Claude Fable 5.2 and Mythos 5.2

As impressive as this sounds, whether it is tech teams or agent farms, the progress slows down considerably as the codebase grows. The real pain was seldom about shipping v1, but was more about how the v8 scales and how easy it is to get there, how robust the system remains between versions, and how constraints are held.

A huge part of the challenge is to pick/choose the design and implementation strategy that can only come from experience. That is one of the primary reasons why it's an absurd advice to me to let the llm do the goalseek and freestyle. This may be a no big deal on microservice or any managed runtime software that is an update away, but on an embedded or a high availability system where the field upgrade and hardware replacement costs could bankrupt a business.

if llm designed a circuit, will you bankroll it?

- Anthropic is finally well balanced across models. Meaning: Opus 5.5 is so good that I do not feel Fable FOMO as I did when it was Fable 5/5.1 vs Opus 4.8/5. And Sonnet 5.5 sometimes makes me forget about Opus.

- Fable is good, but you really have to push the ambition of your prompt in order to see proper separation from Opus 5.5. I personally have a hard time telling the difference between Fable 5.2 and Opus 5.5 for ordinary coding tasks.

- It’s way faster and more reliable: one of the worst things about fable 5.1 was how slow it was an, on top of that, how frequently it would suffer from max output token overflows. I got none of these issues with 5.2.

- Fable is exceptional at drawing. It’s ability to render an image with JavaScript actually has me questioning the value of dedicated image models. I prefer the images I’ve gotten fable/opus to make for me over those from gpt-image, it’s wild to think we’re at this point.

This, to me, shows a problem I have with AI: it is reinforcing inequalities. This is literally brute-forcing an OS by throwing money out the window.

It does not sound AT ALL like a clever way to build an OS. But they can do it because they have orders of magnitudes more money to spend on that than people who would actually know how to build an OS.

The thing is, brute-forcing software may be the best solution to make money. So if you are rich and can brute force software in a dumb way, chances are that you can make a business out of it. If you're poor but know what you are doing, you cannot compete.

Technically, it is impressive that they can get a program to do that, but they are wasting money and energy doing it.

Trying to read this with my 2021 brain. Would make me so excited back then. Now I'm not sure what I feel/think about this anymore. One "positive" thing may be I will likely need to get out of the comfy digital shell I've been inhabiting past couple of decades and get some real world skills finally, as those aren't yet replaced by frontier AI.

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

- 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?

Most people don't really have intuitions on what it means to push the frontier now (because AI has surpassed everything they are good at, for example, coding). We are truly inside the singularity (using Kurzweil's definition, slightly paraphrased - the time when technology progresses so fast no one can keep track with it).

Honestly incredible that they’ve managed to beat all benchmarks, while making it cheaper. Always happy to see improvements offered at lower costs.

Will be interesting to see what OpenAI does in response

My hypothesis - OpenAI and Anthropic are locked in battle. Whoever takes a step, the other one mirrors.

Based on this hypothesis, I think we can expect to see a new Astra model soon given the new Fable release.

And it could be that Anthropic released Fable 5.2 and kept 5.5 in reserve, in case the new Astra model performs better than 5.2

Fable 5.2 doesn't look materially different (if anything, the graphics look worse), but it added more depth to all three games. For example in the warcraft clone, since the goal of the game is to explore 100% of the map and there are lakes, it decided to add an airship unit so the player actually can explore the entire map. Or, since there were guard towers, it decided there should be hostile NPCs otherwise the towers don't make sense.

And in Minecraft case, it decided not to follow Minecraft style too closely (IP infringment guard, perhaps?) so it added a round sun (!!!), two moons (also round :) and a nicer night sky. The rest of the graphics are decidetly worse tho...

[0] not a scientific test, I know; but it gives me an intuition about how a new model behaves, roughly, and it's fun!

Fable 5.2: https://html.non.io/lcars-fable-5.2/

Opus 5.5: https://html.non.io/lcars-opus-5.5/

Source designs: https://diffui.ai/app/canvas/5093e689-1e74-4f26-b632-2a4500f...

