As usual, Fable is the guardrailed model for use by anyone, and is available today across all of Anthropic’s surfaces, while Mythos is for those «pursuing beneficial applications in cybersecurity and biomedical research» and you get to apply for access.
Fable 5.2 jumps straight to the top of Artificial Analysis’ benchmarks, while acing other common benches in computer use, knowledge work, agentic tasks, and cybersecurity, to name a few.
Anthropic would rather tell you about the model’s practical and scientific results, though — as these are more tangible for advancing the frontier.
During testing, Fable wrote specifications, selected components, routed four-layer circuit boards, wrote the software and designed the enclosures for consumer electronics that worked on the first try.
It then planned and modeled four complex molecules from scratch, which Anthropic says is the «hardest job» in organic chemistry. It solved open, complex traffic and scheduling problems and then found solutions to two open problems on black hole physics.
Anthropic recommends Opus 5.5 «for most coding and knowledge tasks,» as their «strongest daily driver,» and only moving on to Fable 5.2 if beneficial.
Fable and Mythos cost $8 per million input tokens and $40 per million outputs. They are also 30% faster in their responses compared to 5.1.
Read more: Anthropic’s presentation, launch post. CNBC, The Verge, and Mashable. Discussion on Hacker News and r/Singularity.