Anthropic released the latest Claude Opus 5.5, with performance benchmarked against Claude Fable 5.1 and running costs cut by 40% compared with Opus 5.
To save face, it went out of its way to stress that this is "the first release since the call to set the pace for frontier development" — talking tough while its actions remain remarkably candid.
Key features are as follows:
- On coding performance, an internal test completed a 680,000-line code migration within a single day
- Achieved the best safety results of any model to date
- Opus 5.5's capabilities in biology and cybersecurity are on par with Claude Mythos 5.1, so the same level of safety safeguards has been deployed
- Input and output per million tokens are $4 and $20 respectively, 20% cheaper than Opus 5; cache reads are $0.2 per million tokens, 60% lower than Opus 5. Opus 5.5's output generation speed is also more than 30% faster than Opus 5
- More natural communication; based on the description, Opus 5.5 received feedback on writing and this time made special optimizations for writing
Other interesting points:
- Anti-distillation attacks — something Claude has been doing all along; Opus 5.5, like Fable 5.1, has made special provisions in this area, preventing extraction of Claude's reasoning process during API calls
- Opus 5.5 can provide zero data retention, likewise accompanied by watermarking
The model is now live on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure, under the model name "claude-opus-5-5".
The rest is content related to model benchmarks, which I won't go into here; those interested can check the official blog.