SAN FRANCISCO, September 22 — Anthropic unveiled Claude Opus 5.5 on Monday, introducing the first model in its new Claude 5.5 family with performance matching Claude Fable 5.1 on most work while costing significantly less to run. The model was tested by external evaluators including Frontier Design and METR before release. On Anthropic’s automated behavioral audit — its most comprehensive alignment test — Opus 5.5 emerged as the strongest-performing model evaluated to date, incorporating safeguards developed for the company’s most capable systems.
In practical testing, Opus 5.5 showed major gains over its predecessor. One tester completed a 680,000-line code migration in under a day, a task that would have taken an engineering team weeks.
When asked to reduce load times across every page of a web app, Opus 5.5 succeeded 39 out of 40 attempts, while Opus 5 delivered smaller improvements that also unintentionally altered the app’s behavior. In a creative benchmark where multiple Claude models built a game from a single prompt, Opus 5.5 scored higher than any other model on graphics and polish.
On safety metrics, it achieved the best scores ever recorded on the automated behavioral audit, is far less likely than recent models to take irreversible or boundary-violating actions, and shows greater resistance to prompt injection than Opus 5. Anthropic expanded its alignment testing to include longer tasks, impossible tasks, and real-world incident scenarios, though it acknowledged the model still has limits.
Full evaluation details are published in the Opus 5.5 System Card. Because Opus 5.5 performs comparably to Claude Mythos 5.1 in biology and cybersecurity, it is being deployed with safeguards similar to those on Claude Fable 5.1. Qualified organizations can apply to the Life Sciences Verification Program to use Opus 5.5 for biology research, and access to the Cyber Verification Program will expand in the coming weeks for verified cybersecurity professionals.
On efficiency, Opus 5.5 requires less compute than Opus 5, with pricing reflecting this improvement. At default settings, it costs 40% less than Opus 5 on typical workloads. Input and output tokens are priced at $4 and $20 per million respectively — 20% less than Opus 5.
Cache reads, which dominate costs in agentic and coding workloads, are $0.20 per million tokens — 60% less than Opus 5. The model also generates output more than 30% faster than Opus 5.
Anthropic increased five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans and gave subscription users a rate limit reset that can be saved and used at their discretion. Early testers noted clearer, more natural writing that was easier to follow, addressing prior feedback about Opus 5. One tester said, “it writes the way I do,” a clarity that supports both usability and safety by making work simpler to review and verify.
Claude Sonnet 5.5 and Claude Haiku 5.5 are expected in the coming weeks with comparable upgrades. On benchmarks, Opus 5.5 leads in agentic coding, computer use, and knowledge work, though Anthropic cautions that benchmark differences have become less reliable indicators of real-world performance at this level.
Internal use shows the gap between Opus 5.5 and Claude Fable 5.1 is narrower than benchmark scores suggest. Unless otherwise noted, all Opus 5.5 results use adaptive thinking at maximum effort. Terminal-Bench 4.0 results for Opus 5.5 are reported at xhigh effort with a standard error of ±2.6 points.
AutomationBench results were conducted and reported by Zapier from its early access evaluation.
