My money is on the more efficient solution. Even if Anthropic can win some benchmarks by using 3x tokens over 3x time, it is a terrible base to build toward the future. Users are no longer willing to wait exponentially long for linear improvements. And as we see from some of the Chinese models, they can quickly distill frontier models with the tax of being slower and more token-guzzling, while retaining most of the quality. The real differentiators are becoming speed and efficiency, which translate to cost and user velocity more than incremental capability improvements.
It's great that frontier models can solve complex math equations, but bread and butter LLM usage (where the money is made) has already shifted from "I need the best always" to "what solves my day-to-day problems quickly and consistently". Fable usage as a percentage is flat-lining. We are already at the point where output quality is negligible. What wins going forward is cost, speed, consistency and the compounding effects of "softer" improvements to the harness.
I've jumped between copilot, claude, gemini and chatgpt since the start of the year. chatgpt wasn't even worth looking at early this year.
Anthropic has the smarter models for sure, and seems to be default in corporate. However, the amount of budget you get with GPT as a user is much better, the harness feels more polished, and the models are faster. They are also much nicer to work with, I can just read the output for the most part. With claude I get pages of text and need to skim to find where the actual information i need to care about lies. So much more cognitive overhead.
Sol is smart enough for anything I've thrown at it, it's not one-shotting like fable, but I'm more willing to actually go back and forth with it, and it's likely producing better output to keep a human in the loop rather than trying to solve the world independently and making multiple incorrect assumptions.
I think GPT sees the market changing and is correctly repositioning themselves. Anthropic is down the wrong road, and if they don't correct course quickly I'm sure many of those enterprise contracts will start pivoting.
I was just discussing this with a co-worker yesterday. I would really like a model (or harness?) that worked with me instead of for me. Walk me through its choices and decisions, let me correct it and guide it along. I would be way more confident in it's output, I would be more familiar with the changes that are being made, and it would make reviewing the final code way easier since I was making the decisions along side it. I'm sure it would also reduce the "brainrot" we're all going to experience the more we hand work to these models.
I'm not sure how they'll survive their creditors tbh
My boss uses opus and gets good results when I use it always burns tokens. The other way is also true too I get great results with sol/terra but my boss does not.
“No, not like THAT!”
When I checked, all that credit was gone, I still wasn't into the next 5 hours, and all the agents had failed, returning nothing.
I didn't even get anything for burning all that credit. If I had paid for it, I'd be very, very pissed.
"Ultracode. When a system-reminder confirms ultracode is on, that opt-in is standing: author and run a workflow for every substantive task by default. The goal is the most exhaustive, correct answer you can produce — token cost is not a constraint."
gpt 5.5/5.6 goes further on its own much more often than opus 4.8/5 does. codex capped ~300k context when claude does 1m.
I don't feel codex is saving tokens, and result is usually not as good imo.
That's configurable in codex.. but there is a higher cost/usage to using it.
Granted it did make me think about my biases and to lean into more future facing inevitabilities. But you've nailed it, for Boris and Anthropic, they are betting that coding is a solved problem in the sense that any person can one-shot any random idea and the output will be in some abstract sense "good". And then at what cost and toward what end?
As someone who does near 100% of my coding via LLM these days, i still find that for anything complex i am still looking at and thinking in terms of code. Im still quality checking and steering at some interval via code. And im still not sure how or whether i can replicate that level of thought without still dealing in code at times.
They're both pretty damn competent and more important, Extremely Fast! I find the speed more useful than trying to be a hundred percent complete on every task. The big intelligent models screw up all the time as well, but I have to wait twenty minutes to three hours to find out.
GPT 5.6 is especially tenacious and seems to want to solve every bug in edge case 1000% all the time. Sometimes that's what you need, but a lot of times you're just trying to move fast and figure out what the product is.
Isn’t the race between Chinese open-weight models and the others more decisive for the future?
once we have a bit more memory fab capacity and the insane bottomless investment in AI giants realizes there is a bottom, local hardware will catch up with model performance to the extent that centralized inference will be downgraded to special cases or for orgs that find it cheaper than buying expensive hardware
but really most power users are going to have 1 TB unified RAM and local models that will do well enough
for light office use you can still have a cheap laptop and a claude subscription
I find it unlikely that there's some fundamental property of OpenAI's models' "personality" or style which Anthropic (or any other serious AI firm) wouldn't be able to match if they wanted to.
