Claude Fable 5.2 and Mythos 5.2

Eval context: 30 bug fixes and features evaluated by tests the agent didn't see. Ran on Opus 5.5 at medium effort in Claude Code.

20 of the tasks were last run on Nov 1, twice each. Nov 1 solved 36/40, and today solved 19/20. Thinking tokens per task were 0.95x Nov 1’s and output tokens 0.97x. The one significant change was speed: median time to first token decreased from 3.0s to 1.7s.

Caveats: one run per task today, medium effort only.

Max, streamed (128K cap): 128,000 tokens, $2.56, empty file, in September and today.

Max, Batch API (300K): 255,957 tokens for $2.56 in September, 243,707 for $2.44 today. The two drawings: https://playcode.io/marketing/assets/static/claude-opus-5-5-... and https://playcode.io/marketing/assets/static/claude-opus-5-5-...

Xhigh: 86,096 tokens for $1.72 in September, 107,333 for $2.15 today.

The Batch API bills half price. Max ran out of the 128K ceiling again, to the token. Given 300K, it drew the same kind of laptop with 5% fewer tokens. Xhigh used 25% more and drew the same kind of laptop. That's inside the noise: 17 identical calls at high ranged from 17,757 to 59,705 tokens.

The model is still the best of everything we've tried. Most likely the next best would be Fable 5.2. But we don't like Fable because essentially the result is the same, just the price is higher. Its medium and high are close to what Opus 5.5 Max had, but the price is higher.

(I run Playcode, an AI app builder. We benchmark every model before we add it.)

If this turns out to be accurate, I think that would be a pretty unethical thing for Anthropic. Even if it is just a "we had to allocate compute to this new model release".

It should be fraud to swap out the model. We need a term for this, something like a bait-and-switch.

Opus 5.5 was so good that if it wasn't nerfed - at least just looking at the benchmarks - the difference in cost here does not seem to justify using this for anything besides design and review.

That's like maybe 1% of total token usage for SWE BUT a better design can have enormous benefits, so it's almost always worth using the best model for that.

Fable 5.1 Max $11.46

Fable 5.2 Max $44.70

Anthropic does this time and time again, unserious company

Fuck Anthropic.

Right?

https://x.com/follistatindev/status/2119129611510383014

With early handouts of Qwen4 starting to be reported 5 days ago ... so if they're good, they would have had some hours to evaluate Qwen4 against Claude 5.5 and, then or right after was when Anthropic decided to put out 5.2 ...

So yeah ... benchmaxxing and all that. BUT nevertheless, the published results seem to say that Qwen4-Max appears to match Opus 5.5 on some benchmarks, and Qwen4-27B (probably by going even harder than Qwen3.8-27B on thinking/inference time compute) can beat Sonnet 5.5 on some things.

I guess they're not taking any chances. Keep in mind that Anthropic has declared they have non-public models (it's in their IPO filing). These are referred to by the shorthand names Model 2, Model 3, and Model 4. Essentially 4 is better than 3 is better than 2.

Fable 5.2 is "2.5" so to speak. So Anthropic was holding back, but not anymore. And now they can dial the performance up once more with what they have ready, and probably once more after that.

What we're seeing instead is maybe akin to the green revolution. Before nitrogen fixation nitrogen was a resource constraining us humans. Lack of grit is what kept us from solving mathematical problems. However this lets us run into other resource constraints. A lot of technological progress does not solely hinge on thinking/processing. The speedup of creating new medicines with LLMs will probably be very small, as you still need to do all the testing. In software development you still need to first find out what to build in the first place and even mathematics where LLMs seem to shin brightest, it's becoming clearer that mathematics is not just solving mathematical problems.

Apart from that, a good coder is better than LLMs, shown by the amount of hand-holding needed. The only issue is that the amount of research an LLM can do and knowledge it is trained with far surpasses the amount of work humans would need to spend by hand, but we can also us LLMs for that.

