alt Hacker NewsClaude Fable 5.2 and Mythos 5.2
Comments
> Our first formalization of Fermat’s Last Theorem, performed by an unreleased research version of Claude, took 11 days. It relied on a purpose-built harness along with guidance from a human researcher. By contrast, a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours, with minimal human intervention. Fable 5.2 started from minimal prompts, built its own ways to coordinate, and used no mathematical references, only a list of intermediate theorem statements to score progress. It needed one-third of the action tokens as our first formalization and produced a proof seven times shorter. As can be seen in the figure below, Opus 5.5 (with the same 1,000-agent setup) completed only 26% of the formalization in 18 hours, and barely made any more progress after 30.
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).
I’ve had a chance to use fable 5.2. Some observations:
- 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.
Pelicans for thinking levels low, medium, high, and xhigh:
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.
- Why are they stacking Fable xHigh against Carbon Med?
- 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?
Setting aside the comments about nerfing, do normal mortals out there have workflows that really require Fable/Mythos and don't just work fine on Opus 5.5?
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?
Does this coincide with the claimed nerfing of Opus 5.5? https://news.ycombinator.com/item?id=50422069
We've been early testing Fable 5.2, and have been surprised by its coding ability. It was able to work with an esoteric 600k line fortran codebase to speed up some computational material science work for us by 36 times (!). Given that this could be done in 2 days, we've decided to open-source the results https://github.com/discovered-materials/dm_qe
Maybe it's just me, but I found Claude's writing style terrible compared to GPT or Gemini. It's "smart", but the way it communicates seems somehow off compared to other models.
There were more than a few "leaks" saying it would be called Fable 5.5, so I'm curious if a) the leaks were just engagementbait or b) Anthropic changed it, despite the fact that Fable 5.2 being better than Opus 5.5 is confusing.
Surprisingly the model is available immediately in my Claude Code: usually there's a delay, either for rollout or safety.
“Testers have found that Fable 5.2 and Mythos 5.2 communicate more clearly than past Fable and Mythos models. They generate outputs up to 30% faster than Claude Fable 5.1 and Claude Mythos 5.1, and they’re available at a lower price: $8 per million input tokens and $40 per million output tokens.”
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
Ethan Mollick's take and early testing with Fable 5.2: https://www.oneusefulthing.org/p/fable-52-and-jagged-agi
Image to html comparisons of Fable 5.2 and Opus 5.5 (the previous SOTA for this test):
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.
> Claude Mythos 5.2 is our most capable model for cybersecurity engineering. It catches more vulnerabilities than our previous models, and is better at reproducing and fixing them (note that Claude Fable 5.2’s safeguards will block most cybersecurity-related tasks).
Why even rank Fable 5.2's capability in Cybersecurity if you won't allow us to use it for that?
https://artificialanalysis.ai/#intelligence-efficiency-tabs this is the chart that matters for most of poor developers and companies. Pareto for a given intelligence index and looks like only openai is across the quadrant with muse at the bottom. They have really cooked efficient models!
Looks like some plans get extra fable usage: https://support.claude.com/en/articles/17506494-claude-fable...
Looking at Artificial Analysis index, the price is absurd to me.
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.
Here's a 20 page SmallDoc presentation on the most interesting aspects, including:
* 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...
Every time there's a flagship release like that, I wonder what simple "benchmark task" could I give it to see the the capabilities change in the most obvious way, instead of trusting the official blog post.
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?
I can't understand why people continue using opus 5.5. Significantly more expensive than other models that produce exact same results (for coding results only)
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.
Yet Opus 5.5 is only a few % under all of these benchmarks (cept the robotics and labor ones)
Do we need new benchmarks or are the models just plateauing in general? Or is it something I just don't see?
Nerfing Opus 5.5 was obvious signal that something new is coming and here we are. I appreciate the resets and lower API price though. Please add selective context compaction, it would probably take an hour max for you to implement it.
https://www.youtube.com/watch?v=vdYynH177h0
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.
I've very interested in the improvements to robotics control capabilities. This model was able to outperform specialized robotics models. It's the bitter lesson coming for robotics.
Is this the Model 4 internal name?
Reference: https://x.com/METR_Evals/status/2118412205662314935
Those theoretical physics papers look like the real thing (i.e. not mathematical physics like the OpenAI Yang-Mills result). It looks like AI is finally branching out of math.
This part is mega concerning to me as someone who researches AI for hardware accelerator design and EDA:
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.
Its cheaper at $8 per million input tokens and $40 per million output tokens which is also pretty sweet.
> To help ensure safety and alignment keep step with growing model capabilities, we ran extensive prerelease safety testing on the model that underlies Fable 5.2 and Mythos 5.2.
So, the difference between them is just the harness?
Doubling Remote Labor Index (RLI) - at 41.3% compared to around 20% for the other models - I wonder what kinds of new agentic work this is going to unlock.
On FrontierCode 2.0 Fable 5.2 xhigh is compared against Gemini 4 Carbon medium... Is there a reason for that?
They're adding 1 to 3 resets for Fable only for all paid accounts. My work's Claude Max got 2 resets. My personal $20 plan got 1 reset.
It's shocking to me that when I check threads for new frontier model releases on HN, it's full of people complaining. I, for one, am happy to see the frontier being advanced. These tools are life changing, and I'm extremely thankful for them.
why would xhigh -> max increase the cost but lowers the score on frontier 2.0 bench?
Context window maxes out at 200k in Claude Code instead of 1M on my $200 ultra plan? is this intended?
My first impression with Fable 5.2 -- it is faster than I expected. Asked it to do a large review and it responded with a good overview super fast.
Gemini carbon low costs more than medium! Weird results on that chart, care to explain?
Assuming no implicit phacking a 15 point jump on tbench scientific is a big deal.
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)
>> Fable 5.2 and Mythos 5.2 are a step above Opus 5.5 in almost all the capabilities we evaluated.
Was this before you lobotomized Opus 5.5 or after?
FrontierCode 2.0 is such a poor benchmark for Gemini since it's the only model forced to use Devin CLI while other models get to use their preferred harness.
I wonder why that team singles Gemini out like that.
The "CVE Rediscovery" benchmark chart feels... totally meaningless. Quoting the description:
> 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?
I'm not sure whether the claimed jump in capabilities makes it worth the doubling in input/output prices.
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).
Yay! New model day. The most likely time the model isn't nerfed and we get a glimpse at the latest capability.
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.
Just read the other thread about opus being nerfed and thought "cue the new release"... paha
> 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.
Obviously. Well, we know what's the "secret sauce". AI models building better AI models and better kernels to optimize inference and training.
Maybe at last it will learn to check git state right before claiming obsolete and incorrect info.
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
So does it cure cancer now?
We'll see what my bank account has to say about this:
> Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task
I didn't realize that Opus 5.5 was more capable than Fable 5.1 in virtually every category.
Damn people were right about 5.5 being nerfed in preparation for a version release. This is not at all a good look.
- https://news.ycombinator.com/item?id=50422069