Claude Opus 5.5 launches 40% cheaper as its maker calls for AI caution

Claude Opus 5.5 does something Anthropic’s previous models didn’t: it costs less and runs faster while matching or outperforming a bigger, pricier system. The model handles agentic coding, computer use, chart reading, and multidisciplinary reasoning at levels that, according to Anthropic’s own data, exceed those of Claude Fable 5.1, the company’s previously most capable model, on several measures. For people and developers who found Opus 5 useful but expensive, the gap between what the model can do and what it costs has shifted significantly.

The price difference is specific. Opus 5.5 charges $4 per million input tokens and $20 per million output tokens; Opus 5 cost $5 and $25 respectively. Cached reads fall from $0.50 to $0.20 per million tokens. Anthropic says typical workloads run about 40% cheaper overall. Speed tracks the same direction: the standard model generates text more than 30% faster than Opus 5. A separate fast-mode option, priced at $8 input and $40 output per million tokens, reaches up to 2.5 times the speed of standard Opus 5 — a useful setting for latency-sensitive applications.

Anthropic’s published performance figures place Opus 5.5 ahead of Fable 5.1, Opus 5, and GPT-6 Astra on four of six benchmarks. On Terminal-Bench 4.0, which tests code execution in a live terminal environment, Opus 5.5 scores 66.4% against Fable 5.1’s 55.8%. On Humanity’s Last Exam, a test spanning academic and professional knowledge across disciplines, the result is 67.7% against 65.6%. OSWorld 2.0, which measures computer-interface operation, shows 81.8% against 80.7%. A Deloitte test found that Opus 5.5 at its lowest effort setting detected 72% of known coding errors in review, against 56% for Opus 5 at high effort. These numbers come from Anthropic and its selected partners; stated margins of error run from ±1.6 to ±5 percentage points, and no independent third-party verification has been published at launch.

Two areas break the pattern. On AutomationBench, GPT-6 Astra scores 41.4% to Opus 5.5’s 40.0%. On Terminal-Bench-Science, the gap is wider: Astra reaches 64.6% while Opus 5.5 lands at 58.7%. Both differences fall within the stated error margins — they may be noise — but they may not be. Anthropic includes these rows in its own published table, which is more than most AI companies do at launch. The numbers still need independent scrutiny before they can be treated as settled.

On safety, Anthropic says Opus 5.5 was evaluated before release by external teams including METR. In the company’s own internal behavioral audit, nearly 2,000 scenarios designed to test whether the model tries to work around constraints placed on it, Opus 5.5 is described as 85% less likely than Opus 5 to attempt such circumvention. Stronger models tend to be more capable of finding workarounds, which makes the improvement meaningful if it holds. Deployment experience will be the real test.

What makes this launch unusual is its timing. Earlier this month, Anthropic CEO Dario Amodei published an essay in which he wrote that he had become convinced that fully addressing AI risks requires “pacing the rate of capabilities advancement so that risk prevention has time to keep up.” OpenAI’s Sam Altman and Elon Musk expressed agreement with the general concern. Nvidia’s Jensen Huang said the underlying science doesn’t support it. And then Anthropic shipped a model faster, cheaper, and more capable than its predecessor. One reading: a model with better safety evaluations and wider access is precisely what Amodei had in mind. Another: the essay named capabilities advancement as the concern, and capabilities advanced. Both readings sit in the same set of facts, and Anthropic’s benchmarks alone cannot resolve which is right.

Claude Opus 5.5 is available as of September 22, 2026 under the API model ID claude-opus-5-5. It runs inside Claude on Pro, Max, Team, and Enterprise plans, and is accessible through AWS, Google Cloud, and Microsoft Azure. Anthropic is raising usage limits on five-hour subscription tiers. Sonnet 5.5 and Haiku 5.5, the remaining members of the 5.5 family, are scheduled to follow in the coming weeks, with no pricing or specific release date announced for either.

Tags: Anthropic, AI safety, AI models, Dario Amodei, API, Claude Opus 5.5