Published: 23 September 2026
Anthropic released Claude Opus 5.5 on 22 September 2026, priced at $4 per million input tokens and $20 per million output tokens. Anthropic says it costs 40 percent less to run than Claude Opus 5 on typical work and generates output more than 30 percent faster. It is available now on the Claude Platform, AWS, Google Cloud and Microsoft Azure.
What is new in Claude Opus 5.5?
The headline is not a new capability but a new price point. Claude Opus 5.5 matches or beats Claude Fable 5.1, Anthropic’s own flagship, on most published benchmarks while costing $4 and $20 against Fable 5.1’s $10 and $50. Anthropic’s documentation now recommends Opus 5.5 for most workloads and reserves Fable 5.1 for demanding reasoning and long-horizon agentic work.
Speed is the second change. Anthropic cites a 680,000 line code migration finished in under a day, and claims 2.5 times better token efficiency on large codebases. A fast mode doubles output speed again at $8 and $40 per million tokens. The model also carries a new safeguard Anthropic calls preserved thinking, which is designed to stop competitors distilling its reasoning traces.
What you get with it:
- API model name: claude-opus-5-5
- Context window: 1,000,000 tokens, with up to 128,000 tokens of output
- Training cutoff: June 2026
- Cache reads at $0.20 and cache writes at $5 per million tokens
- Watermarking for EU AI Act compliance
- Claude Sonnet 5.5 and Claude Haiku 5.5 promised in the coming weeks
Claude Opus 5.5 benchmarks against GPT-6 Astra
Take this from the table: Opus 5.5 wins the coding rows clearly and loses the two rows where OpenAI is strongest.
| Measurement | Claude Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Terminal-Bench 4.0 (command line work) | 66.4% | 57.9% |
| FrontierCode v1.1 (coding agents) | 54.4% | 53.3% |
| AutomationBench (business workflow tasks) | 40.0% | 41.4% |
| Terminal-Bench-Science 0.1 (scientific tooling) | 58.7% | 64.6% |
| Humanity’s Last Exam (hard expert questions) | 67.7% | 57.2% |
| GDPval-AA v2.1 (real professional work, Elo) | 1846 | 1542 |
These are Anthropic’s own evaluations, published in the launch post, and the model card does not say where the GPT-6 Astra figures came from. Anthropic reports standard errors of plus or minus 2.6 points on Terminal-Bench 4.0 and up to 5 points on Terminal-Bench-Science, which is wider than several of the gaps in the table.
What Claude Opus 5.5 means for Claude Fable 5.1
Anthropic has undercut its own top model. Opus 5.5 beats Fable 5.1 on Terminal-Bench 4.0 by 10.6 points, on CursorBench 4.0 by 6 points and on FrontierCode v1.1 by 4.1 points, at 60 percent less per token. Fable 5.1 was released on 1 September 2026, so it held the top slot for three weeks.
Anthropic is unusually direct about the limits of its own numbers. The launch post states that at these capability levels benchmark margins have become a less reliable guide to real-world differences. It also discloses that Opus 5.5 was evaluated with production safeguards switched on, and that when those safeguards intervened, cybersecurity tasks were completed by Claude Opus 4.8 and biology tasks by Claude Opus 5, which Anthropic says likely lowered the scores on those benchmarks. That is a caveat most vendors bury.
What this means
Worth testing now, and worth testing as a replacement rather than an addition. A platform team running coding agents on Claude Fable 5.1 should run its own evaluation on Opus 5.5 this week, because if the results hold the same work costs 60 percent less per token and finishes faster. The migration is a model string change, not a rewrite. Teams on Opus 5 have an easier call still, since the cheaper model is also the better one on every published row.
The one group that should wait is anyone whose work sits close to the safeguard boundary, such as security research or biology tooling. Anthropic has said outright that safeguards intervened during its own evaluations and handed those tasks to older models. If that is your workload, the published scores do not describe what you will get. For everyone else the more interesting question is what Anthropic does with Fable 5.1 now, because a flagship that costs 2.5 times as much and loses to its own cheaper sibling on coding does not have an obvious place in the lineup.
For more information, visit the official announcement of Claude Opus 5.5 on the Anthropic blog.
