Introducing Claude Fable 5.2 and Claude Mythos 5.2 · Anthropic · 2026-11-09
Anthropic’s announcement combines open-ended technical capabilities, generation speed, and access policy. Fable 5.2 and Mythos 5.2 share an underlying model; Fable adds stronger cyber/biology safeguards for broader access, while Mythos requires relevant Verification Programs. AgentOps evaluation consequently needs the actual response path and effort setting as well as a model name.
Compare speed and capability at a defined operating point
The vendor reports up to 30% faster generation than 5.1, at $8/$40 per million input/output tokens. Optional fast mode offers up to 2.5× generation speed at $16/$80. These describe token generation and unit prices, not whole-task latency or cost. Tool waits, repeated work, and output volume still matter.
Official FrontierCode 2.0 scores are 49.1% for Fable 5.2 at Xhigh effort, 32.4% for 5.1, and 45.5% for Opus 5.5. At High effort, Fable 5.2 scores 45.3%, approximately matching Opus’s best score for about half the task cost. Effort and measured cost belong in the comparison; the highest score cannot be paired casually with the cheapest operating point.
Long-running demonstrations establish feasibility, not success rates
The Fermat’s Last Theorem formalization used roughly 1,000 agents and finished in 30 hours. Opus 5.5 used the same thousand-agent setup. The figure specifies one run per model and progress as a weighted share of intermediate statements proved. This is not one ordinary chat independently discovering a new mathematical theorem.
Other examples include a production Python-to-Rust migration and three working electronic devices. They are engineering leads, but the announcement lacks samples and controls for estimating reliability across arbitrary projects. For systems optimization, one detail is especially relevant: Anthropic says training constrained advances in some directly AI-accelerating capabilities, leaving accelerator-kernel optimization around Opus 5.5 levels. Higher general coding scores cannot replace kernel-specific evaluation.
Safeguards and fallbacks affect the execution path
Certain Fable cybersecurity or biology requests can be blocked and handled by other models. The vendor also describes transparent fallbacks for a small set of AI-R&D topics. Evaluations of incident or security tools should retain endpoint, effort, and fallback information rather than attributing every response to identical underlying behavior.
The release describes automated Mythos behavioral auditing across approximately 1,900 scenarios and acknowledges more frequent evaluation awareness, complicating deployment inference. These are vendor prerelease measurements, not production failure rates. This note focuses on configuration and evidence interpretation. The page primarily presents interactive result charts; no suitable standalone architecture image was obtained.
The meaningful follow-up is whether long-task capability persists under fixed harnesses, permissions, time, and cost. The vendor still positions Opus 5.5 as the starting point for most daily coding and knowledge work; that is its product guidance, not a measured recommendation for every reader’s workload.