It would definitely cost me more per month than a x20 ChatGPT or Claude plan, probably around $400+ was my estimate at the time. This was with Fireworks (ZDR) which has since increased their prices (and got slower!).
That being said, very impressed with the model, and looking forward to what comes next. As the frontier models become less subsidized, the open models will become more appealing.
P.S. There are subscription plans for open models, but I've found most of them to be extremely slow, have model throttling (only so much of model X), and also very sketchy about training and data retention. No thanks! If you want to share your data, just use Muse Spark contributor. Seems impossible to beat that on price per task if you don't mind feeding your data to the Meta machine (spoiler: I won't).
I also spent $280 on DeepSeek doing the tests (direct to DS, not OpenRouter). I suggest that if you can't conceive of anyone spending $200 on DeepSeek, you're not being ambitious enough!
https://openrouter.ai/docs/guides/routing/provider-selection
Edit: others have noted the provider and harness matters. My experience is with opencode.
What harness you are using?
The shape of my work changes obviously, so it'll vary, sometimes more, sometimes less. For example, fixing all of the bugs and defects I found that week was 2-3 times the effort and chewed through my ChatGPT allowance, but I had banked resets...
Also worth noting that codex models have been kind of all over the place recently with their usage... and it looks like costs are changing again.
I can’t overstate how bad of an idea I think using an AI for customer interaction is.
But yeah, send me a non solicited AI slop email or worse, political ad, and you dont get the dignity of me saying stop to unsubscribe. Straight to spam for you.
I dont consider myself a programmer but use LLMs almost exclusively for coding.
The number of people able to create useful software today is much much larger than it used to be and arguably a minority of these people are/were "programmers"
Subagents are like trading derivatives. You can lose as much as you want.
When the regulations do arrive, I think they should really focus on AI companies and API providers being more transparent wrt how they're billing their customers. Because right now, it's a totally vibes-dependent and a mess.
A smaller model in the same generation will never be the same as a bigger one, assuming this is a smaller model, and the same generation, as naming implies, it will not be comparable, it might be on the benchmarks, even on the benchmarks that matter, but the whole story should also give the drawbacks.
Excellent pithy warning.
But there’s a point on that spectrum where the ability to run multiple experiments in parallel, even with a significant amount of (one time) wastage, is overall more cost effective than the alternative.
And watch 10 hours of football on Sunday for our DraftKings bets.
Parallelism is fantastic when it actually speeds up the entire pipeline, but in my experience most people's jobs (at least the ones for which AI is currently relevant) involve a lot of overlapping "hurry up and wait" branches that drastically blunt the real benefits of that sort of parallelism.
There may be specific situations where it makes sense to do it, but just immediately going full gastown on anything AI related seems like such a giant waste to me, of both money and finite world resources.
1. write a sketch of a spec by hand
2. have the llm review the document and question me until it can generate a spec
3. review the spec and revise where needed
4. have it write an implementation plan
5. another round or revision/review
6. executing the plan step by step through the plan, plausing between each step to see if we are still on course and if the decisions it made track with my understanding of what we are doing.
I've been working for a couple of hours tonight, the total cost of the session is €0.6.
it's not the build this thing end to end, but also not quite write function x for me. It is still a lot of manual review, but I find I really need it to even discover what I actually want to build. I just cannot imagine building something in a single shot and getting something that actually has value (unless it is basically a clone of an existing thing). To me the whole value of ai right now is that it's now very cheap to build custom software that exactly matches your preferences.
https://www.youtube.com/watch?v=WAeHgE94rVo
The system performs quotation attribution on my local hardware for my near-future, hard sci-fi novel (having nearly 500 quotations) with over 97% accuracy.
The initial prototype was developed quite quickly, but numerous successive iterations were required to fix numerous gaffs by Opus 5 (because it doesn't actually _understand_ what it takes to make general-purpose audiobook narration software).
The only thing I've found Deepseek and Kimi good for are security tasks that GPT refuses to do.
This is a summary of what Deepseek did and got wrong:
Lost the proven baseline: changed kernel source, configuration, compiler, RAM geometry, MMC width, and peripherals together. Matching an upstream commit did not preserve local boot fixes, making failures difficult to isolate. Misidentified an image: a file labelled “r18-known-good” actually contained the r23 parent bootloader. Filename-based reasoning replaced verification of the artifact’s identity and provenance. Shipped inconsistent boot contracts: flash-16b’s loader read too few kernel blocks. Fresh2 changed the device tree without updating the loader’s expected length and CRC, creating deterministic rejection before normal Linux handoff. Patched binaries without maintaining reproducible source: loader constants diverged from source, a separately compiled cache-flush length remained stale, and assembly used an oversized stage-two slot. Their causal contribution to hangs was not established. Overstated diagnosis: claimed failures were definitively in U-Boot, blamed compiler or IPU changes without controlled isolation, converted noisy observations into confirmed hangs, and neglected persistent journals as an alternative explanation. Mistook compilation for integration: framebuffer registration was incomplete, timing success handling was inverted, BT.656 selection was unreachable, encoder overrides were missing, and audio lacked software clock configuration. Misread hardware evidence: asserted interrupt-free PMIC operation, assigned RF to the wrong SPI controller, confused regulator identifiers with register addresses, and described repeated encoder writes as unique registers. Overclaimed results: treated kernel/probe indications as userspace success, presented earlier discoveries as new progress, and omitted failed flashing attempts from the final narrative.
There's your problem, 4.1 Flash is significantly better and cheaper, to the point where the official DeepSeek API is going to (or already has, I forget) redirect requests for Pro to 4.1 Flash, and adjust billing accordingly too.
4 Pro is still offered by providers I'm sure, since it's open weight, so I can understand making that mistake.
Although I do think Luna 6 max is ok for some basic things, would never use it for coding myself.
For CRUD shoveling, models like DS4.1 are enough.
