The bill falls, and some steps finish on an older model
Anthropic opened Claude Opus 5.5 on 22 September and says a typical workload at default settings costs 40 percent less to run than Opus 5. Input and output are priced at 4 and 20 dollars per million tokens, and a cache read at 0.20 dollars per million. The company writes those as 20 percent and 60 percent lower, and says output is more than 30 percent faster. That cost is booked to the buyer's bill.[1]
The same announcement says that when production safeguards intervened in the evaluation, cybersecurity tasks were completed by Claude Opus 4.8 and biology and frontier-model development tasks by Claude Opus 5. The Opus 5.5 price is not the name of the model that finished those steps. Anthropic also writes that, in its own use, the text became easier to follow and to check. The person who checks it is still the person reading the result.[1]
A 680000-line day is not a headcount count
Anthropic writes that one tester finished a 680000-line code migration in less than a day, work it says would have taken an engineering team weeks. In a test that cut page-load times, Opus 5.5 succeeded in 39 of 40 tries. Those two sentences come from the company's testers. Headcount, hours and who read the diff are absent from that comparison.[1]
The speed gain and the review time are separate ledgers. Anthropic says the text is easier to check in its own use; it does not include a measurement that review hours fell for customers. The 40 percent serving-cost cut and an hour off a worker's day are not the same figure. Reading them side by side does not build the bridge that shows who receives the gain; the breakdown of which step finished on Opus 4.8 is what builds that bridge.[1]