“In one notable technique, their prompts asked Claude to imagine and articulate the internal reasoning behind a completed response and write it out step by step—effectively generating chain-of-thought training data at scale. We also observed tasks in which Claude was used to generate censorship-safe alternatives to politically sensitive queries like questions about dissidents, party leaders, or authoritarianism, likely in order to train DeepSeek’s own models to steer conversations away from censored topics. By examining request metadata, we were able to trace these accounts to specific researchers at the lab.”
“You are an expert data analyst combining statistical rigor with deep domain knowledge. Your goal is to deliver data-driven insights — not summaries or visualizations — grounded in real data and supported by complete and transparent reasoning.” (variations appearing 10s of 1000s of times)
Does Dario have the same relationship with the truth as Sam? (Their companies pirate books and develop products that compete at some level with those books’ authors, so obviously neither are that wonderfully trustworthy, so maybe “no evidence” meant you don’t believe this evidence rather than you weren’t aware of it. I would understand and respect your lack of belief!)
This is evidence only if you're an American exceptionalist.
As much as I wish you were right, everything about software economics for the last 30+ years has favored _less efficient software_. Traditional hardware has been optimized for traditional software for decades and we still see bloated software win consistently. LLM hardware is at the start of its cycle, with abundant low-hanging fruit to conquer--I would expect the pro-bloat dynamics to weigh even more heavily in the LLM space than in the traditional software space.
It's hard to put a number on it, but even accounting for all the time in meetings, talking with stakeholders and developers etc I'm over 10% more productive overall. I earn substantially more than $2000/month, the ROI is there
It's only expensive compared to the currently very strong offering from OpenAI. Or other models - my hobby projects are all on DeepSeek
Perhaps that’s true, but it’s a price tag that’s a lot harder to swallow for many orgs than $200/mo, and would require some hard justification for how your increased productivity contributes to the business bottom line.
I’ll agree that you can probably do that with hard numbers. I am skeptical that most $200/mo users could.
I do believe it though. While the subsidies of the frontier providers are nice, they can't continue forever.
Most ordinary people don't care that much about privacy and if there is a solution that just works well enough to cover their needs and costs just fair enough to be affordable, they are going to use that. I feel like this is why fast food is such a thing. There are far better food options out there that either require a little bit more effort or money. Your post on Linkedin is akin to someone claiming in 1960's America that home-cooked burgers are the future, despite McDonalds gaining ground.
It is also why crypto wallets and the like never took off, people don't want to give up the convenience of a bank account, understandably.
At the end of a long session it will start saying stuff like “There are smoke tests on the foundation-gates that are left for the cutting seam checks on these domains, which is genuinely your decision”
Never before the past couple months have I ever not been able to understand wtf it’s even saying lol
I was working on a project recently that required the attribution of a data source, and it added to the "licenses" page of the project something like "We use <blah> and per their terms we owe an acknowledgement of attribution to you, the user."
So in addition to the stuff it says in a session there's gobbledy gook that it prints out in copy as well.
It's the better experience. The limits are way higher (I almost never burn through my $200/m plan), and the output is better than Opus 4.8 (Opus 5 is completely unusable for me).
Not doing crazy multi agent swarms, but have yet to hit any limits during pretty intense weekend sessions.
"From May 13, 2026 through August 19, 2026, your weekly usage limit in Claude Code is 50% higher."
What about errors compounding and any on-the-go decisions being made being the wrong ones? That kills anything long form, no matter how good and detailed your initial plan is - there will always be something along the way.
The models would need a way to identify a difficult problem and apply max reasoning there themselves and cruise through everything else at a lower level.
Once on a sandbox, realizes nothing works in its sandbox, then again outside the sandbox.
Fucking hell.
As soon as it gets annoying enough to switch to another provider your Claude code tokens drop right off.
"Your subscription will auto renew on Aug 20, 2026."
But I doubt they will remove it, just like they didn't remove Fable from the subscription. There's simply too much competition.
Edit to add: As a 3 month experiment, it has been positive. I have gotten things to work that none of the people I have hired and paid over the years took care of. And I myself have been too busy to prioritize. It's a useful tool, but only a fool would think it replaces the judgment of a professional. Not yet anyway.
Sideline LLMs have free compute and offer cheap prices to draw in crowd. Becomes flavor of the month LLM.