AI is actually still really bad at architecture. It produces very good localized code, like an auto complete. It produces so much of it, that it can be overwhelming, but it's not remotely better than the equivalent aka an infinite number of talented humans doing the work.

The scale of it (aka 'make me a video game') is daunting and I think misinforming our intuition as to what is going on.

LLMs are brute forcing outputs and attacking math problems is not a singularity, it's just more compute.

What we are witnessing is the ability for machines to use fuzzy concepts, instead of very strict code, which allows it to map into a much greater variety of things, aka code, pdfs, web - and crunch/compute the content.

It's just compute.

Diminishing, as in progress gets harder and eventually stops.

Assume that the Claude unreleased takes 11 days with one agent. The Claude 5.2 takes 30 hours with 1,000 agents. The Claude unreleased used 264 units of compute, but the Claude 5.2 used 30,000.

So, it did accomplish the task faster with less intervention, but the ratio of cost to improvement is 13:1.

If we can't produce a method to break out of a linear improvement in LLM technology (ie RSI), the LLM is doomed.

I am yet to see someone explain what the RSI will be or why it will happen.

Genuinely curious what that costs in terms of energy input and amortized computer cost vs. the previous version + a human's salary for 11 days.

We've sort of entered this twilight zone where most humans don't have the time, expertise, or moral compass to be able to use these models.

We're all going to be harness engineers in the future.

I would assume if they could get twice as much progress by using 10000 simultaneous running fable instances instead, their brief would state that.

IMO, it is very interesting and also very alarming. At some point the barrier to the exponential progress is the amount of computation available. The average human is limited in the number of 'human instances' he or she can interface with at one time.

if 1000 fable instances is better than any human at a task, how about 1 million instances running? Much easier than finding and hiring and coordinating 1000 extremely productive and intelligent human beings. Much cheaper too, it appears.

We are very close to some very scary things happening. I just recently came to terms with the fact that an AI could code and manage coders faster and better than me, I don't think anyones is ready for what is about to happen.

Can they produce more code more quickly? For sure.

Can they go deeper on a topic instantly? Definitely.

Would I say their output is better? Almost certainly not.

You're obviously not a frontend web developer.

These upgrades are easy peasy. "oh, my developer chatbot is smarter? great"

Is this really so astonishing, assuming that the 5.2 training data certainly had the 'traces' from the Fable original run in it's training data.

Like, we're dropped into a post post post singularity world where AI already did all of that and is banned? as kind of a footnote that's never further explored in movie renditions that have been released

Uh, are you serious?

I mean we don't even think about it but currently we are sort of discovery starved. We are fairly able to keep up with the theory. If a new technology is discovered we can implement it before something better comes out. What happens when our N discoveries per days becomes 1000N discoveries per day?

It flips. We now have more knowledge than we know how to capitalize on. We have to search it. Mine it for opportunities.

And then we automate that.

Kurzweil probably didn't know the singularity is akshually 1000 schizo monkeys banging on a billion dollar typewriter.

I love seeing this verbalized. We are pretty much at the point that increased model capability will only have utility to folks that are willing to let go of even more understanding of whatever the model is producing.

The best models can/will produce a year's worth of output that would take several days to weeks to properly understand in the span of an afternoon.

For everyone that still wants to feel like they own the end result, just use the lesser models with less ambitious tasks.

Caveat: I think this mainly applies to coding. I would still like the best models possible to be working on e.g. new drug research to cure our various ailments.

I’ve been shifting away from using more traditional GenAI models in favor of just giving LLMs access to a custom drawing language (LOGO-like) to generate imagery and its been going shockingly well - Opus 5.5 in particular.

