Opus 5.5 puts the cost of completed work in focus
A new model with new rates
On 22 September 2026, Anthropic introduced Claude Opus 5.5. Published input and output token rates are lower than Opus 5’s; overall improvement claims come from the vendor’s evaluations.
Count the complete task
A development model consumes more than one answer. It may read code, call tools, retry a correction and verify its output. The useful cost measure for a team is therefore an accepted task, including attempts and human review time.
A meaningful comparison would reuse previously solved problems: a reproducible defect, a bounded feature and a documentation change. It would keep tools, rules and acceptance criteria constant, then observe generated changes, errors and required interventions.
Benchmarks are a starting point
A stronger evaluation score does not establish autonomy in a particular repository. Permissions, context quality and available tests strongly affect results. Before rolling out a model change, teams should verify its value on real work and retain a way to compare against their previous configuration.
Editorial retrospective: event dates are stated in the text.