From the source
Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, two large language models it calls its most capable to date, with the same core capabilities but different guardrails.
- Fable 5.2 blocks certain biology and cybersecurity use cases, while Mythos 5.2 has more relaxed guardrails and is available only to trusted researchers through two limited access programs.
- Anthropic tested Fable 5.2 against Fable 5.1 across 14 benchmarks, saying it beat its predecessor by more than 10% on most tests and scored 74.6% on Terminal-Bench-Science 0.1, a 22% improvement.
- The company says Fable 5.2 cut a mathematical proof workflow that took 11 days with an unreleased model to 30 hours by spinning up 1,000 agents, and that Mythos 5.2 accurately modeled 28 of 34 AAV capsids, structures that other biology-optimized models such as
Google DeepMind's AlphaFold2 cannot model because they comprise dozens of proteins.
- Fable 5.2 sends blocked biology and cybersecurity requests to Opus 4.8 or Opus 5, and Anthropic says it is 55% less likely to erroneously block safe prompts than Fable 5.1.
- The models outperform the previous generation by more than 13% on two coding benchmarks, FrontierCode 2.0 and FrontierSWE 2.0, and an Anthropic engineer had Fable 5.2 build an operating system with 1,000 agents in about eight days.
- Both are priced at $8 per million input tokens and $40 per million output tokens, 20% cheaper than the previous generation, with a high-performance mode that responds 2.5 times faster at double the price.