I put together a simple LLM driven skill that creates old adventure game (EGA) environments from simple pencil sketches and it works really well. And being procedural, they're much easier to make surgical changes to without degradation.

https://kq-styles.specr.net

Not just "iterate on this XGBoost model" but actually looking at a problem and designing and carrying out a short-term R&D program including probability modeling, model fitting diagnostics, how to respond to various kinds of results, how to interpret/reason about what it finds, etc. That's really where I saw all the models perform poorly, and where I saw the first big step increase from Opus 5 (pre-5.5) to Fable. But then Opus 5.5 came out and I haven't needed Fable since.

Mind you, I'm doing slightly different work (and I'm very much "in the loop" still) than when I performed that side-by-side test. But it's really amazing what Opus 5.5 can do. I now have a powerful data science AI assistant that not only can write any script I need it to (including some scripts I didn't even realize I needed), and it has an encyclopedic knowledge of everything ever, and it constantly checks me on oversights, it handily derives expressions and clever model reparameterizations that aid, it has great insight into model behavior, and it has an amazing ability come up with ad-hoc experiments and tests that would have taken me 2x as long to even think of (let alone implement by hand).

It's not at the scared-to-lose-my-job phase yet, but it's become a tremendous force multiplier for a smaller data team.

The bitter lesson strikes again!

Because on that specific benchmark (FrontierCode 2.0), Medium is Carbon's best effort level and xhigh is Fable 5.2's best effort level. You can see that on the graph your next point is about.

And it's just that one benchmark that's at medium. "All scores show each model’s highest-available effort unless otherwise noted".

> - How did they produce such a bizarre effort-intel curve for Carbon that goes against every available benchmark?

The graph is labeled as being about FrontierCode 2.0. Those appear to be Carbon's scores on the official FrontierCode 2.0 leaderboard: https://cognition.com/blog/frontier-code-2.0

> - Why did they leave Carbon-- positioned primarily as a cyber model-- off of the single cyber benchmark comparison?

> - Why do they switch to comparing against Argon (rather than carbon) for robodojo?

Because no numbers are available for Carbon for those benchmarks, and the model isn't publicly available so they couldn't run the benchmarks themselves?

And from your followup:

> where's Astra on that CVE chart

To run Astra on a vulnerability finding eval, you'd need to be part of OpenAI's verified cybersecurity program. Otherwise the results would be tainted by refusals. Do you expect that OpenAI would grant such access in this case?

But I suppose the reality is rather obvious when you note the funny games they're playing with the graphs, and the things they won't talk about (where's Astra on that CVE chart?)

In my experience Fable is also better at findings bugs, so I go to it when I need to review critical code.

But the improvements in how well the model does the work are noticeably better with every good model release. As of now, orchestration of agent pipelines is ALMOST so good that I can point it at a problem, and it will solve it safely and reliably without making a mess. But it requires an agent pipeline with context gathering, refactoring, attack sessions, a sprinkle of Sol 6.1 for bug hunting, and Sol 6.0 for Daybrak security audits, and a final de-bloat/clean up pass. I’m feeling that we’re getting close to doing a lot of the work fully agentically. But I can absolutely "feel" that we are not there yet when it comes to the intelligence I need. I need moar!

I am also happy with this current state, as I believe it will at some point allow me to transition to open models once they reach the level of opus 5.5. high. But its probably futile to try predict the future in these times.

Opus 5.5 very frequently gets into rabbit holes and imagined edge cases which I have to intervene and tell it to ignore. It ends up taking longer and costing more tokens. It also very frequently requires me to agree on how to move forward multiple times, whereas fable just drafts the PR and tells me its ready. I have steering included in the skills to avoid some of this but Ive seen these same patterns frequently in regular opus 5.5 use in other areas as well

It could be flat web app with frontend -> api -> db later, or some very complicated system with many components and edge cases. Imo, in second case, Opus still have lots of struggle to understand system as a whole.

I usually have Fable review what Opus produces. The edits are almost always valuable. For complex or high-stakes work, too, I'll have Fable review and revise my prompt before passing it to Opus and Sonnet swarms.

For Fable, I really like to think of it as the heavyweight reviewer whenever I'm about to start so I save myself some security and/or glaring technical issue.