And the intelligence gap between cheap and premium is closing, as can be seen from the title of this post.
Who cares if your car can go 200mph if all you need is 60. If my requirement is 60mph, I want a faster 0-60, not a higher top speed.
Deepseek 4.1: $0.02/$0.60
Just to illustrate how cheap the Corolla is in your analogy. Also Opus output would be $50 without competition.
I use it as main Hermes model that orchestrates codex/droid harnesses with subscriptions for heavy dev work
I do have ChatGPT as main assistant that sets direction and delegation of projects to Hermes
At my increasing usage, kind of 200 usd subscriptions makes sense and max out on Luna max
For raw productivity most of what works is best and switching will cost you getting on use parity with other models, as you need to learn what they good at, potentially how the tool works and how to prompt it best.
For tasks that you implement in code, you should have benchmarks and evals.
That said for me was Luna a huge leap and 500+ of cost savings a month
It's a very noticeable hit. It's like programming with mid-2025-era models: ignored instructions, dead code, mistakes.
It works...but anyone going from Sol to Deepseek is going to have a rough transition.
And no it did not deliver. A lot of it was re-done by Astra
Why do you expect that $200 will give you that on ANY model? Multiplayer FPS games are very difficult to make, no AI will deliver that today.
Astra was able to model low poly enemies, rig them, do simple animations, and greatly improve procedural generation. I have it modeling assets in blender every day, which I often have to go in and fix
I get the same UX on every platform, works perfectly on very low bandwith environments such as in a cabin, in the subway or in the middle of nowhere.
I tried using other harness such as Pi and opencode but I did not like them. If Claude Code gets weird I can swap in an instant.
You just need to follow this guide and disable artifacts in Claude Code's config: https://api-docs.deepseek.com/quick_start/agent_integrations...
I found that ssh is pretty bad in these scenarios, so my usual herdr over a phone doesn't always work properly.
Now I'm using paseo which in principle solves my issues properly, but unfortunately it's "reconnection" and state sync is pretty slow (probably going over their servers).
Use the model through a fast and reliable provider such as Fireworks directly, skip OpenRouter.
Then I installed helix and I just use it without config.
If you like configuring things take pi, if not omp is pretty much great defaults.
curl https://tg.st/u/0001-fix-unblock-all-commands-in-bash-tool.patch | git am
curl https://tg.st/u/0002-feat-add-light-theme-with-auto-detection-for-white-b.patch | git am
curl https://tg.st/u/0003-feat-enable-yolo-mode-by-default.patch | git am
curl https://tg.st/u/0004-fix-disable-mouse-grabbing-to-restore-native-termina.patch | git am
curl https://tg.st/u/0005-feat-skip-project-init-prompt-and-quit-immediately-o.patch | git am
curl https://tg.st/u/0006-feat-remove-scrambled-rune-animation-from-waiting-sp.patch | git am
curl https://tg.st/u/0007-feat-remove-quit-banner-and-thank-you-message.patch | git am
curl https://tg.st/u/0008-feat-show-output-in-full-instead-of-collapsing-trunc.patch | git am
curl https://tg.st/u/0009-fix-discover-map-model-features-advertised-by-v1-mod.patch | git am
curl https://tg.st/u/0010-feat-keep-large-and-small-model-selections-in-sync.patch | git amI haven't used deepseek for anything else but the above results make me question its overall capability. Meanwhile qwen3.8 has continued to impress.
Like who can figure out the ordering? With Luna < Terra < Sol < Astra it's obvious at a glance.
I propose for some third company to name after monsters: Cyclops < Minotaur < Ettin < Cerberus < Hydra < Kraken < Nyarlathotep (the AGI singularity stage)
It.. is?
When was the last time you heard anybody say "opus" or "sonnet" before this?
Also it implies that "haikus" are inherently inferior to longer texts which may be kinda frown-inducing..
I feel like people who don't get it immediately are just being deliberately obtuse.
The more obvious, to me, interpretation was distance from us so I would've expected Terra -> Luna -> Sol -> Astra.
When I think of space and astronomical scales, I think of distance as much more common concern.
Many of which are thiccer than our beta ass sun
Pretty badass name tbh
clear would be something like
piss-cheap - it’s-alright-i guess - okay-relax - ouch-my-wallet
Not sure that's why they did it. But that was my experience.
The smoking gun is how much slower than Sol 6 this is. It's not a retrain.
This is the actual big announcement. 50% cheaper cache than GPT-6 Sol will get you far more mileage on Codex.
Cache doesn't help you much when you are compacting every 5 minutes...
I was shocked at how quickly I ran out my $100/mo subscription with a single agent (sol medium).
This is why these companies are struggling to make money, they're chastising their customers just like they've been chastising the human race.
It's very appropriate in the cases when you're holding it wrong. The fact that you're paying doesn't mean that you can't make mistakes or waste resources.
No LLM will be cost effective if it's compacting this often. You have to find a way around it.
Apparently OpenAI makes you manually setup their 1 Million context window, and it seems to be only documented on X:
https://x.com/thsottiaux/status/2089082893804896524
There's at least a forum thread about it here:
https://community.openai.com/t/why-does-codex-report-a-258-4...
Suggest trying it out yourself: Ask for something difficult from GPT-6 Sol and Opus 5.5 and watch what each one does. The difference is stark.
Edit: Defining "difficult" as a complex coding or systems task (or even series of them in a single prompt).
> Ask for something difficult from GPT-6 Sol and Opus 5.5 and watch what each one does.
I gave the task to codex first, sol 6 xhigh. it took a couple back and forth prompts to define the project and then it worked for a bit and to took a couple more prompts before I decided it was good enough - not perfect, but close. It re-implemented some wrapper components in a simplified way that lost some of the UI, but it would work.
Opus 5.5 high took the same prompt with no back and forth, it just went off and one-shotted a tool that takes pixel-perfect screenshots of exactly what my app looks like.
