Claude Fable 5.2 and Mythos 5.2 cement Anthropic's AI lead just before its IPO

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its most capable models to date, timed just before the company's IPO. The two share the same underlying model but differ in who can use them. Fable 5.2 ships with safeguards that block risky cybersecurity and biology tasks and is available to all users. Mythos 5.2 lifts certain of those restrictions and is open only to organizations that apply through Anthropic's Verification Programs for Cybersecurity or Life Sciences. Opus 5.5 remains the company's recommended model for everyday use.
Pricing is $8 per million input tokens and $40 per million output tokens, down on paper from the $10 and $50 charged for the 5.1 generation. Actual spend depends on token consumption, which varies with task, settings and software environment. Anthropic says outputs are up to 30 percent faster than Fable 5.1 and Mythos 5.1. An optional fast mode, available to select organizations on Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry, raises token throughput by up to 2.5 times at double the price.
On Artificial Analysis's Intelligence Index v5 at the highest compute tier, Fable 5.2 takes first place with 65 points. That is five ahead of Gemini 4 Carbon at 60 (High tier), which Google has not yet released, six ahead of Claude Opus 5.5 at 59 (Max tier), and fifteen ahead of Fable 5.1, which shipped only in early September. Fable 5.2 wins five of nine individual tests, including 80 percent on HLE-Diamond (expert-level academic questions), 73 percent on Terminal-Bench-Science (an agent independently running scientific workflows in a terminal) and 49 percent on FrontierCode 2.0, which asks whether an agent can land useful code changes in real open-source projects.
Its weaker spots are agentic knowledge work and terminal-based coding. On AA-Briefcase, a multi-step knowledge work benchmark, it scores 1,785 Elo, behind Claude Sonnet 5.5 at 1,813 and within the confidence interval of Opus 5.5 at 1,800. On Terminal-Bench 4.0 it reaches 60 percent, tied with Opus 5.5 but behind Kimi K3.1 Pro at 67, Gemini 4 Carbon at 65 and Sonnet 5.5 at 64. Peak quality is expensive: per-task cost at the top tier averages $26.50, roughly 2.6 times Opus 5.5's $10 and about double Fable 5.1's $12.99. At the lower High and Xhigh tiers Fable 5.2 still scores 62 points, ahead of every other model at any tier per Artificial Analysis, while using far fewer tokens: 98,000 output tokens per task versus 173,000 for Opus 5.5 at Max. The cost-efficient tier may matter more to product teams than the headline score, since it suggests near-frontier results without top-tier spend.
Mythos 5.2 sets the security numbers. On Anthropic's internal CVE Rediscovery test, which gives a model a project's pre-fix state and asks it to find published flaws, an October 2026 run over 1,829 CVEs found 10.2 percent more vulnerabilities than Opus 5.5 and 14.4 percent more than Mythos 5.1, leading at every compute tier. Anthropic says one security firm reported that one in six confirmed bugs Mythos 5.2 found had been missed by all other frontier models tested. Because the public Fable variant blocks most security work, including penetration tests and exploit code generation, blocked cybersecurity requests are routed to Opus 4.8 and blocked biology requests to Opus 5. That split means defensive security teams outside the verification program cannot use the stronger capability directly.
Anthropic also says it deliberately limited, during training, the models' gains on capabilities that could most directly accelerate AI development, the recursive self-improvement (RSI) worry now widely discussed. On optimizing compute kernels for accelerator chips and improving large language model architectures, the new models perform roughly on par with Opus 5.5. Related areas such as agentic coding, hardware development and interpretability research are unaffected, the company says. Claude now writes more than 90 percent of the changes merged into Anthropic's core codebase under human guidance, and the company says the model programmed an operating system from scratch in an internal test. This is a notable data point for anyone assessing how much frontier development is already model-assisted; a deliberate throttle on RSI-adjacent work while coding output stays high is, in my reading, a containment strategy rather than a capability gap. How such a throttle is measured and enforced is not something the release details.
The most eye-catching demonstration is mathematical. About 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in Lean, the language that lets a computer verify a proof step by step, in 30 hours. The theorem holds that a^n + b^n = c^n has no positive whole-number solutions for n greater than 2; Andrew Wiles presented a proof exceeding a hundred pages in 1995, and machine verification has now been completed for the first time. An earlier unpublished research version took 11 days with a purpose-built environment and a human researcher's guidance. The agents reportedly started from brief instructions, organized their collaboration themselves and used no mathematical references, tracking progress only through a list of intermediate lemmas, and finished with one-third the action tokens and a proof seven times shorter. A matched run of Opus 5.5 with the same agent count reached just 26 percent of formalization in 18 hours and barely progressed through hour 30.
Why it matters: teams selecting models now face a three-way split rather than a single flagship: a cost-efficient mid-tier, an expensive peak tier, and a gated unrestricted variant for vetted security and biology organizations. The lower-tier numbers and token savings look like the most practical lever for product teams, though the per-task cost at the top tier suggests reserving it for work that truly needs it. Security and life-science firms in verification programs may get materially stronger vulnerability discovery, while the rest must route sensitive work to older models, which in my inference could widen capability gaps between vetted and unvetted users.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Anthropic is releasing Claude Fable 5.2 and Mythos 5.2, its most powerful models to date, in two versions: one freely available and one reserved for verified organizations. The models solve open physics problems, show stronger cybersecurity capabilities, design functioning operating systems, and write more than 90 percent of the changes merged into Anthropic's core code, according to the company. Anthropic says it has deliberately curbed any capabilities that would let the AI accelerate its own development. The art