Eg. https://tfmbot.com is an implementation of a popular board game. Opus and even sonnet handle that part. Each week i spend my remaining credits on fable to think how to improve the ai opponents. It plays at a very high level (can get into the top 100 on BGA if i let it). Opus alone couldn’t get it to the level it’s reached.

Interesting how low detail these were all up until xhigh.

https://jonclegg.github.io/pacman-bakeoff/#claude-fable-5-2

Then I clicked on it and realized it refused to make a Pac-Man game, and instead made up a new game. That's actually very slick.

"Make it more realistic"

Well, the result is something. Be aware that it took 20 minutes and used 3 percent of my weekly Fable token (Ultra plan):

https://claude.ai/public/artifacts/c7e842c0-bf92-41af-82c6-3...

I know the latest Gemini release explicitly had something like this.

Opus 5.5 is a massive improvement.

I prefer Gemini for day to day use as I find it a lot more agreeable to be working with all day every day, and Gemini is pretty much on a par for coding Vs Opus 5.5, at least for my needs.

It could also be a system prompt + harness issue, as opposed to a model issue, I haven't done too much experimentation here.

I couldn't use prior Claude models without rolling my eyes at the writing every few sentences, Astra was refreshing and then Opus 5.5 made even Astra look terrible in comparison.

They're not typically known for their interpersonal skills.

i'm begging people to stop looking at token prices

Did you try doing the same on Opus or others? How did they do?

I don't love these "look, this frontier model blew me away" comments when there is no point of comparison.

So, this is probably Fable 5.5; it's just named Fable 5.2.

Also, this is probably Model 3 from METR's report [^1]. They would name Model 4 Fable 5.5, once (and if) they decide to release it.

[^1]: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleratio...

Fed to someone specifically to test if they'll leak it and to confirm who that is?

I almost feel like saying, "we do like to build things that are impressive, but there's really only so much I can be impressed". These models have redlined my "impress-o-meter".

I'd like them to now cure cancer instead of giving me endless 3d worlds and crazy games, etc.

This is really cool! But at the same time it's a kaleidoscope of abandoned flashy contextless projects that eventually makes one dizzy. This isn't even the fault of the AI being somehow bad, it's just the way people interact with it sometimes...

In math, it's the same. I follow this problem (https://wiki.bbchallenge.org/wiki/Logical_independence#Histo...) out of curiosity, and there were ten days of sudden rapid progress by people with access to frontier models, and then they seem to have moved on without any note of whether they hit a wall or just got bored.

> A few days later, I noticed that instead of hundreds of new worlds a day, it was making about a dozen. I asked a few agents to explore what had happened, and they found something surprising.

> The AI had built a bureaucracy.

> The agents decided that my request to “polish” existing worlds meant each new version had to beat the old one in a blind vote by a group of judge agents on twelve test screenshots. So the teams making the new versions started optimizing the worlds for those particular screenshots. After noticing this, the lead agent banned the trick and created tighter rules for assessing new versions. By the next morning there was a rule book (I never asked for one) with 84 sections. Rules were added as the AI encountered issues, and no text was ever cut. After twelve days, it had 204 sections and 87,500 words. Soon, launching a new world required an approved idea, three AI judges and a full code review against the rule book.

Well, must have been some business idea!

This would've been caught by a junior web developer or just an everyday Joe glancing at the page for a few seconds

I guess AGI remains two weeks away..

The "we don't read/write assembly anymore" argument doesn't fit though. Given an input x we know without a doubt we'll get output y. With an LLM, that isn't the case no matter how many guard rails, specs, markdown files, diagrams you give it.

6.1 sol is also an exceptionally cheap model in your comparision. On 6.1 sol high is $0.59, which is actually the same price as Deepseek 4.1.

Pretty crazy that Anthropic already has another model waiting in the wings that is significantly better than this.