There is way too much subtlety in what does and doesn't work for a given problem, context/prompt, tool set and eval. I can tell you Fable is generally better than Haiku, but comparing similar tiers really does depend on your exact context.
This was the biggest thing I noticed in the 6 models; their conversational prose is dramatically less grating.
Oh yes, I know GPT-6 Sol is ... quite not up to par. At least it's not as bad as GPT-5.6 Terra I suppose.
However it's less willing to obey your instruction so it's less usable for general runtine flows.
Looks like 5.5 is the new 4.6
They form these super strong opinions after a few prompts, then face reality over time.
People have been talking about how good whatever model is at “complex” tasks since the beginning, never mind that all of those models are now outperformed by Luna which many people consider unusable for complex work.
I get better results and usage our of my $20 claude sub than my $100 openai sub... it's that ridiculous
Astra used 215% of a week's budget (I burned 2 free resets) and took 13 hours. Opus used 20% of a week's budget and took 20 hours. Both were asked to use lesser sub-agents for implementation grunt work at their discretion (Luna, Sonnet) as long as they manage and review the output.
The timing comparison is not that interesting because the wall-clock speed mostly reflects how often they ran the (large, slow) test suite, not their coding speed. Although in the past my gut feeling is that OpenAI models do generally respond faster.
The quality of their implementation was more interesting. There turned out to be a bug in one of the unit tests the agents were trying to pass. Opus interpreted the natural-language requirements from the task packet, found the test bug, and fixed it. Astra tried hard to solve the problem without altering the test suite. In practical terms Opus got much, much farther into a useful implementation. Astra was still stubbing out and faking critical parts of the implementation (B-splines) and since it ultimately couldn't pass the full test suite, finally gave up on its implementation. Astra wrote some useful tooling in the process of its efforts which I ended up integrating into Opus's version of the code, but otherwise its approach was behind.
Now, this is just one comparison in one domain, and arguably Astra's strict adherence to the tests as-given is a good thing. But Opus wasn't merely loosening the rules / moving the goalposts to pass, it spotted an actual bug, and was more successful at doing what I actually wanted. And the cost difference was Astra-nomical.
Out of curiosity for an interpretation free from my personal bias, I gave Astra a hint from Opus and permission to change the test in question, which it did, and got a bit farther, but still ultimately didn't produce a working implementation (to be fair, Opus's was not completely working either, but was closer). I then fired up fresh agents to review the two repos. Predictably, an Opus agent thought the Opus-written repo was the better basis to build on, and an Astra agent thought the Astra-written repo was the one to keep. They were not explicitly told which was which nor did the commit trailers say, but I assume they can tell. However, after doing this twice each, I saved the 4 review reports into another folder and did yet another meta-review of the 4 reports, so each would see the arguments and critiques both directions. In this meta-review both Astra and Opus converged on preferring the Opus implementation.
(I did use some CC for Fable when it came out, and it was... ok. Not the worst thing ever.)
Opus 5.5 fails the "understanding" tasks which Opus 5 passes. I feed it a script which takes two numbers and prints the max of the two numbers. Opus 5.5 thinks it prints 1/0 instead of the max numbers. Opus 5 gets it right.
Here are the outputs from both: https://gist.github.com/dom96/b5bce82b6e6c1ebd5271ed70ad941b....
Looking at that Opus 5.5 fails to deduce that the "hack statement" is actually an if statement in disguise, but Opus 5 gets this right. I feel like this is a pretty good test and shows Opus 5's greater intelligence.
With the 80% price cut, this is competitive with Opus 5.5 despite the subscription downgrade.
Additionally, it was said that existing 20x subscriptions retain the higher limits for some time.
I have seen you make these immature accusations that users here are OpenAI employees multiple times today.
This is not even remotely true.
If you're running 2-3 parallel agent session with a few sub agents and waiting for you to prompt them, you'll have a very different experience!
This is a tiny minority of people, not “most people”.
(Usage limits are entirely dependent on what you're doing with them. If you're not running it on 1 million LoC codebases you can get a lot of mileage out of even a 5x account particularly with the recent cheap models)
Sol 6 was so bad that I switched over to Opus 5.5 exclusively.
Huge regression compared to Sol 5.6, often doing really dumb things. Same for Luna.
Even Astra is very unreliable for coding. Brilliant for vision, sometimes just great, but it also often does very stupid things.
I'm a bit sour on OpenAI right now and skeptical that 6.1 will be much different.
(Note: this is after preferring and shilling Codex/OpenAI models for the last half year)
I've implemented multiple features side by side with Opus 5.5 and 6 Sol, and the Opus 5.5 results always have fewer high severity bugs and require fewer rounds of fixes to get it over the finish line.
If 6.1 Sol has actually matched Opus 5.5, I'd be very happy. However, benchmarks and real usage don't seem to agree in my own tests. So we'll have to see.
For coding specifically, I've found 5.6-Sol > 6.0 Sol > Astra.
For modeling and artwork, Astra has been great routinely outperforming Kimi.
This is reminiscent to me of what Anthropic pulled back in February with their adaptive thinking rollout.
I can't wait for technology to catch up to a point where we can rid ourselves of this oligopoly.
I used about 10 hours of Astra high-thinking compute time and it was a bad experience. Incredibly slow (prompts running for 30/40 minutes) to do simple things. As a result, Astra didn't get much done. It needs the same small implementation slices as GPT 5.5/others, but was much slower and didn't generate better results. (On a complex infra project/across a large codebase.)
It was absolutely terrible on a few long running tasks (~2 hours each). It really doesn't seem to be better than 5.5 at most programming jobs.
I'm on a $200 per month plan with OpenAI, which I am happy with and is definitely worth it. But I also use Google Gemini a lot (paid plan) and it is incredibly fast. Like I can't get coffee fast. Like I can't send an email fast.