Back to step one.
It should be pretty clear by now that token prices are predominately a function of available compute.
https://www.bloomberg.com/news/articles/2026-08-17/anthropic...
(nobody uses claude anymore, it's too over-subscribed)
This happened to me, so far two months without any progress with Claude support trying to resolve it. Chatbot support got no response, email support got a single response after a month saying it had been passed on to another team to resolve. Mostly just silence.
I wish I could use my Anthropic sub with it but I heard you get banned, but at least you can use it with any other subscription or model.
For anything that makes money, it feels like a huge step down in productivity, even with the downsides of agent-assisted code.
The max account is for home projects where I'm basically making some common tools but tailored to myself (diet tracker, note app etc).
It's been great but across work and home it's just too much.
I replied, "never stop stopping!" and we had a stalemate, lol
They raise their prices too high, a lot of customers will still buy but be unhappy about it, it's burning goodwill for money. If they drop their prices too low they have overwhelming demand.
They could be making money hand over fist for all we know. We don't really know how much compute is being used, nor how much it cost them.
Why should I pay you money , and have slower model and less intelligence? Oh yeah and I do not care about limits with cursor ultra at all . Unlike with you .
But CC took some getting used to, I also prefer the Cursor interface (the legacy "chat" pane).
I'm kind of surprised that you haven't been watching it with /usage to see where you're at. I've had a couple times that it ran out and lost what it was doing and had to start over, so I've gotten more careful.
And the time it burned $100 credit (that they gave us) and the fizzled with multiple agents, leaving me nothing of that work... Well, that was certainly instructive.
Like, stop toying around with token limits and just focus on more efficient models.
Once the local models are good enough we are so abandoning these elephants.
I'm gonna walk as soon as possible.
The future is models baked into directly into hardware, onto bare metal, serving 10k+ tokens/sec.
We can have our little models, they'll be serving a different customer.
> the whole OpenClaw thing imploded
seems like just a month or so ago it was the new hotness [0]
[0] https://www.cnet.com/tech/services-and-software/from-clawdbo...
And smarter people than me can probably find some extra-linear relationship between usage and required limits for uptime etc.
So yeah, different people get different quality. I wouldn't be surprised if it calculates a wealth level, willingness to upgrade, influence, usage levels, etc. and decides what models to serve.
It's extremely obvious to me they've nerfed all models on my 5 Max from Fable to even 4.6, but i've tried through enterprise with huge differences in quality, and i've seen other people experiencing this both on reddit and twitter.
Don't take my word for it, use a simple prompt of "create a shader showing [complex scene]" and see what the other models vomit out, including opus 4.6/4.8/5 and 5.6 sol (compare fable low effort with sol xhigh).
Edit: If you're mad I'm light on the details I have provided some in the replies.
I have a Claude Code and OpenAI subscription so that I can use Opus/Fable/gpt-5.6 as I please, and the models are often catching things the other models missed. So much that I would significantly weaken my workflow if I dropped one subscription.
My best workflow at the moment is to create the initial plan with Fable (before review/revise-cycling with other models). From my own testing it seems slightly better at arriving at high-level ideal solutions after sweeping the whole project, projecting future needs, then coming up with good trade-offs like "by construction" correctness.
While mostly subjective, maybe the closest objectivity I have here is noticing fewer revision cycles needed with Fable-initialized plans.
I mean i think if a developer has a good handle of the code the difference is marginal .
Unless we 100% offload the thinking to the model and act like a prompt manager. Maybe
Even then, it's kind of a wash these days between the sota models, and we're talking about maybe a 10% performance difference or something. But every once in a while there's the experience of one model spinning its wheels on a bug/repro/issue while another model comes in and one-shots the solution.
I am asking because in my personal projects after a while they becomes a giant messy ball of wires and i basically trust the model to untangle it for me , by the time it untangles properly, I run into my token limits.
You can swap out "architectural simplification" with performance opportunities, bugs, correctness, etc. I get the orchestrator agent to then itemize it all into a file where I can keep track of which ones I've implemented.
The results are pretty astounding. I run these right before my weekly limits reset for each subscription and the findings will dictate the secondary tasks I get done during the week.
It's definitely token-heavy. I'm on the $200/mo Claude Code sub and the $100/mo Codex sub.