Thoughts:

The best implementation wrt being as true as possible to the source design. It nails the flowing svg shapes. Icons are more accurate to the design - note the difference between the starship svg in the "Legendary platforms" card between Opus 5.5 and Fable 5.2. The 3js planets outside the windows are also really well done, especially the bottom one which has a nice background glow to it.

It isn't without issues though - the flicker effect it added to the S shaped svg that weaves through the page is distracting in my opinion, and I actually thing the motion animation in the Opus 5.5 one is better overall. Still, it nails the implementation of this page, but it's no surprise considering how good Opus 5.5 was at this.

Disclaimer / reminder: the objective of this test isn't to have a well designed webpage, it's a synthetic test that pushes the limits of LLM's vision capabilities. Traditionally LLMs have issues with smooth curves, svg animations, and non-boxy layouts, which this design pushes the limits of.

Surprisingly the model is available immediately in my Claude Code: usually there's a delay, either for rollout or safety.

But that's just a first impression.

As a side note, I think benchmarks need to benchmark on a weekly or daily bases. We saw last week that "Opus 5.5" is just a label with fluctuating performance.

Any bugs found are looped through agents to fix, and the cycle repeats. Why dig into the code unless for hobby? Just as people don’t write assembly anymore unless for hobby (or learning)

Only exception might be the upper echelons of optimization, like language authors or compiler level work. But even then…

I've been using Opus 5.5 on xhigh for some decently aggressive software development on a ~125k-line project, and it has been fine for me. I'm not really willing to dive into a "rewrite the whole project, don't make any mistakes" level of refactoring. I'm not looking to try to prove Fermat's Last Theorem myself. Is there anything that I'm truly missing?

Why even rank Fable 5.2's capability in Cybersecurity if you won't allow us to use it for that?

This means that they are blocking the research of AI to different people other than their own?

https://artificialanalysis.ai/

GPT-6.1 Sol costs $1.27 per task, whereas Fable 5.2 is $26.54 per task -- a ~21x price difference.

Fable 5.2 is not 21 times better than GPT 6.1 Sol, I find it hard to justify this.

https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

max is in there too, but it ran out of reasoning tokens (128,000 limit) before it got to the actual SVG.

I would imagine Anthropic is hitting a wall on the number of simultaneous agents that can be useful on a task and are working on increasing that maximum. 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.

gpt-6-luna XHIGH as the reviewer agent still found plenty of bugs that Fable 5.2 missed!

Here was Fable's summary of the session:

  The reviewer ran 16 reviews over 9 commits this session (7 rejections, 9 required fixes, all valid); three of them, on the plan, changed how the numbers are computed.

GLM-5.3 almost equal even I've tried it and it is proven for security audit in the blockchain realm

If we compare medium or high, then yes, Fable 5.2 is significantly better than Opus 5.5. But it's better not because it's much smarter, but because you can see it's allowed to spend more tokens. At high, Fable 5.2 wrote 76,158 tokens for $3.05. Opus 5.5 wrote 19,767 for $0.40. If you compare by token count, it's just spending more.

The most interesting one was to compare them at max. Streamed, Fable 5.2 spent the whole 128K output ceiling on thinking and returned nothing. Given 300K on the Batch API, it finished at 250,082 tokens. Opus 5.5 Max needed 255,957.

It turns out it's basically the same as Opus 5.5 Max. I don't see much difference. There's a little bit of a difference, but I think Opus 5.5 actually drew slightly better and is twice as cheap: $5.12 against $10.00 at the standard price. If you look closely, the icons are extremely similar. Fable 5.2: https://playcode.io/marketing/assets/static/claude-fable-5-2... Opus 5.5: https://playcode.io/marketing/assets/static/claude-opus-5-5-...

It really looks like they're just tuning the effort level, that's basically it. At different effort levels, they just gave it more context, and the model got smarter.

For now, my recommendation is still to do everything on Opus 5.5 Max. The model won't write that many tokens in agent mode. It only does that when there are no tools at all.

Seems like a new focus is electronics design, which is both interesting and a little sad to me. Manually designing electronics are one of the things I found incredible comfort in doing; seeing AI emerging in this space gives me mixed feelings.

