Anthropic released Claude Opus 5.5 on September 22, the first model in the Claude 5.5 lineup. The company's pitch comes down to two claims: performance on most tasks roughly matches Claude Fable 5.1, and overall running costs are 40% lower than Opus 5.
The price sheet breaks down by line item: input drops to $4 per million tokens and output to $20, each down about 20%; cached reads fall from $0.50 to $0.20, a 60% cut. On speed, Anthropic says output is more than 30% faster than Opus 5, and there's also a fast mode priced at $8 and $40 that runs up to 2.5 times faster. The model is live on the Claude platform API (model ID claude-opus-5-5), AWS, Google Cloud and Microsoft Azure, with consumer apps rolling out across platforms at the same time; maximum output per request is 128,000 tokens.
Benchmarks
Here's the comparison Anthropic published on its release page.
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% |
| GDPval-AA v2.1 | 1846 Elo | 1735 | 1708 |
| OSWorld 2.0 | 81.8% | 80.7% | 74.0% |
Third-party evaluator Artificial Analysis ranks Opus 5.5 first on its intelligence index, with a score of 58. Claude Code's 2.1.280 release, shipped the same day, has already made Opus 5.5 the default Opus model.
The release page also quotes feedback from several customers who trialed the model: GitHub says it solves the same terminal tasks in less than half the steps Opus 5 needed; Optiver says quality is on par with Opus 5 while turn counts fall by roughly half and costs drop 40% to 50%; Deloitte says the model caught 72% of known defects at its lowest thinking setting, versus 56% for Opus 5 at its highest setting. All of these figures come from customer quotes relayed on the vendor's own release page, with no independent verification.
What Anthropic says about safety
Anthropic says Opus 5.5 passed automated behavioral audits with the strongest alignment score of any Claude model to date, and attempts to circumvent sandbox boundaries fell 85% compared with Opus 5. Most cybersecurity-related tasks are still routed to Opus 4.8, and the biosecurity safeguard level matches Fable 5.1. The model keeps its reasoning-trace protections against distillation, carries a watermark in line with the EU AI Act, and offers enterprises a zero-data-retention option. METR and Frontier Design took part in pre-release evaluation.
Specific failure rates, refusal rates and red-team results will have to wait for the full system card; the release page itself doesn't list those numbers.
Reading it against the same day's other price sheet
OpenAI also shipped new models on September 22, GPT-6 Sol and Luna, with API pricing cut roughly in half from the prior generation: Sol comes in at $2 and $10, Luna at $0.10 and $0.50. Side by side, Opus 5.5 still costs twice as much per token as Sol — Anthropic is selling scores and fewer steps at the same price tier. Going by Optiver's and GitHub's numbers, fewer turns per task partly offsets that per-token gap on the actual bill. How much it offsets depends on the shape of the task, and there's no general answer.
One counterexample has already surfaced from developers. Blogger Simon Willison tested Opus 5.5 on generating a complex SVG at the highest thinking setting, and the model burned through its full 128,000-token output budget without finishing the task. High thinking settings paired with long outputs can eat an entire output budget fast, and that ceiling needs to be factored in before wiring the model into any automated pipeline.
For teams already running Opus 5, the switching cost is mainly in two places: the model ID needs updating, and changes in output speed and turn counts mean existing timeout and retry settings may no longer fit. The price cuts matter most for high-frequency use cases, especially that 60% drop in cached-read pricing — the longer a standing system prompt and the denser the request volume, the bigger the savings.
Sources: Anthropic's official release page, Artificial Analysis, CocoLoop, Simon Willison's blog; per-tier pricing, benchmarks and safety details verified against Anthropic's release page.