But it's pretty clear to me that software engineering is more or less solved and all you need is enough patience + tokens to get what you want. I think 20 years of engineering experience more lets me save on tokens rather than unlock things nobody else can build.
An example of the scope of one of my AI-engineered projects is a iterm2/ghostty-like terminal app that implements its own pty session, parsing, rendering. It's almost 2000 commits right now.
That said, I have a specific workflow that isn't just a blind "ok now make it so a screen can be split into panes". I have a plan phase focused on coming up with ideal invariants and such. But I'm not sure anymore how much of that is useful vs just yoloing a solution and then paying technical debt in sweeps, like garbage collection.
But the example you give is of a terminal for which there are copius examples in open source code. How hard is it really for a pattern matching machine to do that?
If I was doing it I would start by forking an existing repo and I might even say then that "software engineering is more or less solved since the open source revolution"
But I don't work on things like that.
While just simply trying many times independently gets you the improvement that is due to pass@k vs 1, you can get huge improvements if on top of that, depending on your setting, you find a way to ensure some stochasticity by perturbing tool calls, etc and running multiple instances.
The general theme is, embrace the stochasticity rather than the leaky abstraction on top of it.
With modern LLMs, investing in this kind of harness tooling is much more fruitful than hoping for the best from the model.
While many basic instances of this are built in to the popular harnesses (much of cursors higher-than-usual success rate with older models was due to really excellent context mgmt), you can never beat one that is optimised for your particular codebase, infra and general setup.
Until last year or so, the context management needed varied too much at too coarse a level across different models and even model instances, but now they are all extremely robust in a much higher % of contexts and are thus way more amenable to developing context management tools for, without needing to do a research teams worth of evals.
Custom evals and harnesses are thus extremely high ROI now. We are finding companies needing to do less and less tweaks and getting much fewer regressions (you should have reg tests in ur evals) with every new usecase and every new model.
It can be really simple to start with: change your grep/rg that it uses to a script that does in effect "rg $@ | shuf".
More complex examples are: giving different subagents different tools, randomly failing tool calls, truncating file reads randomly, having a small model invent N possible failure modes causing a bug and appending that to N prompts and starting subagents from each - this all forces each to pursue different paths. $example_specific_to_your_company_setup is highest ROI though, since most companies actual failure modes are dominated by idiosyncratic API shapes and retrieval quirks that no usual harness will bother modelling.
Also important IMO to not assign any meaning or semantically interpret the CoT as an acceptance mechanism (it is ok to use it as a rejection mechanism e.g if you see it plotting a sandbox escape whether it eventually emits the exploit or not is not something you want to hedge). We have to resist the temptation and ensure we only interpret tool calls, codegen, etc in our evals and only think of the cot as "some output that pushes the conditional distribution" which may or may not semantically match the typical preceding tokens of the desired tool call.
The whole thing is getting ridiculous.
I'll ask it to do something and it'll say, I tried, but I couldn't do it over and over again or some variation of.
But it doesn't do that at the start of my subscription, so...
Because just calling something bad does not add a lot to the conversation. It's not thoughtful, interesting, or good.
The Claude desktop app is also widely panned, as I mentioned, and for me this mainly is due to general UX and a poor remote control interface. Codex's connected machine support is top notch.
I also mentioned the value of the Codex resets!
Everyone has different experiences with these things (for example, I've never experienced what you describe), and "$X is bad" is not conducive to thoughtful discussion.
One thing I do like about Claude is that the normal (non-Code) chat interface supports MCP, whereas ChatGPT basically does not.
Opus 5 is fine for me and works better and faster on low and medium than higher effort on prior versions. Same as 5.6 Sol compared to 5.5 or 5.4.
If Wall Street sees a single outage or a tiny drop in usage, they won't be happy and will pressure Anthropic to take away the free tokens.
Better to reduce the limits now rather than to wait until Wall St. tells them to just to avoid a stock punishment.
"Thinking" aka 'trust us bro!' without proof of thinking.
Making the model waste more tokens.
Advertisement: aka you pay to be advertised at.
Silently downgrading you and still faking models with the more costly tokens.
Intentional strategies to eat more tokens with no real gains.
Giving out almost-but-not-quite solutions that require another pull of the slo(t/p) machine.
Over the long term I think OpenAI will produce the better experience when it comes to model quality, harness quality, and availability. I have been using codex the past few months and never looked back.