Anthropic's Claude Fable 5.2 tops AI benchmarks at $26.50 per task

Anthropic's new flagship model Claude Fable 5.2 scored 65 points on the Artificial Analysis Intelligence Index v5, taking first place five points ahead of Gemini 4 Carbon and fifteen ahead of its predecessor Fable 5.1, released only in early September. The launch of Fable 5.2 and its restricted twin Mythos 5.2 comes shortly before Anthropic's IPO and reclaims the AI performance lead, in some cases by a wide margin.

What are Claude Fable 5.2 and Mythos 5.2?

Both models share the same foundation but differ in access, like their predecessors. Fable includes safeguards that block risky tasks in cybersecurity and biology and is available to all users. Mythos removes certain restrictions and is open only to organizations that apply through Anthropic's Verification Programs for Cybersecurity or Life Sciences.

Pricing sits at $8 per million input tokens and $40 per million output tokens, a cut from the predecessors' $10 and $50. Actual cost depends on token consumption, which varies with tasks, settings, and software environments. Anthropic positions the new models for the most demanding technical work and continues to recommend Claude Opus 5.5 for everyday use.

The models produce output up to 30 percent faster than Fable 5.1 and Mythos 5.1, Anthropic says. An optional fast mode for select organizations speeds up token output by up to a factor of 2.5, costs twice as much, and runs only on certain platforms: Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry.

How far ahead is Fable 5.2 on benchmarks?

On the current Artificial Analysis Index v5 at the highest compute tier, Fable 5.2 takes first place with 65 points. It leads Gemini 4 Carbon at 60 points on the High tier — a model Google has still not released — Claude Opus 5.5 at 59 on the Max tier, and Fable 5.1 at 50.

Fable 5.2 scores best in five of nine individual tests, according to Artificial Analysis. It hits:

• 80 percent on HLE-Diamond, expert-level academic questions
• 73 percent on Terminal-Bench-Science, testing whether an AI agent can independently run scientific workflows in a terminal
• 49 percent on FrontierCode 2.0, which evaluates whether an agent delivers useful code changes in real open-source projects

The model is weaker in agentic knowledge work and terminal-based coding. It reaches 1,785 Elo on AA-Briefcase, a benchmark for multi-step knowledge work, just 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, which measures multi-step terminal tasks in software development and security, it scores 60 percent, tied with Opus 5.5. Kimi K3.1 Pro leads at 67 percent, followed by Gemini 4 Carbon at 65 and Sonnet 5.5 at 64.

Peak quality carries a price. Per-task costs at the highest tier average $26.50, about 2.6 times Opus 5.5's $10 and roughly double Fable 5.1's $12.99. At lower High and Xhigh tiers, Fable 5.2 still scores 62 points, ahead of every other model at any tier, while consuming far fewer tokens: 98,000 output tokens per task versus 173,000 for Opus 5.5 at the Max tier.

What do Anthropic's own numbers show?

Anthropic's benchmark table places Fable 5.2 and Mythos 5.2 ahead across most categories:

