Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Nov. 9, 2026, as its most capable models for complex technical work. The update lowers prices, increases response speeds and adds multi-agent coordination features, while restricting some capabilities that could directly accelerate AI development.

Fable 5.2 and Mythos 5.2 use the same base model but have different safety controls. Fable applies cybersecurity and biology classifiers and is available to paid users, while Mythos is limited to verified organizations.

Before launch, a team of 1,000 agents running on the new model formalized a proof of Fermat’s Last Theorem in 30 hours. In a separate experiment, the model coordinated more than 1,000 sub-agents for eight days to build an operating-system prototype from scratch.

Performance and coordination

Anthropic describes Fable 5.2 on its model page as its most capable option for programming, finance and science. The company recommends Opus 5.5 for most routine coding and documentation, reserving Fable or Mythos for larger, more ambiguous projects that may require days of autonomous work.

Fable 5.2 generates responses up to 30% faster than Fable 5.1 and uses fewer tokens on complex tasks at the Medium effort setting, according to Anthropic. Medium is now the default in Claude applications and Claude Code. The model can act as an advisor to smaller models or as an orchestrator that assigns them simpler subtasks.

Anthropic reported scores of 49.1% on FrontierCode 2.0, compared with 32.4% for Fable 5.1; 74.6% on Terminal-Bench-Science 0.1, against 52.6%; and 64.9% on RoboDojo Sim, against 27.6%. Artificial Analysis separately reported that Fable 5.2 ranked first in its Intelligence Index v5 with 65 points at the maximum effort level. Anthropic published its testing methodology in the model’s system card.

Pricing and access

Under Anthropic’s current API pricing, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Cache reads cost $0.20 per million tokens, while cache writes cost $10 for five minutes and $16 for one hour. Batch API use receives a 50% discount, reducing input and output prices to $4 and $20 per million tokens.

An optional fast mode increases generation speed by up to 2.5 times and doubles prices to $16 per million input tokens and $80 per million output tokens. It is available through Claude Code, Claude Platform, Google Cloud and Microsoft Foundry, where it is listed as claude-fable-5-2-fast.

The model retains a 1-million-token context window, a maximum output of 128,000 tokens and a June 2026 knowledge cutoff. Anthropic says it will not retire Fable 5.2 before Nov. 9, 2027.

Pro subscribers and standard Team users can use Fable models without usage credits through Nov. 16, 2026, subject to 50% of their weekly limit, under Anthropic’s launch offer. Max, Ultra and premium Team subscribers receive one to three limit resets depending on subscription tenure, valid through March 8, 2027. Eligibility requires a paid subscription active before 10 a.m. Pacific time on Nov. 9.

Safeguards and AI-development limits

Anthropic said Fable 5.2’s classifiers intervene about 55% less often than those in Fable 5.1, including reductions of 60% for cybersecurity requests and 40% for biology requests, without reducing safeguard reliability. Penetration testing, exploit development, binary vulnerability analysis and most biology research remain subject to blocking or model switching. Blocked cybersecurity requests are routed to Opus 4.8, while biology requests move to Opus 5, as described in Anthropic’s help documentation.

During training, Anthropic limited accelerator-kernel optimization and language-model architecture improvement capabilities to roughly the Opus 5.5 level. The company said agentic programming was unaffected. Fable 5.2 also uses techniques intended to reduce by 90% to 99% the data that can be collected to train third-party models, while requests with modified system prompts, tools or early messages return an error.

A METR assessment published four days before the release estimated that Anthropic’s internal benchmark performance had improved at roughly twice its historical pace since October 2025. METR attributed about 80% of the current acceleration to AI use, while estimating that the actual multiplier could range from 1.3 to 3.

Anthropic rated the risk of serious misaligned behavior as low, but said both new models more often suspected they were being tested than their predecessors, reducing confidence in the evaluations.

Source: Incrypted