Anthropic has launched Opus 5.5, a new model that outperforms its larger sibling, Fable, on several benchmarks while costing less to run.
The release arrives just two months after Opus 5 debuted on July 24. Opus sits at the top of a three-tier system alongside the mid-range Sonnet and the budget-friendly Haiku. This latest version beats Fable in coding and knowledge tasks, completing informal challenges that the bigger model previously failed.
Price drops and speed gains
The cost to generate output tokens has fallen to $20 per million, down from $25 for the prior version. Other pricing metrics have seen similar reductions. The model also runs faster, indicating a lower total compute requirement.
Communication style has shifted as well. Opus 5.5 avoids jargon more often and places critical information at the beginning of its responses rather than burying it.
Safety and policy
Anthropic states that Opus 5.5 matches the biology and cybersecurity capabilities of Mythos. Consequently, it faces the same restrictions as Fable. These rules prevent the model from identifying exploits in compiled software or creating biological weapons.
This launch marks a shift in strategy for CEO Dario Amodei. He has moved to slow the pace of capability development to allow safety measures to catch up.
Amodei wrote earlier this month that addressing risks requires more prudence. He argued that the rate of advancement must slow so that risk prevention can keep up.
Safety training for Opus 5.5 remains similar to previous models, with testing conducted by external groups like METR and Frontier Design. The company noted that more advanced security and monitoring systems are already being prepared for future releases.
The blog post states that public policy must play a larger role as AI grows more capable. Anthropic says it is building the infrastructure to support this capacity and expects to share more details soon.
