Claude Opus 5.5 cuts typical agent costs by 40%, Anthropic says

Anthropic introduced Claude Opus 5.5 on September 22, saying its new flagship reaches the level of Claude Fable 5.1 on most work while costing about 40% less to run than Opus 5 on typical workloads. The striking part is the pairing: a model pitched for difficult, long-running agent tasks is also being sold as a cheaper one to operate.

The price cut is about more than the token list

Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens, each 20% below Opus 5. Cache reads fall from $0.50 to $0.20 per million tokens. That matters for coding agents, which repeatedly reuse context as they inspect files, test changes and return to earlier material.

The company's 40% typical-workload saving is a broader estimate, based on lower serving costs and fewer tokens used per task. It is not a promise that every API bill will fall by that amount. Anthropic also claims output is more than 30% faster than Opus 5. A separate fast mode costs more, at $8 per million input tokens and $40 per million output tokens.

A stronger agent, with a narrower test than real life

Anthropic says Opus 5.5 leads its selected evaluations in agentic coding, computer use and knowledge work. In one internal app optimization exercise it succeeded in 39 of 40 attempts. An early tester reportedly migrated a 680,000-line codebase in less than a day. Those are notable examples, but they do not establish how a typical organization will fare with its own code and constraints.

The company itself cautions that benchmark margins have become a less reliable guide to real-world differences. Its comparisons also use different effort settings and, for some rivals, figures reported by other companies. Buyers should test task completion, review burden and total cost on their own workloads before treating leaderboard positions as procurement answers.

Safety moves closer to the release decision

Opus 5.5 is Anthropic's first model since chief executive Dario Amodei called for pacing frontier AI development. Anthropic says external groups including METR and Frontier Design tested it before launch. On the company's automated behavioral audit, it scored better than recent Claude models on most measured signs of misaligned behavior, including boundary crossing and prompt injection resistance.

Those are company-reported evaluation results, not proof of safety in every deployment. Anthropic says the model can appear to recognize evaluations, which makes unknown real-world settings harder to assess. It is routing many cybersecurity tasks to another model and offering verified access paths for eligible defenders and biology researchers.

Opus 5.5 is available across Claude's platforms and major cloud partners. Sonnet 5.5 and Haiku 5.5 are due in coming weeks. The open question is whether the lower cost and stronger safeguards hold up when independent users put long-running agents to work.