Last verified: October 2, 2026 (Pacific).
Direct answer: Anthropic released Claude Opus 5.5 on September 22, 2026 — the first model in the Claude 5.5 family. It matches Claude Fable 5.1 on most work while costing 40% less to run on a typical workload than Opus 5. API pricing: $4 per million input tokens, $20 per million output tokens, 20% lower than Opus 5 across the board. Cache reads dropped 60% to $0.20 per million tokens. Opus 5 is now legacy.
What actually changed
Three moves at once: cheaper, faster, and smarter at using less effort.
- Price. $4/$20 per million input/output tokens, down from Opus 5's $5/$25. Cache reads fell from $0.50 to $0.20 per million — a 60% cut.
- Speed. Output generation is more than 30% faster than Opus 5. A faster serving mode runs 2.5x faster at double the price ($8/$40 per million tokens).
- Effort. Opus 5.5 defaults to *medium* effort, one step below Opus 5's high default. Anthropic's testing found medium-effort 5.5 matches or beats high-effort Opus 5 on coding and knowledge work. Thinking cannot be turned off on Opus 5.5 — requests that try get an error.
The benchmarks, with the caveat they deserve
All numbers below are Anthropic-reported, not independently reproduced. Treat vendor benchmarks as direction, not gospel — validate against your own task corpus.
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 55.8% | 52.3% |
| GDPval-AA v2.1 (professional work, 44 occupations) | 1846 Elo | 1735 | 1708 |
| AutomationBench (task completion) | 40% | — | 26.9% |
| FrontierCode (max effort) | 54.4% | 50.3% | — |
Anthropic also claims Opus 5.5 beats Opus 5 at maximum effort on FrontierCode at about a fifth of the cost, and matches GPT-6 Astra on Terminal-Bench at roughly 40% of the cost. Cost-per-task is the more honest frame than headline margins.
Prompting: effort is the first dial
Anthropic published prompting guidance alongside the release, and the headline advice is simple: stop reusing your Opus 5 effort settings. The effort setting — low, medium, high, xhigh, max — is now the first control for balancing quality against speed and cost, before prompt changes.
Two concrete changes worth making:
1. Test effort levels directly. Don't assume high is better. In Anthropic's testing, medium 5.5 matches high Opus 5.
2. Reconsider "think carefully" lines in chat apps. Opus 5.5 decides for itself how much to think, with effort as the main control. Anthropic's own test showed removing a think-carefully instruction made replies start sooner with no clear quality drop.
Opus 5 prompts still work without edits — the existing guidance is a reasonable starting point. Adjust effort before rewriting prompts, and reserve xhigh and max for tasks where quality really moves.
Availability and context
Opus 5.5 ships on Anthropic's platform and through AWS, Google Cloud, and Azure. It carries a 1M-token context window with 128K maximum output. Before launch, external evaluators including Frontier Design and METR tested the model, and Anthropic says it recorded its best result to date on the company's automated behavioral audit.
This is the first release since CEO Dario Amodei's September essay calling on the industry to pace frontier development — and Anthropic shipped it with the same cybersecurity and biosafety measures as Fable 5.1, plus published evaluator access, which is the pacing commitment in concrete form.
The rest of the family is filling in: Sonnet 5.5 landed September 28 at $2/$10 with the same $0.20 cache reads and 30%+ speed gain over Sonnet 5, and Haiku 5.5 is listed as coming soon. For the full lineup and legacy status, see our Claude Release History (Sept 2026).
Leave a Reply