Key Points
- Anthropic launched Fable 5.2 and Mythos 5.2, saying both beat Opus 5.5 across most capabilities it evaluated.
- The two models share one base; Fable ships with tighter safeguards, Mythos is gated behind verification programs.
- Anthropic says Fable 5.2 matched Opus 5.5 coding results at roughly half the cost per task.
The latest:
Two new Anthropic models, Fable 5.2 and Mythos 5.2, outperform Opus 5.5 on most capabilities the company evaluated, and generate outputs up to 30% faster than the 5.1 releases. Anthropic said both models scored at or above its recent models on most alignment measures. Mythos is available only on request, through verification programs for cybersecurity and biomedical work.
Details:
- Coding: In preliminary testing on FrontierCode 2.0, Fable 5.2 running at High effort matched the score Opus 5.5 posted at Max effort, according to Anthropic, at roughly half the cost per task. The company framed the result as an efficiency gain rather than a ceiling increase.
- Math: Anthropic said 1,000 Fable 5.2 agents produced a fully formalized proof of Fermat’s Last Theorem in 30 hours. It said Opus 5.5 reached 26% of the same task over 18 hours, a comparison the company presented as a measure of sustained multi-agent work.
- Biology: Mythos 5.2 arrived at a new method for predicting the structure of protein capsids used to deliver gene therapies, Anthropic said, and accurately modeled 28 of 34 capsids. The company did not say whether the method has been reviewed outside Anthropic.
- Robotics: On the RoboDojo benchmark, which uses a simulated two-armed robot, Fable 5.2 outscored the other models tested and most dedicated robotic control systems, according to the company. Anthropic did not detail how the simulated environment maps onto physical hardware.
- Speed: Both models generate outputs up to 30% faster than the 5.1 generation, Anthropic said, a gain it presented alongside the capability claims rather than as a trade-off against them.
- Safety split: The two releases are built on the same underlying model. Fable carries the stricter set of safeguards and is broadly available, while Mythos is released only to applicants cleared through Anthropic’s cybersecurity and biomedical verification programs.
- Alignment: Anthropic said the two models performed comparably to or better than its recent models on most alignment measures. The company did not break out which measures they did not improve on.
- Outside testing: Both models were tested before launch by external evaluators, including the AI Security Institute and METR, Anthropic said. The company did not publish those evaluators’ conclusions alongside the launch.
- Defaults: Medium remains the default effort level in the Claude apps and Claude Code. Anthropic recommends starting most tasks on Opus 5.5, despite the benchmark claims for the new pair.
Between the lines:
The headline claim and the usage guidance point in different directions: Anthropic says the new models beat Opus 5.5 on most evaluated capabilities, yet still recommends Opus 5.5 as the starting point for most tasks. The sharpest results cited — half-cost coding parity, a formalized Fermat proof — come from preliminary testing and large agent swarms, settings further from ordinary use than the default Medium effort level most users will encounter.
What’s next
Cybersecurity firms and biomedical researchers can apply for Mythos access through the two verification programs. Watch for published findings from the AI Security Institute and METR, and for FrontierCode 2.0 results beyond the preliminary round.
