Cheaper tokens and faster answers are the headline on Anthropic’s newest Claude release. Opus 5.5 arrives with output priced at $20 per million tokens, down from $25 for the version that shipped on July 24, and the company says it needs fewer tokens to reach the same quality. Answers come back more than 30 percent faster because serving the model costs less compute.
The tier is smaller than Fable, yet Anthropic claims it beats that larger system on many benchmarks and finished informal tasks Fable failed. Communication changed too: less jargon, and the important material placed up front. Box AI measured a third of the tokens Opus 5 burned, with replies 40 percent shorter and no accuracy loss. Engineers at Ramp said a design spec came out usable after minimal edits.
Subscribers get more room to work. Five-hour usage limits rise 20 percent, and because the model drains quota more slowly the practical capacity gain lands near 50 percent. Sonnet 5.5 and Haiku 5.5 follow in the coming weeks.
On biology and cybersecurity the model rates alongside Mythos, so it inherits the same gating that restricts the Fable system, including limits on exploit discovery and dangerous biological work. Alignment testing and pre-release review went to outside groups including METR and Frontier Design. It is the company’s first launch since CEO Dario Amodei argued for pacing capability gains so risk prevention can keep pace.
