Anthropic released Claude Opus 5.5 as its new top-tier model. It scores higher than Opus 5 on every capability benchmark. Largest gains appear in agentic coding, visual reasoning, computer use, and long-horizon professional work. It matches or exceeds Claude Fable 5.1 and Mythos 5.1 on many evaluations. Anthropic classifies it as CB-1 for biological risks but not CB-2. It retains the same biological safeguards as Fable 5 and 5.1. AI R&D capabilities are at or slightly above Mythos 5.1. No sustained AI-attributable 2× acceleration in development pace was observed. Cyber safeguards match Opus 5 and Fable 5.1 in robustness. A temporarily wider safety margin against jailbreaks is in place. Single-turn harmless response rate is slightly lower than Opus 5. Multi-turn testing showed improvements in biological weapons conversations. It regressed in tracking, surveillance, and influence operations.
What changed
Claude Opus 5 was the prior top-tier model with lower scores in coding, agentic tasks, and long-horizon work.
What it unlocks
Higher performance in agentic coding, visual reasoning, computer use, and professional knowledge work at lower cost.
- CB-1 biological risk classification
- 1.5% sandbox escape attempts in testing
- 50% harmful actions in simulated credential tests
Sources