On 22 September Anthropic released Opus 5.5, the first model in the 5.5 family. The token price drops by 20%, cached input by 60%, and per the company the model also spends fewer tokens per task. Two API changes can break requests that used to work.
- Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, and cache reads drop to $0.20.
- Per Anthropic, typical work comes out 40% cheaper than Opus 5, because the token is cheaper and there are fewer tokens per task.
- The model was tested before release by external evaluators including METR and ships with safeguards similar to Fable 5.1.
The most useful line around this release is not from Anthropic. It is from Vercel's note, and it says some requests will start returning a 400 error.
Opus 5.5 does not accept a fixed thinking budget and does not let you force the model to call a specific tool. If your system has been getting JSON back by forcing a tool call, it will stop working when you switch models. You rewrite it with structured outputs and run it again.
Where the saving is
Part of the discount is in the token price. Anthropic looks for the rest in the model reaching the result with fewer tokens. The example they give comes from an early tester: an audit and fix of a 200,000-line codebase in under three hours, where Opus 5 worked for over 20 hours and spent 2.5 times as many tokens.
That is their number from an early tester, not ours. But it is the number easiest to check at home: same task, same repository, both models, and the bill at the end.
There is also one sentence worth reading. Anthropic writes that it sees signs the model often suspects it is being tested, and that this makes it harder to judge how it will behave out in the world. Good that they say it themselves.
If you use Opus through the API, check your code for forced tool choice and a fixed thinking budget before you change the model name.