Claude Opus 5.5 launches 40% cheaper as its maker calls for AI caution

A launch piece by Susan Hill for Martin Cid Magazine on Claude Opus 5.5, which Anthropic released on 22 September. It lists the numbers. Input and output cost $4 and $20 per million tokens, down from $5 and $25 for Opus 5; cached reads fall from $0.50 to $0.20; fast mode is $8 and $40. Generation is more than 30% faster than Opus 5. On Anthropic’s benchmarks the model leads Claude Fable 5.1 on Terminal-Bench 4.0 (66.4% to 55.8%), Humanity’s Last Exam (67.7% to 65.6%) and OSWorld (81.8% to 80.7%), and trails GPT-6 Astra on AutomationBench and Terminal-Bench-Science. In Deloitte’s code reviews it caught 72% of known bugs at its lowest effort setting, against 56% for Opus 5 at high effort. External testers included METR, and an internal audit found it about 85% less likely than Opus 5 to try to get around its constraints. The article says plainly that these figures come from Anthropic and its chosen partners, and that no independent verification had been published at launch.

The angle is in the headline. Earlier in the month Dario Amodei had argued for pacing the advance of capabilities so that risk prevention can keep up, and here is a more capable model, cheaper and faster. Hill offers two readings: a better-evaluated model in wider hands is what Amodei meant, or capabilities advanced, which is what the essay named as the concern. She says the benchmarks cannot settle which.

The prices, the speed claim and the benchmark figures match Anthropic’s own announcement, with one correction to the headline. The list price falls by 20%, and cached reads by 60%. The 40% is Anthropic’s estimate of the saving “on typical workloads” at default settings, a blend that depends on how much a workload reuses its cache. Anthropic’s table calls the computer-use benchmark OSWorld 2.1, names Frontier Design beside METR as external evaluators, and puts the 85% against Claude Mythos 5.1 as well as Opus 5. The juxtaposition is not only the magazine’s: Anthropic’s launch post itself cites Amodei’s essay. Nothing in what was read had been replicated independently.