As for Mistral - I got really excited when they said Large 4 was focusing on being #1 in cybersecurity, because that's somewhere that they genuinely could edge out Anthropic & OpenAI. Have it actually solve problems, instead of Anthropic flagging "you tried to find a null pointer exception bug in your own code, we're now reporting you to the US government". But on the Mistral benchmarks I'm seeing, this looks very disappointing, but at least they haven't entirely given up. I genuinely thought Mistral had given up on new general models. They need to learn the bitter lesson all over again.
The capabilities of all models increasing so much all the time means there are simply less and less tasks you need a frontier model for.
Even if Opus 5.5 is 500x better than Deepseek, if deepseek can solve all my problems, why do I need to pay for more?
Obviously any model will do if you use it as a better autocomplete.
I believe that there is a large gap in expectations between different workflows.
Until the AI like reads my mind and produces perfectly production ready apps with minimal intervention from my side, there is still going to be room for improvement.
Now my provider serves DeepSeek around 300 tokens a second. The code is shit but I can have a few more rounds of corrections.
DeepSeek was maybe a dollar for the full PR. Opus/Astra over 50 dollars. And double that if you use their fast variants which matches the DeepSeek speeds on certain providers.
Yes, it is so good I rarely test new models anymore. Or think about cost.
Uhm, yes?
It's about 12000 lines of code. I could probably understand every line of it if I spent a couple of months on it, although I'd need to learn some things about WebGL, numerical computation, and solid-state physics. But I elicited it in four days. Probably I'd be better off spending the next couple of months doing something else instead.
So, I was faced with the question of how I could keep the artifact thus created from being completely valueless. My solution was to export the heat-equation solution produced by the solver as (gzipped!) CSV, so that I can use my own code (that I do understand every line of!) to check that the solutions found by the solver are, in fact, solutions.
If that check checks out, the backward Euler solver may be of some value even if I don't understand every line of it.
This morning I elicited a microkernel operating system from Opus 5.5. Well, mostly. It doesn't implement task switching yet; we'll see if it runs into a wall at some point. But it boots in QEMU, and it's running a user process in ring 3 and serving web pages.
I have enough real problems in life. I don't need to invent new ones just because a new technology is available.
Many of my problems in life are fully solved far past my satiation point by a 3b model that costs me nothing to run.
Many others are not.
But in either case, when I am acting and living wisely, almost all of my problems exist prior to the existence of technological solutions to those problems.
This is also true for the customers and employers that I care to work with. This has changed in me over time, but I now try my best to avoid inventing new problems. The world has enough big, important problems already.
> This morning I elicited a microkernel operating system from Opus 5.5.
This is cool but also a good example. I don't need a personalized microkernel just because it's possible to have one.
Maybe I need one and I don't know it, but the problem statement definitely isn't "I have inherent desire for a personalized microkernel".
You're talking about definition 1 in https://en.wiktionary.org/wiki/problem, "A difficulty that has to be resolved or dealt with," with the examples given being racism, addictions, and lack of access to health care. Those aren't the kind of problems AI can help with.
The kind of "problem" that AI can help with is definition 2, "A question to be answered, schoolwork exercise." That is the character of engineering "problems", although they are more open-ended than schoolwork exercises, because there are many defensible tradeoffs. "How can I build a bridge here?" or "How can I improve the fuel efficiency of this vehicle?" is a "problem" in the sense of a question to be answered, not in the sense of being similar to racism or addiction.
If the questions you're thinking of are so easy to answer that they can be easily answered by a hypothetical AI model 1/500th as good as Opus 5.5 — well, think harder.
But if on the other hand, I mostly use my human intelligence and just need a dumb model to complement my human intelligence at low cost and high speed (say review every commit to catch obvious bugs), I have a much better chance of building an actual moat than you do.
But outside of coding, it’s even more clear that you don’t need frontier intelligence. My customer service agent is very happy with a 100B param Deepseek flash model, thank you!
I’m using subscription models for exactly that, better models catch more subtle bugs, and they catch them faster. It works out far better in terms of work-hours saved.
Also A/ then OAI slashed token pricing by 2x~5x on their latest models
Look at the context in which I used that term 'good enough'.
What i was saying is that there are tasks for which a dumber model can be good enough, and for organizations with sovereignty/ privacy concerns, those concerns can be strong enough to incentivize the use of a dumber model.
I had the exact same experience. And unlike Fable, it doesn't gobble up your entire usage limit in a few hours.
I always wonder what the "good enough" people are actually using it for.
If Opus can’t 1-shot it, then it must rely on our human intelligence which can be complemented well enough with a dumb model as a frontier model.
Prices for the same task (ARC-AGI-2) have since dropped by factor 10 if you compare to Opus 5.5 Low.
