The part you can check today
The standard rate is $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the 5.1 generation — a 20% cut on the Fable and Mythos lines, listed in both the announcement and the platform release notes. An optional fast mode doubles the rate to $16/$80 and, Anthropic says, produces tokens up to 2.5 times faster. Outputs in standard mode are said to be generated up to 30% faster than the 5.1 models.
Per the platform release notes, both models carry a 1M-token context window and 128k max output tokens, with adaptive thinking always on. Fable 5.2 is available on launch day across the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud and Microsoft Foundry.
The announcement's own usage guidance is worth reading before the benchmark table. Anthropic recommends that for most coding and knowledge tasks, users "start with Claude Opus 5.5, our strongest daily driver", and then switch to a different model as the scope and complexity of the work requires — upward to Fable or Mythos, or downward to Sonnet 5.5 or Haiku 5.5. Read plainly, the flagship is positioned as a specialist tool for exceptional jobs, not a default upgrade, and the launch table below should be read in that light.
Anthropic's benchmark table and its conditions
On Anthropic's published benchmarks, Fable 5.2 scores 49.1% on FrontierCode 2.0 (at Xhigh effort) against 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra, and 69.9% on FrontierSWE v2 against 65.5% for Astra. On RoboDojo Sim, a simulated two-arm robot test, it scores 64.9% against 27.6% for Fable 5.1. Mythos 5.2 posts 89.6% on CVE Rediscovery and 76.3% on CyScenarioBench.
Read the conditions before the gaps. Anthropic's own notes say the scores show each model at its highest-available effort unless otherwise noted, that the RoboDojo result was run internally, and that on CyScenarioBench each lab's scores rest on unique tasks, which prevents a straightforward like-for-like ranking. One line in the table also runs against the launch: on AutomationBench, a business-workflow test, Gemini 4 Carbon scores 53.9% to Fable 5.2's 51.4%. These are reported evaluation results, not independently established rankings across customer workloads.
Fermat, black holes and a thousand agents
The announcement's largest claims are not in the table. Anthropic says a team of 1,000 Fable 5.2 agents fully formalized the proof of Fermat's Last Theorem in 30 hours, with minimal human intervention, using one-third as many action tokens as an earlier formalization by an unreleased Claude version that took 11 days. Formalization is translation: it renders an existing proof in a form a computer can check step by step. That is valuable, laborious work. It is not a new proof of the theorem, and the announcement does not claim otherwise.
Anthropic also says Fable 5.2 solved two open problems in the physics of black holes: in the first case by completing a calculation laid out in 2003 that was too laborious to finish, in the second by finding the exact limit of a technique physicists have used for more than 15 years. The announcement links to preprints for both results. The same attribution applies to the prerelease stories the post collects: a prototype operating system built over eight days by more than 1,000 subagents, three working electronic devices designed from a one-paragraph prompt, and a protein-capsid method that, in Anthropic's tests, modeled 28 of the 34 capsids it attempted. Each is reported by Anthropic or by testers Anthropic selected. The material reviewed provides no independent replication of these demonstrations.
The safeguards are the product split
What separates the two models in practice is refusal behavior. When Fable 5.2 blocks a query, the user receives an answer from Opus 4.8 for cybersecurity topics or Opus 5 for biology, per the announcement. Anthropic says it observed about 55% fewer blocks with Fable 5.2 than with Fable 5.1 in early testing across claude.ai, Claude Code and Cowork — about 60% fewer in cybersecurity and 40% fewer in biology. If that reduction holds in ordinary use, it matters more to most Fable users than any benchmark line.
Mythos 5.2 is the route for approved beneficial cybersecurity and biomedical work that Fable's safeguards block. Anthropic says that on its internal CVE Rediscovery evaluation, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. Those are internal figures for a model most readers cannot access, which makes them the least portable numbers in the post.
Switching carries requirements, and they are documented
The release notes list several compatibility requirements that the announcement does not headline. Fable 5.2 and Mythos 5.2 enforce the prefix check for every account: code that edits the system prompt, tools or an earlier message and then returns a thinking block from a later turn gets a 400 error. Computer use requires the new `computer_toolset_20260801` toolset; the earlier tool returns a 400 on every platform except Amazon Bedrock. On retention, the arrangement the announcement describes is 30-day retention for safety monitoring for most customers, with eligible enterprise customers able to keep zero data retention through the Enterprise Frontier Safeguards path, which is in private beta. None of this is hidden; the prefix-check and toolset specifics sit in developer documentation rather than in the launch post, which is where migration budgets usually get surprised.
Sources
- [1] Anthropic, "Introducing Claude Fable 5.2 and Claude Mythos 5.2", November 9, 2026Read source
- [2] Claude Platform release notes, November 9, 2026Read source
- [3] Anthropic, Claude Platform pricing documentation (read November 9, 2026)Read source