAI on the other hand has different models optimized for coding, GPT-x-codex, Anthropic doesnt have this distinction
But they detect it under the hood and apply a similar "variant", as API results are not the same than on Claude Code (that was documented before by someone).
Yeah GLM’s great for coding, code review, and tool use. Not amazing at other domains.
I use it and think its intelligence compares favorably with OpenAI and Anthropic workhorses. Its biggest weakness is its speed.
I have the M3 Max MBP with 128 GB of memory and the 40 core GPU. What's the best local model I can run today for coding?
You can try https://github.com/AlexsJones/llmfit
I think the benchmarks and numbers need to be easier to read. Those benchmarks are useless to the regular consumer.
A bit weird to be comparing it to Opus-4.5 when 4.7 was released...
Very impressive!
I am trying since one week to subscribe Alibaba Coding Plan (to use Qwen 3.6 Plus) but it's always out of stock.
They brag about Qwen but don't let people use it.
ToKeN PrIcEs ArE gOiNg tO PluMmEt, InTelLigEnCe WiLl Be AfForDaBlE FoR EvErYOnE
I tried it asked to write it an SVG with a cat holding a guitar it wrote a pic of my gradma's look alike taking a poop. Seems alibaba has it on the spot! Lolz try it for your selves for remarkable svg's and png's!