Opus 5.5: judge complex work by accepted outcomes

The release focus

Anthropic introduced Claude Opus 5.5 on September 22. Its announcement emphasizes complex coding, repository migrations, longer tasks and clearer collaboration, supported by vendor evaluations and early customer reports. Source 1

Budget by operating mode

The announcement lists standard prices of $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20. Fast mode lists input and output at $8 and $40. Keep these modes separate in estimates and confirm the actual provider’s bill.

Our analysis: include unintended changes

For a cross-module refactor, code volume and elapsed time are incomplete measures. Check whether behavior is preserved, failures can be recovered, permissions are respected and the result can be maintained by the team. These acceptance criteria reveal whether faster execution produces useful work.

Treat examples as bounded evidence

Customer stories and benchmark results depend on their tools, settings and workloads. We have not reproduced the vendor tests and do not infer a fixed saving for a new project. Longer runs also make checkpoints and budget limits more important.

Start with a reversible task

Choose a real task with meaningful automated checks, run it in an isolated branch or test environment, and retain the model version, settings and tool logs. Include human review time when comparing the total cost of accepted outcomes.