Anthropic's Fable 5.2 and Mythos 5.2 top the benchmarks. Here's what's new, how the two models differ, and where they fit against Opus 5.5.
What are Claude Fable 5.2 and Mythos 5.2? #
Fable 5.2 and Mythos 5.2 are Anthropic’s newest models, released the same day and built from the same underlying system. Mythos 5.2 is the full-strength version, restricted to cybersecurity testers and research partners. Fable 5.2 is the version available to everyone else: the same core model with additional safety guardrails layered on top. Both beat Claude Opus 5.5 on most published benchmarks, and Fable 5.2 is now available through the Claude apps, web interface, and API for anyone on a paid plan.
TL;DR #
- Fable and Mythos are the same base model with Mythos reserved for cybersecurity testing and Fable released publicly with extra guardrails.
- Anthropic doesn’t want Fable to be your daily driver : the company explicitly recommends starting with Opus 5.5 for most coding and knowledge work, and treating Fable as an advisor or orchestrator for harder problems.
- Benchmark gains are real but uneven , with roughly 5 to 15 percent improvement on agentic coding and business workflows, and bigger jumps (around 15 to 20 percent) on agentic scientific research.
- Anthropic says it deliberately capped certain capabilities during training , specifically around accelerator kernel optimization and improving language model architectures, rather than letting the model train to its full potential in those areas.
- Fable 5.2 is expensive to run at full reasoning power , with Artificial Analysis clocking an average task cost around $26.54 at max settings, though extra-high reasoning gets similar results for noticeably less.
- API pricing dropped slightly from Fable 5.1 , from $10/$50 per million input/output tokens to $8/$40, undercutting GPT-6.1’s roughly $10/$50 pricing while scoring higher on most benchmarks.
- Real-world demos show strong agentic coding and hardware design skills , including an internal Anthropic test where the model coordinated over 1,000 sub-agents across eight days to build a working operating system from an empty folder.
pass the audit.
How does Fable 5.2 differ from Mythos 5.2? #
The two names describe one model split by access level. Mythos 5.2 is the unrestricted version, available only to a narrow group of cybersecurity testers, because its raw capability in that domain is considered too risky for open release. Fable 5.2 is what the public gets: effectively the same model with more guardrails around sensitive areas like cybersecurity. For everyday users, “Fable 5.2” is the one that matters, since Mythos isn’t accessible outside that testing program.
How does Fable 5.2 compare to Opus 5.5 on benchmarks? #
Across most published evaluations, Fable 5.2 beats Opus 5.5, but the margin varies by task type. On the Frontier Code 2.0 agentic coding benchmark, Opus 5.5 scored 45.5 percent against Fable 5.2’s 49.1 percent, a roughly 10 percent relative improvement. On Frontier SWE v2, Opus scored 62.3 and Fable scored 69.9, another step up but not a leap. Business workflow benchmarks showed Gemini 4 Carbon actually edging out Fable by a couple of points, while one knowledge-work benchmark had Opus 5.5 narrowly ahead of Fable too.
The bigger jumps show up in harder, more specialized domains. Agentic scientific research went from 58.7 percent on Opus 5.5 to 74.6 percent on Fable 5.2, a substantial gain compared to everyday coding tasks. Fable also set new highs on cybersecurity, protein design, agentic search, computer use, robotics, and multi-disciplinary reasoning benchmarks. On Artificial Analysis’s combined intelligence index, which aggregates many benchmarks into one score, Fable 5.2 currently ranks as the top model overall, ahead of Gemini 4 Carbon, which hasn’t had a full public release yet.
One detail worth noting: more reasoning effort doesn’t always mean a better score. On Frontier Code 2.0, Fable 5.2 actually scored higher at “extra high” reasoning than at “max,” suggesting the model hits diminishing returns, and sometimes gets worse, when pushed to think longer. That matters for cost, since max reasoning mode is also the most expensive setting to run.
Is Fable 5.2 worth using over Opus 5.5? #
For most day-to-day coding and knowledge work, Anthropic itself says no. The company’s own guidance recommends Opus 5.5 as the default “daily driver” and suggests reaching for Fable only for the most ambitious, open-ended technical problems: theoretical physics, protein folding, mathematical formalization, hardware and robotics design, and similar frontier research tasks. Anthropic also suggests using Fable in an advisory or orchestrator role, where it directs smaller models like Haiku or Sonnet to handle routine subtasks rather than doing everything itself. mission-criticalenterprise applications.