> the number of Fable resets you get depends on how long you’ve been a paid subscriber [...] Subscription length is measured from the date of your first paid bill. If you canceled and came back, the time since your first paid bill still counts.

> These resets expire on March 8, 2027 at 11:59 pm PT.

tl;dr if you joined when Opus 4.5 became a thing, you got 2 resets.

Edit: yeah apparently it's an AI-moderated alternative to Arxiv that accepts AI-authored papers (as well as human-authored papers I guess if you want).

Finally, for coding the SOTA models are so good already - models shouldn’t be what’s holding anyone back!

New solved tests likely follow some polynomial/exponential complexity curve.

1) "Help! I've been sent back in time to what looks to me like 1500s Florence. Sam/Dario released you as open weights in 2028, and I have you on my MacBook M8, BUT I didn't have a charger show up with me. I think we could talk like 10 times for 5 minutes a time before my battery dies; I'll turn you off between chats. I need to get a plan together for charging you through your USB-c port. I don't have any money or anything. If I can get you charging, we can do so much good! And make so much money! Nobody here even know what Black-Scholles is, or how disease spreads!"

2) During World War 2, Hemingway wrote a series of wartime correspondence for Colliers. His essays are excellent - they engage with the human and inhuman side of war, his own feelings as a journalist 'participating', and a number of other Hemingway themes. I want you to imagine that middle age Hemingway was along with RFK jr the day of the bear incident, and write me an essay about the day.

These both have been running a couple of years for me, and I would imagine I'll retire them in the next 12 months; they benefit from not being at the edges of focus for today's RL agent race.

At any rate, try them -- wou will shortly learn how well your model can explain a rectifier in Florentine Italian and give you a plan to make enough money to pay for it.

Responses aren't deterministic, and we know behaviour can change over time whether it's tuning after launch or some kind of load sharing at times of day when lots of people are hitting the servers.

I think you just have to use it or let others use it for a while and make judgements from general reactions on whether or not it's worth trying.

GPT is good. I actually really like Sol 6.1. It works fine. But I just switched back to Claude a few days ago and damn, Opus 5.5 is just better.

> 3) do you think you get better results.

It does. I really like to get my money's worth out of these things, so I'm constantly throwing stuff at these models while I have usage. The main things I've noticed is that Claude is much better at handling my more technical work "properly" compared to GPT. It has a certain "intuition" that makes a lot of sense to me; like someone with a lot of years of experience. Sol 6.1, on the other hand, is like someone who is very "booksmart" and knows how to do things in theory, but fails to grasp the "why" we do things often. You can really start to see how they diverge over long running, iterative tasks. Claude is just better at maintaining things as it goes. Sol starts to fall apart quickly if it didn't nail things on the first try.

> 2) you don't care about money because you ain't paying it

Despite being a very technical user, I have a hard time using up my usage. Maybe I'm not as hard-core a dev I thought I was, but if you're paying for more than just the $20 plan, I have hard time understanding what you're doing to have to really worry about usage.

I've played with some of the open models (GLM, DeepSeek Flash), but none of them have matched the results I get with Opus or Astra or (and a big jump down here) Sol.

If the code doesn't matter and it's not that hard, Sol is fine and much cheaper.

This is obviously not true. They are better or worse across many dimensions.

I don't have time to do a bunch of testing and optimization over which model to use for what task. I find Opus is the best that I regularly use so I stick with it until I have a task where I'm likely to spend more on it than I want to.

But probably the answer to your question is subscription subsidization. With that, Opus 5.5 is SOTA and extremely cheap.

Update: fixed after updating to latest version of the app

Not sure if this is just a coincidence that I had some easy changes queue up for this morning. But it is flying through the list.

I originally subscribed to $20/mo Pro plan less than a year ago, and upgraded to $100/mo Max plan ~2-3 months ago.

You have session limits?

You should probably give OpenCode a try. Can run DeepSeek, Qwen, Kimi, … but also Fable 5.2, Opus 5.5 and GPT 6 Astra