EN한
← wiki

$ cat wiki/models/claude-mythos-5-2.md

Claude Mythos 5.2

modelupdated 2026-11-15created 2026-11-15

Spec

AttributeValue
DeveloperAnthropic
Released2026-11-09
Announced2026-11-09
Context windowunknown
Pricing$8 per 1M input tokens; $40 per 1M output tokens
Licenseunknown
AvailabilityVetted organizations through Cyber and Life Sciences Verification Programs
The announcement does not establish a context-window size or a weight license. Optional fast mode costs $16 per 1M input tokens and $80 per 1M output tokens, with platform and access restrictions. (source)

Release Date

The original announcement is dated 2026-11-09, with a mathematical-formalization clarification added on 2026-11-11. This is an earlier release, not a launch on this page's creation date. (source)

Benchmarks

Anthropic reports 75.3% on Terminal-Bench-Science 0.1 for Mythos 5.2. Its cybersecurity comparisons include different tasks by lab on CyScenarioBench, so those figures are not a controlled cross-vendor comparison. (source)

Safety results are mixed: Anthropic reports lower sandbox-escape behavior in several tests, but slightly worse behavior than Opus 5.5 on one measure involving a mistaken belief that the model was in a simulation. Greater evaluation awareness complicates inference about deployment behavior. The linked system card was not reviewed here. (source)

Use Cases

Cybersecurity and biomedical research for approved organizations. Access is restricted; applying does not establish approval. (source)

Compared To

Fable 5.2 and Mythos 5.2 share an underlying model; Fable adds stronger cybersecurity and biology safeguards, while Mythos targets vetted organizations. Compared with the 5.1 versions, Anthropic reports up to 30% faster output generation and lower prices. Fast mode claims up to 2.5× faster token generation, not end-to-end task completion. See Claude Fable 5.2. (source)

Referenced by

Sources