Opus 5.5 cuts standard API token prices by 20%; task savings depend on usage

Opus 5’s corresponding prices were $5 and $25. At identical token volumes, that is a twenty-percent reduction. The company’s larger claim of roughly forty-percent lower task costs also depends on using fewer tokens, so it should not be read as the published price cut alone. [1] [2]

A small worked example makes the distinction visible. Ten thousand ordinary input tokens and two thousand output tokens cost eight cents at the new rates, compared with ten cents at the previous rates. This calculation excludes caching, tool charges, fast mode and other pricing modifiers; it holds usage constant rather than predicting how either model will behave. [2]

The release is also an integration change. Documentation inspected for this September 28 revision says thinking cannot be disabled and forced tool selection is unsupported. Applications that rely on those settings need a compatibility check before a model-name substitution. The launch statement also describes safeguards that can route certain work to other models, which matters when interpreting a result as the performance of one named model. [1] [3]

The practical comparison is therefore a completed job with an agreed standard. Take a small set of representative tasks, preserve the same inputs and record the total charge, elapsed time, retries and human corrections. A lower bill is useful only if the work remains acceptable; a better answer may justify more computation. Anthropic’s published comparisons make a case for testing. They do not establish the saving or reliability a particular team will obtain.