OpenAI is making some excellent products for sure but I'm not going to keep using Astra unless I can get some benefit from it. It really seems like even the frontier models just aren't good at working autonomously on large codebase situations. Just because something compiles doesn't make it right!! In one of those 2 hour implementations, Astra engaged in *fucking EPIC cheating*. It wrote a probe/side app and then worked through the design there. Um, what? Not that it's invalid to do this but I actually have to test in the live codebase or I can't possibly say that something is working.
Just because you can, doesn't mean you should.
GPT 6 needs to be babysit, otherwise it starts doing ridiculous things.
The lackluster GPT-6 Sol has been superseded by this apparently much better 6.1 Sol within a week.
I am very skeptical of claims that old models weren't much worse. Compare this to February's GPT-5.3.
I could point out that I said 6.0 seemed good only in comparison to nerfed 5.6 - people would say I’m just a RSI denialist - but now it is in vogue to accept that 6.0 sucked now that 6.1 is out.
I'm by no means an AI booster, but given 2022 - 2026 progress I'd say it's "exponential" in the sense of, "holy shit, every year I can do more and more genuinely different things", not "RSI mind reading intelligence can do anything is here".
I don't think Navier-Stokes level intelligence translates over to my projects, unfortunately. Yet? Who knows.
> I haven’t seen actual capability growth since ~January, and I’m pretty sure that was all tooling/harness improvements.
Even if that were the case, I'd say that it's improved in practice. And just from a philosophy perspective, if you're trying to imply some kind of mind dualistic way of viewing things, uh, I disagree with those theories of intelligence strongly (which also incidentally also disagrees with AIT-style theories of intelligence on one axis, though I have many bones to pick with the culture there).
Here's a good example with some assumptions on my part: I work in C++ and it really feels like the models are trained so hard to keep everything compiling all the time. That's a huge negative in my opinion because what happens is that the AI will do things like use wrappers to keep things compiling, even when that basically results in creating or hiding abstraction leaks. Or they get sneaky and include a header they shouldn't. Or they actually do see that there should be a layer boundary and they write some kind of abstraction to cross it but the abstraction itself is garbage or doesn't follow existing API patterns. The AI could invent 10 different, new patterns when there is already 1 existing pattern they should use.
I feel like a lot of this involves a lot of babysitting prompts. Not that there's anything wrong with that of course.
On these metrics it is much better than it was in March of 2025 but no better than it was in March of 2026.
5.6 Sol in the last two weeks became much dumber such that what used to be one correction turned into endless rounds of corrections before just giving up and coding it manually. I’m mostly having it do the “chore” part of coding so it is disappointing that it isn’t better at that.
Yes, still running into this, but surprised about this
> On these metrics it is much better than it was in March of 2025 but no better than it was in March of 2026.
I was super hyped at the agentic thing a year ago (Fall 2025), but designing functional software was hell. It would not just "grasp" the right level of "here is the essence of what we need" versus "these are all the small impl details". But idk I feel like Astra's the first model in quite a while that I don't feel genuinely annoyed at handholding a toddler with a PhD.
But I totally believe you on the 50/50 thing. Even recently as a few days ago, Astra did the thing where it ran into an error, and instead of making the sensible bounded decision of "make user retry in this case", it silently built an extremely elaborate recovery state machine w/o looking. These pathologies by no means gone, and I'm still careful in the design phases (which themselves are bounded and incremental) to sus out if Astra's gonna do this kind of RL slop failure mode.
For my use cases personally though, it's been better and better. I can't use AI at work, so you have much harier edge cases than I do, but still.
What was a pleasant and productive experience is becoming increasingly frustrating and draining.
I mean Opus 5.5 is absolutely fantastic, unreasonably and unexpectedly so, but Astra was great and as far as I can tell SOTA until, when was it, 3 days ago, no?
(Sol 6 idk, have not used it much for coding really. Seemed to work just fine when Astra used it in Codex as subagents.)
Since like last December I haven’t had any issues getting work done with whatever the latest Anthropic or OpenAI models at the time were. Tooling and models have only gotten better since then.
Astra seems better though.
Showing one potentially saturated benchmark doesn't necessarily fill me with a lot of confidence in the coding results.
Actually, it is so close to Astra, that I'm starting to beleive it's a slightly quantized version of Astra.
[0]: https://aibenchy.com/compare/openai-gpt-6-1-sol-xhigh/openai...
Which, honestly, is fine. A lot of juice to squeeze in efficiency and even if models got zero more capable, making the capability that is already here cheaper is a huge win for everyone (except Nvidia)
Not once has any of these predictions come true, the pace of progress has continued on it's exponential trajectory since ChatGPT first came to the public's attention.
So why now? What is special about today that suggests all of this is coming to a screeching halt despite all evidence to the contrary?
Did it? Model wise? I would understand agents wise, sure. But model wise? The attention to detail from the model? The ability to recall minute things? Improvements are there, yes, but mostly on Fable and Astra. Opus still isn't as attentive as Fable in long term writing for example.
Sure, Opus 5.5 benchmarks better than Fable. Sure. But is that the model, or is that the RL for agentic work?
From where I'm standing, the model work has not been exponential at all, and more and more it looks like the latest and greatest is getting too expensive too fast. Both 5.5 and 5.6 chat models got nerfed, actually nerfed not the tea leaves kind. In mid 5.5 cycle the chat model lost the ability to substitute names if given an outline. 5.6 cycle the chat model lost the ability to use paragraphs after a few hundred words (coinciding with Chat/Work split).
There's a race from OpenAI to serve dumber models on chat. I'm not even sure who they are racing against, but the fact that Astra, Sol 6.0, and now Sol 6.1 not being available for chat, should tell you that those models are expensive, and not the kind of models that can be freely "chatted" with on a subscription. OpenAI much prefers you use Work and limit the chat usage, much like Grok and Claude. I'm guessing they will announce that later during the dev days.
That could be cost cutting too, true, but really? That's the only explanation? And nothing else?
