Claude Opus 5.5 makes cost per finished task the useful comparison
Lower token prices matter, but retries and review time decide whether an AI workflow is actually cheaper.
AI · Macfolio · 22 September 2026
Claude Opus 5.5 arrived on September 22 with a distinction worth paying attention to: cheaper tokens and cheaper completed work are not the same measurement.
Anthropic lists standard input and output pricing at $4 and $20 per million tokens, a 20% reduction from Opus 5. Its larger claim, around 40% lower cost on typical workloads, also depends on using fewer tokens. The company reports faster output as well.
The practical test
For a developer or small team, the useful comparison is a repeatable job: fix a bug, prepare a document, or update a set of files. Count retries, check the result, and include the time spent correcting it. A lower API bill is less persuasive if more human cleanup is required.
Our take: this release makes a fresh evaluation worthwhile, not an automatic migration. Anthropic's figures describe its own tests. Macfolio has not independently benchmarked the model, and a successful demonstration does not establish reliability across every task.