Anthropic shipped Claude Opus 5.5 yesterday, the first model in the new 5.5 family. It performs at the level of Claude Fable 5.1 on most work, the company says, while costing 40 percent less to run than Opus 5. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks.
Something was coming, and we knew it: three days ago I wrote that Anthropic was weighing a new model. Here it is.
66.4 percent on Terminal-Bench, with a caveat from Anthropic itself
Opus 5.5 scores 66.4 percent on Terminal-Bench 4.0. Fable 5.1 sits at 55.8, Opus 5 at 52.3, GPT-6 Astra at 57.9. On FrontierCode v1.1 it is 54.4 against 50.3 and 48.0. On GDPval-AA v2.1, 1846 Elo against 1735 and 1708. OSWorld 2.0 lands at 81.8 percent.
What stands out is the sentence Anthropic prints right next to the table. At this level of capability, it says, benchmark margins have become «a less reliable guide to real-world differences». In its own use, the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest. You do not often see a company relativise its own chart.
The price falls hardest where agents burn the most
Input runs $4 per million tokens, output $20. Against Opus 5 at $5 and $25, that is 20 percent off each. The real lever sits elsewhere: cache reads drop from 50 to 20 cents, down 60 percent. Run an agent for hours across the same codebase and that is exactly where the money goes.
On top of that, Opus 5.5 needs fewer tokens per task. Together those two add up to the 40 percent. Output also comes out more than 30 percent faster. Claude Code and the Claude Platform get a Fast Mode with up to 2.5x speed at $8 and $40 per million tokens.
And one more thing plenty of people will notice right away: Anthropic is raising five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans. Subscribers also get a rate limit reset they can bank and spend whenever they want.
Eighteen hours unattended
The numbers from early testers say more than any benchmark. One finished a 680,000-line migration in less than a day. Another audited and fixed a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and 2.5 times as many tokens.
Sean Heintz at Clio handed the model a task spanning six repositories and let it run overnight, unattended: «It stayed on task for over 18 hours.» Anthropic itself had Opus 5.5 and Fable 5.1 translate HAProxy from C to Rust. Both passed nearly all of HAProxy’s own regression tests; Opus 5.5 took 9.5 hours instead of 12 and cost 51 percent less.
The same guardrails as Fable 5.1
Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships with the safeguards built for Fable 5.1. Vetted organisations can apply to the Life Sciences Verification Program today, and the Cyber Verification Program opens up in the coming weeks.
On Anthropic’s automated behavioural audit, its most comprehensive alignment test, Opus 5.5 is the strongest-performing model the company has tested. It reaches for hard-to-reverse actions less often and resists prompt injection better than Opus 5. In a test by security firm Gray Swan it ties Fable 5.1 at the top. External evaluators, including METR and Frontier Design, tested it before release.
The first model after the brake call is an efficiency model
Anthropic writes it plainly: this is the first release since Amodei’s call to slow the pace. And look closely, the release really is built differently from the ones before it. The biggest jump is not in intelligence but in price and token consumption. Same level as Fable 5.1, considerably cheaper.
If you work in Claude Code every day, that is the more useful kind of progress. A model that does the same job for 40 percent less changes more about your week than ten points on a benchmark nobody outside the labs recomputes.