Anthropic introduced Claude Opus 5.5 on 22 September. The company says it can handle large coding and knowledge-work tasks with less compute than Opus 5. API pricing fell to $4 per million input tokens and $20 per million output tokens, 20% below its predecessor. Reading previously cached context costs $0.20 per million tokens.

The company estimates that its typical workload costs 40% less to run: it combines the lower unit price with fewer tokens used. That is not a universal discount for every customer. Anthropic reports stronger agentic coding, computer use and knowledge-work results; its own announcement also warns that small benchmark gaps are becoming less reliable guides to real-world differences.

Safety matters here. Anthropic reports expanded testing of task boundaries and resistance to malicious instructions embedded in outside material. External organizations also evaluated the model before release, but the published overall conclusions and most figures still come from the provider. Because Anthropic attributes high biological and cyber capability to the model, it restricts advanced use in those fields and offers special access to vetted organizations.

If its efficiency holds on real data, the lower cost could make longer audits of large codebases, documentation and complex workflows more practical. Prerequisites include repeatable tests on an organization's own cases, measuring cost per correct result, testing failure and security scenarios, and requiring human approval for irreversible actions. Limited pilots could optimistically be assessed within weeks to a few months; a timeline for generalizing to critical processes is not yet defensible.