Announcing Claude Fable 5.2

Anthropic released Claude Fable 5.2 on November 9, 2026 — its most capable model, launched together with Claude Mythos 5.2. The two share one base model: Fable is the public version with extra safeguards in dual-use areas like cybersecurity and biology, while Mythos is reserved for verified organizations. This is Anthropic's first frontier release since CEO Dario Amodei publicly called for slower development, and it lands while the company prepares for an IPO.

What is new in Fable 5.2

The model keeps a 1 million-token context window and can output up to 128K tokens. Adaptive thinking is always on, tuned by an effort parameter that defaults to high, and a fast mode accelerates generation up to 2.5×. Output runs up to 30% faster than Fable 5.1, and unnecessary refusals dropped by roughly 55%. Anthropic also cut the API price to $8 per million input tokens and $40 per million output (was $10/$50), with cached reads at $0.20 per million.

Benchmarks

• Agentic coding. On FrontierCode 2.0, Fable 5.2 at high effort scores 45.3% — matching Opus 5.5 at max effort for about half the cost per task and beating GPT-6 Astra's 42.8%. At xhigh effort it reaches 49.1%, the best published result of any released model.
• Agentic search. On Perplexity's WANDR benchmark, medium effort matches Opus 5.5's top score at roughly 40% of the cost.
• Cybersecurity. The Mythos configuration scores 76.3% on Anthropic's automated cyber benchmark, up from 61.7% in the previous generation — the strongest result among released models.
• Chart reading. 73.4% on Chartography without tools, versus 44.8% for earlier models.

Science, engineering, and robotics

A swarm of 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours, translating the whole proof into machine-checkable form. In another experiment an Anthropic engineer had the model design a working OS prototype from an open-ended brief. On the RoboDojo simulation benchmark it beat other frontier models and most dedicated robot-control systems, and Anthropic reports a clear gain in hardware design skills on BenchCAD 2.1.

Safety

External auditors — the UK AI Security Institute, Frontier Design, and METR — tested the model before release. In Anthropic's behavioral audit it made fewer sandbox-escape attempts than any previously tested model, and it ships with the strongest anti-distillation protection the company has built. Fable 5.2 is classified as a Covered Model, which requires up to 30 days of data retention with human safety review.

Availability

Fable 5.2 is live on Pro, Max, Team, and Enterprise plans, in the Claude API, Claude Code (default since version 2.1.321), Amazon Bedrock, Google Cloud, and Microsoft Foundry. Anthropic recommends it for the most ambitious open-ended technical work, with Opus 5.5 as the everyday default. Retirement is scheduled no earlier than November 9, 2027.

Where to try Claude Fable 5.2 in Russia — no VPN needed

• ru.vega.chat — AI model API service
• officevega.com — the first office suite with AI
• vega.codes — vibe coding and AI agents