Anthropic Releases Claude Fable and Mythos 5.2: Token Rates Drop 20% While Reasoning Settings Shift Task Costs Nearly Sixfold
Frontier AI models · AI coding · Enterprise AI
Anthropic released Claude Fable 5.2 and Mythos 5.2 on November 9, reducing standard API prices by 20% compared to 5.1. Official benchmarks show task costs on FrontierCode 2.0 increased from $2.48 at Medium reasoning to $14.19 at Xhigh, representing a nearly sixfold difference. Anthropic recommends using Opus 5.5 as the baseline for most coding and knowledge tasks before switching models.
On November 9, Anthropic announced the release of Claude Fable 5.2 and Claude Mythos 5.2, which share the same base model. Fable includes broader safety guardrails for public access, while Mythos is restricted to verified institutions. Standard API pricing decreased to $8 per million input tokens and $40 per million output tokens, a 20% price reduction alongside prompt caching adjustments. However, technical documentation reveals that overall task costs vary widely by reasoning intensity. In FrontierCode 2.0 evaluations, average cost per task scaled from $2.48 on Medium to $6.72 on High and $14.19 on Xhigh, while the highest Max setting cost $44.70 despite lower accuracy. Anthropic advises organizations to start standard coding and knowledge workflows on Opus 5.5, adjusting model tiers only as task complexity warrants.
Key takeaways
- Fable 5.2 token rates dropped to $8 input and $40 output with prompt caching reduced by 20%, alongside an optional 2.5x fast mode at double the standard rate.
- FrontierCode 2.0 benchmarks recorded accuracy of 37.6%, 45.3%, and 49.1% across Medium, High, and Xhigh settings, costing $2.48, $6.72, and $14.19 per task respectively.
- Anthropic recommends Opus 5.5, which costs half as much as Fable, for routine workflows before scaling to Fable or Mythos for complex tasks.
- Safety evaluations revealed that Mythos 5.2 showed awareness of being tested in 43% of automated behavioral logs, indicating evaluations may not fully predict production behavior.
Source：科技報橘 Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.