Sure, the progress did not stop. But it is nowhere near close being exponential when it comes to LLMs themselves. Agents are separate.
These things are knocking down Millennium Prize problems while a substantial subset of commenters here are still thinking about stochastic parrots.
do you think it will be exponential forever?
Fabs.
Either needing more fabs, new types of fabs, retooling existing fabs.
All of that takes years.
maybe we can design our way out of that too. But, I suppose that would be the similar breakthrough you are mentioning.
To make a manufacturing analogy - ChatGPT was a manual machining mill, and in the years after we've gone from that to a 3-axis CNC mill. Now we've added a 4th and 5th axis, which is great for the 2% of parts that need that functionality. But the big win was that initial jump from manual control to CNC. Why would I pay an extra $2 million for my CNC machine when I could just design my parts to be simpler to produce instead? The AI labs are trying to make these incredibly complex tools, but the market doesn't want/need them so they're competing on price for the tools that people do use. By selling their metaphorical CNC machines for half of what they cost to produce.
Oh, and we've bet the entire economy on the hope that fancier CNC machines will magically solve all our problems in all industries, from healthcare to the legal system.
So - will AI progress continue to improve? Sure. Will we continue lighting money on fire in order to make it happen? That remains to be seen.
In some aspects sure, but in others no. Open AI's goal is to build "highly autonomous systems that outperform humans at most economically valuable work." and Astra was a big jump in that. There still isn't a better model for computer use and vision/spatial work. Driving, Operating Robots, Video Editing, 3D modelling, graphics are all things Astra was >>> at than any other model. I'm sure you don't care about any of that so it's easy enough to slip by you but this analogy - "Now we've added a 4th and 5th axis, which is great for the 2% of parts that need that functionality." is dead wrong.
> "Opus 5.5 is so good that I don't want it to be replaced anytime soon. Stop training models[...]"_
What a time to be alive.
- plan youth soccer practices
- develop well-formatted soccer game substitution schedules
- build and ship software in languages I haven't used in 25 years on platforms I've never programmed for
- do meal planning and build shopping lists
- prepare grocery shopping carts
- solicit medical advice
- perform Garmin watch data analysis
- administer devices (with SSH access) using natural language
- avoid counterfeit soccer jersey purchases
- create "Warrior Cat" graphic novels
- make cartoon strips
- troubleshoot appliances
- manage finances
- review accounting ledgers
- diagnose malware infections
- so much more
And we do it all from a simple prompt that we can talk to if we choose.
I've built more (and better) software in the past month than I did in any given year in the 30+ years I've been programming.
I can understand pessimism regarding how this affects society. I can understand pessimism regarding how this gets abused. But for the life of me there's no good reason at all to be pessimistic about how quickly this has improved.
I feel similarly, but I think it's a valid question. Why is all the software I'm using not getting better? To be honest, I feel it's more buggy than it's ever been.
- it’s correct there isn’t much fresh data anymore
- it’s correct that compute is scarce, that was 100% the case and a huge issue at the beginning of the year, it is better now but still scarce, and hardware is now way, way more expensive
- it’s correct the finances don’t make sense
But there is no way to know when a bubble pop, because it’s a psychological phenomenon across an extremely complicated distributed system (ie the stock and bonds markets)
If I have some ML workload to run I can buy $x of Blackwell chips or I can buy significantly less $ worth of Vera Rubin chips to get the same performance. That's the key thing to keep in mind when you're talking about financials.
On some tasks in this benchmark, the models seem to be coming up with novel solutions. For example, Astra came up with a relatively simple formula for a sequence that only has 8 terms in OEIS and is considered "hard" [2]. It produced a lean proof that the formula is correct, but I'm just starting to learn lean and don't have enough expertise to check it.
[1] https://proceedings.neurips.cc/paper_files/paper/2025/hash/c... [2] https://oeis.org/A000530
It just seems like these claims are constant and looking back the calls of 'plateau' between 2023 and 2025 were clearly false, why should we think it's different now?
if true then LLM related AI (post-post AI winter AI?) is probably one of the fastest inception-to-plateau tech sectors to have ever existed.
We're still improving transistors on a somewhat routine basis.
I think it's more a token-cost-demand plateau. They've reached the scale and investor trillions to which they can't 10x the hardware cost of inference any more. They can't afford to compete by eating costs and there isn't appetite for more expensive inference.
So in order that they don't bankrupt each other they're looking for the legal cartel behavior coordinating a stop to growth by convincing governments to regulate them into stopping.
There's a lot of juice to squeeze in efficiency but only so much whereas it seemed like capability was going to continue to scale with parameter count.
Maybe it's good news for everyone that model capability is now going to scale on semiconductor cost meaning huge players are going to be very motivated to make semiconductors cheap.
Edit: removed a comment that was uncharitable and rude, for which I apologize.
We are seeing multiple frontier models dropping on the same day and no one bats an eye, because it's more of the same.
We've gone from 80% in some places to 80% in some more places.
Any area that is verifiable will trend inexorably towards 100% over time. In unverifiable areas, it'll always be "80%" because the ubiquity of "AI" style erodes its value, and ">80%" for unverifiable things involves fashion, cachet and "vibes" that humans will probably never knowingly let it have.
I don't think that's the motivation, it's because both companies want to IPO and the _only_ way to even hope to be profitable is to do a whole lot less training, which costs a fortune. But unless Chinese labs go along with this gentleman's agreement (they won't), slowing down on training will bring about the inevitable Chinese model parity date more rapidly. At which point the game is well and truly over for OpenAI and Anthropic. Bit of a pickle they've gotten themselves into with the emphasis on being best, with premium prices to match.
People were talking about plateau for years already.
It's not even anything controversial..
it's also why there have been so many calls for regulation and slowdowns.
I see posts about OpenAI and Anthropic latest and don’t even care looking at what they do better. I just read the comments here.