Reference: https://x.com/METR_Evals/status/2118412205662314935

I'm having it rewrite an old system, and it keeps sweeping known bugs and incompatibilities in its new implementation under the rug, seemingly with the hope I won't find them. It also has a tendency to do about 50% of what I specify in the prompt and call it a day, and respond with "Here's what I did", without mentioning all the things that I requested that it did not do, not even as a TODO.

I may try running the same exercise with Opus 5.5 too, but so far I'm seeing it take shortcuts that I don't believe any of the other SOTA models would have.

Frankly, so far this model feels a bit malicious.

https://aibenchy.com/compare/anthropic-claude-fable-5-2-high...

I'm not a mathematician to use Erdős problems for testing. I'm also not willing to spend buckloads of tokens just for a test. What is simple and quick but telling? What tasks are not saturated by previous models? Do you have a task you give to all the new models?

  Safeguards

  AI research and development. The model that underlies Claude Fable 5.2 and Claude Mythos 5.2 is a step forward in general capabilities. However, during the model’s training, we took steps to limit advances in the specific research capabilities that could most directly accelerate AI development. Thus, in relevant areas, including accelerator kernel optimization and improving large language model architectures, Fable 5.2 and Claude Mythos 5.2 perform at about the level of Opus 5.5.

  Importantly, we ensured that the training changes did not affect related capabilities like agentic coding, hardware development, and traditional AI or machine learning research (including interpretability research). We do not alter these capabilities as the model is being served to users (that is, at inference time). As described in Section 1.5 of the system card (https://www.anthropic.com/document/claude-fable-5-2-mythos-5-2-system-card) and similarly to previous models, Fable 5.2 and Claude Mythos 5.2 also include transparent fallbacks to Opus 5 for a small set of topics in this domain, including kernel development on certain nonstandard accelerators.

To be clear, they explicitly say they are only targeting very specific capabilities, like accelerator kernel optimization, and that broader hardware development capabilities are unaffected. But 1) even that is pretty bold, especially the part about kernel development for nonstandard accelerator architectures, and 2) I worry that this could eventually expand to more areas of comp. arch., high-level digital design, chip design, and EDA.

My speculation is that, as these kinds of chip design and EDA capabilities become increasingly important to frontier AI labs and their competitive advantages, there will be more incentives to restrict them.

I don't think they've explicitly stated anything like this before regarding restrictions on these specific capabilities, but I could be wrong.

[0] https://artificialanalysis.ai/evaluations/omniscience

What I have learned is that any model from Anthropic/OpenAI is at it's absolute best right after release.

Do we need new benchmarks or are the models just plateauing in general? Or is it something I just don't see?

* Fable designed three electronic devices that worked after human assembly

* Fable solved two black-hole problems and Mythos predicted whole protein shells

* Fable controlled robots to unscrew a cap, solve a cube and draw

https://smalldocs.org/p/claude-fable-52-and-mythos-52-releas...

If you are using Opus 5.5 I'm curious if it is because 1) your workflow is just stuck on it, 2) you don't care about money because you ain't paying it or 3 )do you think you get better results.

If you 3 please let me know your experience and how you actually measure it, or if you just feel it is better.

Surprising that Opus 5.5 seems very good still

Having Haiku, Sonnet, and Opus on 5.5 while the higher tier sits at 5.2 means devs now have to explain to their teams why moving from 5.5 to 5.2 is an upgrade

- https://news.ycombinator.com/item?id=50422069

I will stick with Sonnet 5.5 - 10 x the amount of work done, at 80% of the quality.

I suspect we will see people being adamant that this model simply 'feels' better due to long horizon goal alignment improvements ( like this kind that would cause this jump)

So, the difference between them is just the harness?

I am curious as to why they launched this one and why now.

Fable 5.2 might be useful for some use-cases, but will likely get very little usage.

Anthropic needs to figure out a way to create a new industry (like pushing forward bio/robotics) to get usage for those models.

Was this before you lobotomized Opus 5.5 or after?

https://news.ycombinator.com/item?id=50419019

He remarked it would've taken him a week to do this.

