Posted by mfiguiere 9 hours ago
Some people seem to agree and some don't, but I think that indicates we're just down to your specific domain and usage patterns rather than the SOTA models being objectively better like they clearly used to be.
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.
I was very impressed.
Every time codex reviews claude written rust, I can't explain it, but it almost feels like codex wants to scream at whoever wrote it.
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.
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.
Yes, but... isn't the same true for Opus and all the other models too?
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.
It also compresses the context at around 100k tokens.
In case anyone is interested: https://github.com/sebastian/pi-extensions/tree/main/.pi/ext...
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.
That would leave almost no tokens for actual work
4.7 is better, but its also wildly expensive
You’re absolutely right!
Jokes apart, I did notice GLM doing these back and forth loops.
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.
Could you please share more about this
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
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 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.
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.
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.
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.
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’d prefer providers to simply deprecate stuff faster, but then that would break other people’s existing workflows.
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.
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.
We're also seeing that the people up top are using this to cull the herd.
At some point the is a need to have faith in some stable enough ground to be able to walk onto.
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.
East, West, Religion, Practice… From a Zen perspective, you're just troubling your mind with binaries and conflict.
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.
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!
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.
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.
It's a good model though, would be nice with a refresh.
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
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.
Now, given they can't satisfy current volume, they are forced to settle for just having crazy margins.
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.
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.
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.
[0] https://news.ycombinator.com/item?id=47431671 https://news.ycombinator.com/item?id=47322887
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.
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
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.
That's very different from an American SaaS model which relies of free but proprietary software for early growth
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.
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.
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.
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!
I still prefer that over US total dominance.
Let them fight it out.
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.
Then decide ...
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.
You really don't see the difference?
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.
It would be a great day for humanity if people would stopping glazing text autocomplete as revolutionary.
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 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.
Buying the most expensive circular saw doesn't get you the best woodworking, but it is the most expensive woodworking.
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?
You can't do any serious work on it without rationing your work and kneecapping your workflows, to the point where you design workflows around anthropic usage limit woodoo rather than what actually works.
Without this, I run into WEEKLY usage limits on $200 plan, working on a single codebase, one feature at a time, on just day 3.
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.
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.
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."
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.
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.
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
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.
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.
1. Keeping models closed source.
2. Jacking up pricing. A lot. Sometimes up to 100% increase.
How is that different from American?
Oh wait, it doesn't apply to those…
If you overuse LLMs or get excited about them at all, you're ngmi and a complete idiot.
I knew of all the 3.5’s and the one 3.6, but only now heard about the Plus.