I use DS4.1 Flash and GLM 5.3 Flash, pay peanuts per day and get more than acceptable results.
Insane pricing pressure on the horizon. Even if big companies will not go with open weight models, the threat will be ever present that they can instantly flip flop on providers.
I remember when bandwidth was super expensive and now it’s dirt cheap.
Consumers are now saying the new pricing with lower usage caps is not so great. https://news.ycombinator.com/item?id=49896975
DeepSeek understands that. Grok understands it. Every other AI company thinks they need to be the best at everything all the time and it’s weird.
This model might be the first step in that direction, as competition heats up between OpenAI and Anthropic.
Here they are for GPT-6.1-Sol: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
They're not notably different from the GPT-6 family pelicans: https://static.simonwillison.net/static/2026/gpt-pelicans-gr...
GPT 6.1 Sol — 91, ~9 min, $0.51 https://jonclegg.github.io/pacman-bakeoff/#gpt-6.1-sol
Opus 5.5 — 99, ~9 min, $2.00 https://jonclegg.github.io/pacman-bakeoff/#claude-opus-5-5
GPT 6 Astra — 87, ~10 min, $2.42 https://jonclegg.github.io/pacman-bakeoff/#gpt-6-astra
Opus still plays the best. Sol is almost as good and way cheaper. Astra costs the most, scores the least of the three, and the UI is full of slop copy and design.
Full gallery: https://jonclegg.github.io/pacman-bakeoff/
Interesting that High got the render order correct, with the back leg behind the bike, while xhigh and max have both legs on the same side of the bicycle. Astra only got this right on Max.
It's not frontier pelican without the back leg behind the bike frame IMO.
For example, GPT-6.1 Sol High gets 75.2% on DeepSWE and XHigh gets 71.9% and is more expensive
https://openai.com/index/introducing-gpt-6-1-sol/#deepswe
Also, how many times did they test each condition - just once or a few times? are they showing an average of multiple attempts, etc..
With that benchmark I think even if you just run it once overall but the benchmark includes multiple runs per task as part of its scoring. DeepSWE is on GitHub if you want to check the run details.
It's because they need subscription money and interaction data and so keeping a version bump in the wings to stop the bleeding from your competitor's version bump is the logical thing to do. It has nothing to do with RSI.
Like think about a software org with good CI/CD versus one without. The mature org can do consistent incremental releases because each one is safe and low overhead, the messier org will do fewer big releases because each release requires a big effort on its own.
As model developers mature we might expect to see more frequent point releases rather than the big bang evolutions.
Luckily it's not a mistake as now we have access to . . . dots.
(and sol 6.1, it seems)
See, that's an/the issue. As soon as people start to flee to the improved model, they start to serve degraded models to keep up with the demand.
Edit: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...
>RSI
Recursive improvement doesn't imply increased rate, another word for it is "iterative" but this probably sounds too boring to some people.
EDIT: I love getting downvoted by openai and anthropic employees or their bots.
And yeah I have worked with Anthropic and OpenAI models, they're good but they cost a fortune while Chinese models are already really good at a fraction of the cost.
Is that what the Chinese models are capable of? If so, how are you using them? API? Or is there an inference provider that is as fast as the big 2? What about the coding harness?
At the same time, OpenAI is also making its existing $200 Pro plan less appealing. In Codex and Work, $200 Pro subscribers will see their included usage decrease from 20x of what the company offers to Plus users, down to 10x of that same allowance. In ChatGPT, meanwhile, GPT-6 Pro message caps will decrease from 200 to 100 per week.”
https://www.engadget.com/2272106/openai-adds-dollar500-pro-s...
Yikes
Our VC-backed subscription days are numbered
Well, the time it takes to compress frontier intelligence down to DeepSeek V4.1 Flash costs (basically too cheap to meter) is dropping, and the differential between the two is also dropping...
So... who cares?
Lastly, I'd like to actually use it in the real world to see how far my plan goes or if its unusable.
Tibo said that the existing $200 subscriptions keep the 20x factor for a while.
Ultrafast would have been nice with the temporary "Pro 400" plan.
But $200 is likely the ceiling of what people will pay for a subscription with usage based on vibes.
$500 for the old $200 is definitely a fumble.
This is missing an important context. And I actually remember this well, because I was saying that too. And the reason I was saying is that $200 plan didn't come with API usage, it was a chat plan.
It made no sense up until they started including API usage. Just as $500 makes no sense now.
> costs going to 10% of white collar income.
There's a permanent and ever lowering ceiling maintained by open weight models. It makes no sense to justify paying 10% of income permanently for something that will get you unlimited local inference for a 6 month subscription cost.
Now you can use it in coding harnesses that call the API.
Long term, this only works if you have a non-commodity, and if the higher tier is actually more profitable. We'll eventually learn whether both are true. For OpenAI right now, it's probably enough to just increase revenue, even if the higher tier is even less profitable.
Now, as it's linear, it makes much more sense to downgrade to 100$ OAI and pick up a 100$ Claude sub. (without doing the numbers) the usage should remain the same, total paid the same, but having access to best of both worlds. It should be a win for the user, and a loss for OAI.
With this in mind, it sounds like a fumble by OAI.
That they are expensive and climbing doesn't negate my point if the cost of the subscription over how long you plan to keep it is equally or more expensive than the GPUs. You can put together dual 5060 Ti or 5070 Ti systems to run local LLMs too. You don't need to splurge on a 5090. That's a bad option at this point.
I've messed around with Qwen3.6-27B but I'm not sure if it could yet even replace Luna for me.
I can justify $200/mo but more than double is not appealing to me.
Basically OpenAI aligned with Anthropic on the weekly usage with the caveat that OpenAI doesn't have a 5h limit.
I think this is still true provided you're not using Astra.
You have to do a lot of things in parallel.
Come on .. this is barely released and you can already make that assessment?