911 Dispatch Console Harbor County Emergency Services
Cost reinforces that positioning. Artificial Analysis measured an average task cost of roughly $26.54 for Fable 5.2 at max reasoning settings, making it the most expensive model in its comparison set. API pricing for Fable 5.2 runs $8 per million input tokens and $40 per million output tokens, double the per-token cost of Opus 5.5, though cheaper than the previous Fable 5.1 pricing of $10/$50. It’s also slightly cheaper than GPT-6.1’s roughly $10/$50 pricing while scoring higher on several benchmarks. For routine tasks, that price difference adds up fast without a corresponding jump in usefulness.
Why did Anthropic reportedly limit the model’s capabilities? #
One notable claim circulating around this release is that Anthropic intentionally constrained Fable 5.2 and Mythos 5.2 during training in specific areas, namely accelerator kernel optimization and improving large language model architectures, rather than letting the model reach its full potential there. The stated reasoning is that unrestricted capability in those domains could accelerate AI development itself, since a model that’s extremely good at optimizing training code or model architecture starts to resemble a tool for improving future AI systems with less human involvement.
This ties into a broader industry conversation about AI systems contributing to their own research pipelines, an idea sometimes called recursive self-improvement. Reports referencing independent evaluation organizations have suggested that AI-assisted research has sped up meaningfully over the past year, and that labs are now making deliberate choices about which capabilities to expose publicly versus hold back. Whether other major labs adopt a similar approach to capability limits remains an open question, and it’s one worth watching as future frontier models ship.
What can Fable 5.2 actually do in practice? #
Anthropic’s own examples lean into hard, resource-intensive tasks. In one cited experiment, an Anthropic engineer gave Fable 5.2 an empty folder and the goal of designing a prototype operating system. The model coordinated more than 1,000 sub-agents over eight days, writing, testing, and reviewing code, and produced a working OS that boots on a laptop, complete with its own kernel, disk and keyboard drivers, networking, file storage, a windowing system, a login screen and installer, and built-in cloud functionality. Anthropic also cites the model using open-source protein-folding tools to find a new way to predict capsid structures, and using a single paragraph prompt to design and help produce three working electronic devices, including circuit boards, 3D-printed enclosures, and firmware.
Independent testing by creators covering the release showed similarly strong results in agentic coding and simulation work: generating playable games with correct mechanics on the first attempt, building SVG-based web graphics and landing pages, running particle and cloth simulations, and designing real hardware builds (microcontroller boards, wiring diagrams, 3D-printable cases) from a single prompt. Reported weak points were more cosmetic than functional, with some testers noting that Opus 5.5 still produced better visual game design even when Fable 5.2 nailed the underlying mechanics faster.
Frequently Asked Questions #
Is Fable 5.2 the same as Mythos 5.2?
They share the same underlying model. Mythos 5.2 is the unrestricted version limited to cybersecurity testers, while Fable 5.2 is the publicly available version with additional guardrails, particularly around cybersecurity-related capabilities.
How much does Fable 5.2 cost to use?
Through the API, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. That’s double Opus 5.5’s per-token price but cheaper than the previous Fable 5.1 pricing of $10/$50. Average task cost at maximum reasoning settings has been measured around $26.54 by Artificial Analysis, making it one of the pricier models to run at full power.
mission-criticalenterprise applications.
✓ HIPAA
✓ SOC 2
✓ Self-hosted
Do I need a paid Claude plan to use Fable 5.2?
Yes. Fable 5.2 is not available on the free tier of Claude. It’s accessible on paid plans starting with Claude Pro, as well as Max, Ultra, and other paid tiers, and through the API.
Should I switch from Opus 5.5 to Fable 5.2 for everyday coding?
Anthropic’s own recommendation is no. Opus 5.5 remains the suggested default for most coding and knowledge work. Fable 5.2 is positioned for the hardest, most open-ended technical problems, like scientific research, protein folding, and hardware design, where its benchmark gains over Opus 5.5 are largest.
What’s new about Fable 5.2 compared to previous Claude models?
Beyond benchmark improvements in agentic coding, scientific research, and hardware design, Fable 5.2 reportedly had certain capabilities deliberately limited during training, specifically around accelerator kernel optimization and improving model architectures, a departure from the usual approach of training each new model to its maximum measurable capability across the board.