• Agentic coding: FrontierCode 2.0 at 49.1 percent (Xhigh effort) versus Fable 5.1's 32.4, Opus 5.5's 45.5, GPT-6 Astra's 42.8, and Gemini 4 Carbon's 41.1 at Medium effort; FrontierSWE v2 at 69.9 percent versus Opus 5.5 at 62.3 and GPT-6 Astra at 65.5
• Business workflows: AutomationBench at 51.4 percent, behind Gemini 4 Carbon's 53.9 but ahead of Opus 5.5 at 42.5 and GPT-6 Astra at 41.4
• Agentic scientific research: Terminal-Bench Science 0.1 at 74.6 percent for Fable 5.2 and 75.3 percent for Mythos 5.2, against Fable 5.1 at 52.6 and Opus 5.5 at 58.7
• Cybersecurity: Mythos 5.2 scores 76.3 percent on CyScenarioBench and 89.6 percent on CVE Rediscovery, up from Mythos 5.1's 61.7 and 78.3
• Protein design: 85.1 percent on de novo binder design within 24 hours for Mythos 5.2, versus 79.1 for Mythos 5.1 and 82.6 for Opus 5.5 without safeguards
• Robotics: RoboDojoSim at 64.9 percent, more than double Fable 5.1's 27.6 and ahead of GPT-6 Astra at 29.8
• Computer use: OSWorld 2.1 offline set at 89.3 percent, ahead of Opus 5.5 at 87.2 and GPT-6 Astra at 79.5
• Agentic search: WANDR relaxed score at 82.1 percent versus 68.7 for Fable 5.1
• Knowledge work: Remote Labor Index at 41.3 percent with High effort, against 17.9 for Fable 5.1 and 21.3 for Opus 5.5
• Multidisciplinary reasoning: Humanity's Last Exam with tools at 74.2 percent versus 57.2 for GPT-6 Astra
• Visual charts: Chartography without tools at 73.4 percent, behind Gemini 4 Carbon at 72.5 being essentially tied, ahead of GPT-6 Astra at 71.0

Why does Mythos matter for cybersecurity?

The new models set records in cybersecurity, an area of intense debate. Mythos 5.2 finds more vulnerabilities than earlier models and is better at reproducing and fixing them, Anthropic says. Fable, the publicly available variant, blocks most security-related tasks, including penetration tests and exploit code generation. Opus 4.8 responds to blocked cybersecurity requests instead; Opus 5 responds to blocked biology requests.

Anthropic tests this with an internal benchmark called CVE Rediscovery, which measures how many published vulnerabilities from the CVE database a model can independently find when given the project's pre-fix state.

In the October 2026 run covering 1,829 CVEs, Mythos 5.2 found 10.2 percent more vulnerabilities than Opus 5.5 and 14.4 percent more than Mythos 5.1, Anthropic says. It led at every compute tier. One security firm reported that one in six confirmed bugs Mythos 5.2 found had been missed by all other frontier models tested.

Did Anthropic limit the models' capabilities?

Yes. Anthropic says it deliberately held the models back during training from improving at capabilities that could most directly accelerate AI development — so-called RSI, widely discussed as a risk. Fable 5.2 and Mythos 5.2 perform roughly on par with Opus 5.5 in areas like optimizing compute kernels for accelerator chips and improving large language model architectures, the company says.

Related capabilities — agentic coding, hardware development, and traditional AI research including interpretability research — are unaffected. Claude now writes over 90 percent of the changes merged into Anthropic's core codebase under human guidance, the company says. Anthropic also claims the model programmed an operating system from scratch in an internal test.

What did 1,000 agents do to Fermat's Last Theorem in 30 hours?

The most striking result comes from mathematics. About 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem, translating it into the Lean programming language so a computer can verify it step by step. It took 30 hours, the first completed machine verification of the theorem.

Fermat's Last Theorem states that the equation a² + b² = c² has no whole-number solutions for powers higher than squares. Nobody proved this for centuries until Andrew Wiles presented a proof over a hundred pages long in 1995.

An earlier, unpublished research version of Claude needed 11 days for the same task and relied on a purpose-built environment and guidance from a human researcher. The Fable 5.2 agents started with brief instructions, organized their collaboration on their own, and used no mathematical references, Anthropic says. The only progress metric was a list of intermediate lemmas. The model used one-third of the action tokens of the first attempt and produced a proof seven times shorter.

Anthropic ran Opus 5.5 in the same setup with 1,000 agents for comparison. Opus managed 26 percent of the formalization in 18 hours and made barely any progress after that through hour 30.

With Fable 5.2 leading every compute tier on Artificial Analysis's index while Anthropic throttles the capabilities closest to recursive self-improvement, the release sets the benchmark — and the safety posture — the company will carry into its public listing.