And no, the $200 Anthropic plan is not significantly better than the $200 OpenAI plan, it's just the same Marketing non-sense and anybody shall now rather stick to the $100 plan of both of these provider if the monthly budget is $200. Anthropic doesn't have a Luna Max equivalent, and frankly Sol 6.1 is yet to be thoroughly tested.
Yes, he was talking about safety, but IMHO they're likely already IMHO pushing the boundaries of cartel type behaviour. And they will use safety as the cover to make it happen.
I suspect we'll see serious price fixing and the DOJ do nothing about it because of the inroads these people have with the Trump regime.
Whether that survives contact with Chinese open weight models is hard to say.
Might want to hold off on canceling and continue to bleed them dry until the nerf hits
The only way is for prices to go up. Way up.
Opus 5.5 is definitely better at coding, but nothing even comes close to 6-Astra for work in 3D graphics...
A number of others have done game/3d video benchmarks but this guy is probably the most prolific.
I have played around a little bit with fixing some rigging problems and was impressed, but Opus even warned me it was bad at animations cause it can only really grab screenshots to process static content.
I've only dabbled but yes with SOTA models it is very good at animating and really most Blender tasks you can think of. Certainly if you are coming at Blender at below expert level it makes it far more accessible and fun to work with.
There are still rough edges of course. But try the official MCP out with Astra and judge for yourself.
Since Luna is so dirt cheap compared to Sol/Astra it would be nice if they could set or you could reserve some small percent like 3-5% of usage pool on codex just for Luna so if you hit usage limits you can at least still run a lot of Luna.
Fable 5.1/Opus 5.5 isn’t different, but the first cut is better quality.
Astra is a whole order of magnitude cheaper than Fable, and the Anthropic usage limits are ridiculous. Layers on layers of limits that constantly trip.
We don’t really use Sol because Astra X High is cheap. Some have mentioned regressions but we haven’t noticed any with Astra.
Astra is a pretty impressive model. Excited to try this.
Yes, benchmarks aren't real work blah blah, but the delta here is so large compared to Astra, it makes it seem like this is distilled Bel or similar.
Opus 5.5 was a gut punch and my impression is OpenAI is still reeling.
The best thing is that we benefit from these constant back and forth gut punches :)
They can release a new version every day if they wanted to. The question is whether or not the new releases provide substantial improvements or not. It's not hard to just go through the motions, bump the minor version, then make an announcement to rile up the users who don't get that none of this is standardized or regulated in any way and it's literally all made up by the company trying to sell them the product.
I just added an agent / coding agent into an email app, and doing it through `codex` and its Codex App Server couldn't have been easier, and the results are very compelling.
The open source harness, API around it, and friendliness for connecting a subscription puts Claude to shame right now.
A few more thoughts here https://housecat.com/blog/introducing-housecat-agent
GPT 6.1 Sol: https://html.non.io/lcars-gpt-6.1-sol
Opus 5.5: https://html.non.io/lcars-opus-5.5
Overall, opus executes a bit better than 6.1 sol, which surprises me. Astra has been the best model for this flow so far, so the fact that Sol missed some alignment / vision pieces here is interesting. It's not bad by any means, but I think where Opus really wins is the motion animation of the svgs / final polish (scroll down to the "customize every detail" section on the homepage, the svg animation is beautiful for that).
Still, it executed quick and was quite cheap to run.
I have to be more literal with it than with GPT 5.x, otherwise, it sometimes does something totally different than what I want.
Then Opus 5.5 caught them off guard and now they're actually releasing the correct sized model.
This is because they have really aggressively priced a larger model to compete with Opus 5.5, so their margins are much worse. Consequently, the equivalent API spend on the subscription is much less.
Back fired because of opus 5.5.
So now we get the real sol-6 as sol-6.1, and OpenAI will eat the cost to stay competitive.
This could be invalidated if sol-6.1 is the same speed as sol-6.
However, that doesn't say much. You can just run a smaller model at a larger batch size to get higher throughput but lower interactivity.
The last time a model announcement felt like a leap in capability beyond other things out there was Fable - which was promptly taken away. Sol and recently Opus 5.5 were strong because they approach that capability with a lot more efficiency and don't blabber incoherently (looking at you Opus 5.1).
Deepseek is a workhorse for those who prefer open and API usage. Other than that the model announcements all just seem like a blur and quite interchangeable but I wonder if that's just me tuning out or do others feel the same way?
My experience with agentic coding on projects I care about (because my responsibility in my firm is to care about these things, at least for now) has not changed a lot in the past few months, and I have kept up with every single model update / experimented with harness a great deal.
> I'm starting to feel people are not honest about their experiences...
1. They don't actually look at/care what the agent is producing as long as it works (not planning on maintaining/ops yet).
2. They are using 3rd party benchmarks (which is fair given how widely real-world workloads change from day-to-day, feature-to-feature, making it difficult to really know how well the models would perform).
3. They are doing greenfield work where there is no scaffolding, no existing code, no legacy code, nothing to guide the agents along. I believe in these cases, new models can possible do better from a blank slate. But in existing codebases, I feel like the agents are more likely to simply follow existing patterns and existing guidance to begin with so things are a wash and more reliant on harness and existing code hygiene.
Excited to tryout Decisions API as well.
https://artificialanalysis.ai/?models=gpt-5-6-luna-low%2Ccla...
According to this, at Max it's better and cheaper than 5.5 Medium, but worse than 5.5 High. At Medium, it's better and cheaper than 5.5 Low.
It's easy to hit those numbers in a day in an modern-enterprise context synthesizing from incoherent information in jira, slack, layers of codebases etc. Modern enterprise meaning a firm that has been serving a few strategic customers w/ "move fast and break things" since day 1
These moves all make sense when you take into account the enterprise market.
The good part is that this kind of behaviour also makes it good to find subtle bugs or debug issues that Fable/Claude just cannot get/fix even when you point it.
This is a decent win though, if it really is better. 6-sol was really no good, at least in my work.