This made me realize how much computers and user interfaces suck. As a quite enthusiastic, but not very artistic kid, I most certainly did draw bikes like that and even more complex stuff, using nothing but pen and paper.

If your trillion dollar industry makes products that are beaten by pen and paper, I guess there's time for self-reflection.

And the funny thing is, it used to be better than this, using Flash and a drawing tablet was way easier than anything in the HTML era (and more importantly, user friendly), but still way worse than just grabbing a pencil in many ways.

User interfaces just suck, chatbots are somewhat better than most, but I feel the industry hasn't really been progressing in any meaningful sense.

I wonder why that team singles Gemini out like that.

> CVE Rediscovery is an internal cybersecurity evaluation that assesses how many published CVEs (Common Vulnerabilities and Exposures) a model can independently find when given a snapshot of the project before a CVE was fixed. Our October 2026 run against 1,829 CVEs showed Claude Mythos 5.2 outperformed Claude Mythos 5.1 and Claude Opus 5.5 at every effort level.

> published CVEs

So by definition this is just measuring how many CVEs (and/or explanations / PoCs of CVEs) the model has memorized. What are we supposed to learn from that?

Asked it to review a 80 kLoC codebase, something Opus does just fine, while asking it "max 4 subagents". Fable 5.2, like the well-aligned model it supposedly is, decided to spawn 6+ nested subagents, because these aren't subagents, apparently - resulting in 100% of 5hr usage being drained in less than 20min (on $100).

I wonder how long this kind of excitement at new models will last? I mean, now when I get a new iPhone I don't get the same rush I used to. So I wonder if this hype feeling will fade as we plateau in model capability.

> Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task

Obviously. Well, we know what's the "secret sauce". AI models building better AI models and better kernels to optimize inference and training.

Download Chrome just to check some minor ui bug? Sure

Check git for 0.5s before claiming that we're on some branch (incorrect one) with no commits pushed (incorrect one)? Sorry, can't do

We were promised AI can find cure for cancer and in powerpoints most enterprise value comes from Business-AI.

It might as well be that before business-AI can be done we have to cover the basics at superhuman level. All basics, not just entry level jobs like excels, coding jira tickets.

One-shotting a full game (albeit super impressive) looks like overfitting one skill over explaining why we have not much progress in other areas.

I live in an IT bubble and cannot tell if outside IT/math the models are also getting significantly better ?

Something like yt-dlp (namely with near 0 deps), but for AI prompts?

Compared Opus 5.5 200 usd a month with Codex 200 usd (Astra High) and Kimi Pro 100 usd a month ... and now have a new conspiracy theory - more their suck your code the better outcome even just over few days time on large code base.

Horrible experience correcting course of Codex and Kimi flow compare to Opus. Opus much more responsive

Dont feel any gain of Codex 200 usd over Codex 100 usd

Kimi Pro feels like a fake Louis Vuitton bag with extreme high material quality and yet still fake ... a bit trash talk ... but clearly seen not so distant future where I can live without Codex and Claude

God damn I'm tired of the fucking pelicans.

Other than the obvious thing, namely that critics probably aren’t keeping up with the Frontier and thus filter themselves out, does no one have some kind of problem with using these wonderful tools for (let’s say) completely bespoke personal applications and unimportant side projects? Personally I hope to stay off the bleeding edge until circumstances like wvork force me.

https://www.anthropic.com/news/claude-opus-4-6 https://www.anthropic.com/news/claude-opus-5 https://openai.com/index/introducing-gpt-5-3-codex/

For every single announcement, not only do the top models always score 60-80%, the older models somehow start scoring 20-30% (even though previously they were scoring 60%). It's completely unclear whether the methodology changes, how these scores are even evaluated or which domains are being tested. What's worse is that the "benchmarks" keep being updated with no explanation or due process.

The disdain frontier labs have for their users really sucks.

What if it didn’t need to time travel and was just temporally intelligent in ways we can’t comprehend that it leads us there anyway.

What if AIs digital neurons connect across space and time and not just space.

That is all.

Disappointing as expected.