TheModelDesk

Anthropic now says Opus 5.5 beats the larger Fable model on several benchmarks and completed informal tasks that Fable failed to finish. The model was released two months after Opus 5, which arrived on July 24.

Output tokens are charged at $20 per million for Opus 5.5, compared to $25 previously. Opus 5.5 runs faster because serving it requires less compute, according to Anthropic. The company also changed how the model communicates. It uses less jargon and moves important information toward the start of an answer.

Sonnet 5.5 and Haiku 5.5 are planned for the coming weeks, with similar performance improvements. Anthropic considers Opus 5.5 comparable to Mythos in biology and cybersecurity. The model therefore receives the safeguards used for Fable. These controls limit activities such as finding exploits in compiled programs and developing recognizable biological weapons.

Opus 5.5 is Anthropic's first model release since CEO Dario Amodei called for a slower pace in frontier capability development. He argued that alignment and risk-prevention work need time to keep up. Opus 5.5 received safety training broadly similar to earlier models, including alignment testing and pre-release evaluations by METR and Frontier Design. Anthropic is preparing more advanced training and evaluation systems, along with improved security and monitoring, for future models. The company also says public policy should play a larger role in AI safety and that it is building the necessary infrastructure.