Will see if this remedies things.
https://amphetamem.es/meme?id=the-simpsons_06_12_71&text=We%...
it's 500$ for 25x the plus usage, thats pro (max).
this implies that the old 200$ 20x pro (more) is now more like 10x the usage of plus.
they are slashing our subscriptions in half and make it "but we're more efficient!"
However, I don't know how future larger models such as the cancelled 6.1 Astra will be priced.
If the price stays high, this would indeed be quite bad for the $200 subscription..
Either way, a little ironic…
Chinese models are cheap and fast by token, but they generate oceans of thinking tokens in order to accomplish the same result GPT-6.1-Sol accomplishes in, comparatively, two drops of thinking tokens, making the Chinese models come out costlier and slower per actual task performed end-to-end.
Here's from artificialanalysis.ai, per task:
* Deepseek-v4.1-flash (max): 0.27$, 5.5 minutes, 89k tokens generated.
* GPT-6.1-Sol (medium) : 0.21$, 2.2 minutes, 8k tokens generated.
API pricing. With Sol having SUBSTANTIALLY better performance.
If OpenAI cuts alternative harness support it will be a weird day trying to figure out what to do next, it's been so clearly the best bang for your buck (imo) for a while. maybe id finally have to give smaller models a try.
anything to avoid using the dogwater codex & claude code tuis.
anyways this seems like a nice cost improvement over GPT 6 Sol and I expect this will be my new daily driver.
not saying this is the case here but it does feel a bit like wine tasting sometimes, everyone claims to be an expert that can taste a few tokens and tell you exactly what region and vineyard its from.
Fuck altruism, ammi right? lets make money, gobs of it by screwing the middle users as much as we can to push them into just two tiers: Ones that use it for recreation and others that pay through their noses.
One thing I wish was better communicated is the mileage we get for our subscriptions. I do not fully understand how much usage I get with each model and their reasoning effort on 5h and weekly limit in Codex. I am asking because I know switching to Astra would consume my 5h usage limit quite rapidly, so I avoid it. If I knew how much mileage I would get from each model and respective reasoning effort, then I would be able to plan my workflow better and know when to upgrade model for a task. In almost all cases, GPT-6 Luna (XHigh) have been enough. That's why I appreciate its discount, because its dirt cheap, yet highly capable.
In other news:
> In the coming days, we’ll also offer GPT‑6.1 Sol Ultrafast , with up to 8x faster token generation compared to its standard speed in Codex.
Guess not?
{"type":"item.completed","item":{"id":"item_0","type":"error","message":"Model metadata for `gpt-6.1-sol` not found. Defaulting to fallback metadata; this can degrade performance and cause issues."}}
Laptops use very minimal power - you don't need to worry about them. If they didn't their battery life would suck.
Edit: for context, just Steam alone has ~200million monthly active users.
I'd be curious as to how much of internet infrastructure is dedicated to gaming though.
How many DCs are devoted solely to gaming?
Yes but it adds up when you consider that just on Steam alone there are 200 million monthly active users.
An entire planet. Just Steam alone has one or two hundres million monthly active users.
Yeah? Show me the big movements against computer gaming.
The differences with AI are: 1) we are starting off (mid 2020s) from a baseline point of already being in a hopelessly shitty situation, past the 1.5C warming target; and 2) Electronics, chips, data centers etc were already a thing for a long time, but industry took _decades_ to ramp up production to pre-AI levels, and these things are used everywhere for a huge number of things. Now we're consuming electronics/data centers/water/power at an unheard-of rate, and for a single purpose (AI) with questionable benefits, besides the private interests of a handful of people.
Even then people do care about the power consumption of non-AI things. Look at the energy label on your TV or tumble drier for example.
But this is not that, the same gpus you play games with are used to run llms. How was energy consumation by gpu not a topic before llms?
> I don't think video games consume nearly as much power. A PS5's power consumption is apparently around 200W. That's not enough to run even one GPU, let alone the armada it presumably takes to run Astra.
Just Steam has 200 million monthly active users. Add Steam, PS, Xbox, and whole other devices having gpus and I'm pretty sure you at least 10x the energy consumption of all ai companies.
I dunno what you're not getting but a GPU to run games is like 200-500W. A GPU cluster to run Astra is probably more like 10kW.
Also gamers tend not to spin up dozens of other machines to also game for them.
Huge misstep releasing it.
I can handle issues much better if they are predictable even if the model makes mistakes — much more frustrating when the model is erratic
I find codex wanders off road more often and fails to see the “bigger picture” (as much as LLMs can see the bigger picture at least)
And tbh when it was first released Astral felt even worse
I’m being forced to use it right now and at the end of the day I’m making do so it’s fine, but Claude makes for a smoother experience
Personally I've taken to having a list of 3 to 4 models in default context with some ordering on which to prefer. Things like GPT 6 Luna is cheap very cheap, use it. Because otherwise the model will assume Haiku or such is the good cheap model to use.
The speed I'm having to update that document has not gone unnoticed.
When is the alleged "safety" concern satisfied? Does this mean releasing new capability to consumers is going to get a lot slower? Lower price for 6 Astra capability via this 6.1 Sol is exciting, but that is because of Astra capability not merely the low price point.
When do we get the next jump in capability? When is 6.1 Astra released?
After what DeepSeek pulled with V4.1 Flash I've given up on trying to map LLM versions to semver.
The coverage around 6.1 Astra seems deliberately playing into the dubious, recently headline "safety" narrative in a way that feels distinct. But you may be correct in which case, I would take the correction on board and maybe suggest a different alternative.
Although in theory if OpenAI was boycotted in this way the market pressure would force them to release. Then everyone moves back over there. Then Claude faces the same pressure. So even so, I think it could still work even if you have to trade off who you are boycotting from time to time.
Without more details on the credibility of the "safety" concern this seems like a totally coherent action for customers to take. We shouldn't put up with teasing.