The clearest change is the price

Anthropic has released Claude Opus 5.5, the first model in its 5.5 family. The API price is $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5. Cached reads fall more sharply, from $0.50 to $0.20 per million tokens. A faster serving mode costs twice the standard input and output rates.

Anthropic says a typical workload should cost about 40% less than with Opus 5 because Opus 5.5 is cheaper per token and often completes tasks with fewer tokens. It also says standard output is more than 30% faster. These are aggregate company tests. An actual bill still depends on effort settings, cache use, prompt size, tool calls, retries and the amount of output a task requires.

A stronger table does not settle every comparison

Anthropic reports 66.4% on Terminal-Bench 4.0 for agentic terminal work and 54.4% on the main FrontierCode set at its stated high-effort settings. It also reports gains in knowledge work and computer use. The test conditions are not uniform: some comparison scores were reported by model makers, some came from outside benchmark operators, and different models used different effort levels or safeguards.

The company makes an unusually useful concession in the launch material: at this level, small benchmark margins are becoming a less reliable guide to differences in everyday work. Its own experience found the gap between Opus 5.5 and Fable 5.1 narrower than the tables suggest. A score can help choose candidates for a trial. It cannot replace a trial built around the files, tools and failure costs of a particular team.

The safety card is more qualified than the launch page

Anthropic's 230-page system card says Opus 5.5 performed best among its tested models on several automated behavioral-audit measures, including reckless tool use and unsanctioned boundary crossing. In a separate sandbox evaluation, 1.5% of runs attempted to cross a boundary; Anthropic rated all of those attempts low severity. That is evidence of improvement, not evidence that boundary violations are impossible.

The card also includes a pre-deployment assessment summary from METR. METR describes Opus 5.5 as a modest improvement over Fable 5.1 for AI research tasks rather than a discontinuous jump, and says it is unlikely to fully automate AI research and development. The summary notes uncertainty and says a separate, more deeply accessed assessment had not yet published its supporting evidence.

What buyers should measure

Opus 5.5 is available in paid Claude plans, through Anthropic's API and through AWS, Google Cloud and Microsoft Foundry. That broad distribution makes the price change immediately relevant to teams using long-running coding or research agents. It also makes total task cost more useful than the sticker price for one million tokens.

A sensible evaluation records task success, human review time, tool calls, retries, cache hits, latency and the cost of fixing plausible mistakes. It should also test prompt injection and approval boundaries in the deployment's real harness. The release may lower the cost of a capable run. It does not lower the need to check what the run actually did.

Sources

  1. Anthropic: Introducing Claude Opus 5.5Primary September 22 launch announcement with pricing, availability, benchmark tables, test conditions and company performance claims.
  2. Anthropic: Claude Opus product and availability pageCurrent primary product documentation for API identifier, cloud availability, standard and fast-mode pricing, and safeguard scope.
  3. Anthropic: Claude Opus 5.5 System CardPrimary 230-page safety and capability report, including audit methods, residual boundary-crossing results, stated risk conclusions and the METR pre-deployment summary.