The real question is if this model is good enough that it can still accelerate work, and not be a hindrance to real work like older Mistral models often were.
If they can do that, they'll have customers.
If these companies will steal from deep pockets like Disney or Sony (some of the most infamously litigious copyright trolls to ever exist), they won't think twice of stealing every bit of code you upload to them.
If your code passes through an AI company's servers, you can assume you just gave it to them. In turn, when your competitor tries to copy that new feature you just added, the AI is now trained in exactly how to copy you and eliminate your competitive edge. Unlike your employees, the AI isn't bound by the same rules and even if it were and violated them, your company probably doesn't have enough money to prove it in court (and that's if we somehow reverse some of the stupid "AI is the most transformative use of copyright I've ever seen" judges who have drunk the coolaid).
Most companies could build the compute to run GLM or Kimi models for way less than the potential loss due to IP theft from using third-party systems.
A simple loophole, use the code to create an RLVR environment where the resultant code is the end goal / max reward. Technically the customer data is never trained upon, but effectively you’re using it. Even better, use the code as a seed to generate synthetic data similar to it and use that synthetic data as rewards in an RLVR model.
Unless you can host the ChatGPT model on your own servers, which I know some enterprises are doing, I don’t think there’s any hope of protecting your data / competitive advantage from these frontier companies. Better to be paranoid, than be commodified by these companies.
> Following an investigation, we have confirmed that Buckmaster’s Codex prompts over the two months preceding this announcement and paper on September 8, 2026, could not have influenced the system in any way, including through training. The OpenAI internal model used for this result was developed through large-scale reinforcement learning on top of a previously pretrained model. Our proofs also differ significantly. In the Euler case, Alpöge and Buckmaster proved a result with external forcing, while OpenAI’s system proved a result without external forcing.
People who aren't afraid of rolling their sleeves into any code base? The difference is practically zero.
Maybe it's because people stopped watching what their agents are doing and stopped looking at the quality of the output. But I still see agents being absolutely mindless like a junior dev.
Recent example: it updated an an API to add newly released models to the backend. There's a list of models that require specific configuration for the reasoning effort and temperature or the API call fails. GPT 6.1 Sol misses this and code fails at runtime because the newer models need to be added to the list for special handling of temp and reasoning. Fixes it for one model and tests it for that model using an E2E test. But doesn't test the other models that were added for the same error condition...I had to explicitly ask it to do so and it finds them and adds them to the list and says "that's on me."
Yeah, not that smart.
I keep seeing this kind of thing over and over, and honestly it's not got _that_ much better since the big breakthroughs about a year ago.
For sure I happily vibecode stuff without worrying about it when it's a greenfield project, and if the LLM has written it entirely from scratch then usually it's well structured and sane. But making changes in messy, mostly human-written mature codebases is still a minefield.
This works fine with Opus 5.5. But it also works fine with GPT 6.1 Sol, Kimi K3 and MiMo 2.6 Pro.
It doesn't work equally well with Sonnet 5.5, interestingly.
For people who have some expertise, the models accelerate the grunt work, but you’re the one validating it.
"Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast as that!"
Yeah, I could buy a bigger and bigger vehicle when a new one goes to market, but using a semi-truck to drive myself around would not only be significantly more wasteful but also more inconvenient and limiting compared to a sedan. Very much like advanced closed models vs performant open ones.
And yes, open weights are still behind, but are catching up.
It's especially the case as more non-Americans look to self hosted models and domestic cloud inference providers using open models that the US providers who are still leading the charge need to drastically drop their prices and find a path to profitability in order to maintain their lead and retain the advantage they had as AI turns into a commodity (which is happening faster than I think even the frontier labs initially predicted).
I use MiMov2.6Pro, DeepSeekv4.1Flash, GLM5.3, Hy4, Qwen3.8 and KimiK3 at home. Opus5.5 is not a game changer.
I have to admit, Sonnet got really good too.
But Opus just uses tools, a broad spectrum of it, etc. it feels like sure if you add some router behind it you could split it up if you need to but if you give me the choice, its opus allll day long.
I hear this literally every other week about whatever the newest FoTM model is.
Unless you can provide concrete examples of things you can do with them that you simply couldn't do with last week's model, it's absolutely meaningless.
There is no ceiling on what you can accomplish with more intelligence, so there will always be a market for the best models, and that market is likely to just keep growing. If Opus 13.5 can one-shot a profitable company or discover a new disease treatment or whatever you can think of that a swarm of relentless super-geniuses could accomplish, companies (and governments) will throw money at it.
I also think there will always be a market for many sub-frontier models that will continue to grow rapidly as well, because "good enough" is definitely a thing for a given task.
No company would ever release such a thing