Anthropic introduced Claude Opus 5.5 on September 22, 2026, presenting it as the first model in the Claude 5.5 family. The company says it performs at roughly Claude Fable 5.1’s level on most work, costs 40% less to run than Opus 5 on typical workloads, and produces output more than 30% faster. Anthropic also says external evaluators including Frontier Design and METR tested the model before release and that it achieved the company’s strongest result yet on an automated behavioral audit. Those are Anthropic’s claims; the announcement does not include the evaluators’ underlying reports or establish that they independently reproduced every result.
The clearest part of the announcement is the published price schedule. Opus 5.5 charges $4 per million input tokens, down from $5 for Opus 5, and $20 per million output tokens, down from $25. Cache reads fall from $0.50 to $0.20 per million tokens, while cache writes decline from $6.25 to $5. Those list-price reductions are 20%, 60%, and 20%, respectively. They make lower bills plausible, but they do not by themselves prove Anthropic’s broader 40% figure: total cost also depends on how many tokens, cache operations, tool calls, and retries a task requires.
That distinction matters for coding agents and other systems that solve one request through many model calls. Anthropic says Opus 5.5 uses fewer tokens per task as well as charging less per token, producing the claimed net cost reduction. It cites early-tester examples including a 680,000-line migration completed in under a day and an audit of a 200,000-line codebase completed in under three hours. The company also reports an internal HAProxy translation test in which Opus 5.5 finished in 9.5 hours, versus 12 hours for Fable 5.1, at 51% lower cost. These examples illustrate possible gains, but they are selected cases rather than a controlled comparison across ordinary customer workloads.
The benchmark section requires particular caution because the primary page presents inconsistent figures. One table lists Opus 5.5 scores of 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 57.8% on CursorBench 4.0. Later prose gives FrontierCode as 54.6% and CursorBench as 52.5%, without explaining the discrepancy. The page also notes that some results used different effort levels, that production safeguards could trigger failures, and that other models’ figures were taken from their publishers or public leaderboards. Readers therefore have evidence that Anthropic tested the model across several agent benchmarks, but not a single internally consistent set of scores from which to rank every system precisely.
Anthropic’s safety argument extends beyond benchmark totals. The company says Opus 5.5 screens every proposed action with a classifier, runs code in an open-source sandbox that security teams can inspect, and reviews code for vulnerabilities before changes are merged. It further claims stronger resistance to prompt injection across coding, tool use, computer use, and web browsing, citing a Gray Swan evaluation in which Opus 5.5 tied Fable 5.1 for the lowest injection success rate among tested models. These details describe concrete controls, but the announcement does not provide the classifier’s performance data, sandbox audit results, vulnerability-detection rates, or Gray Swan’s full methodology.
The release also arrives in the context of Anthropic chief executive Dario Amodei’s separate essay calling for the frontier AI industry to slow capability advancement long enough for safety work and third-party evaluation to catch up. Amodei proposes embedded external evaluators, common standards among companies in democratic countries, and eventual global coordination. Opus 5.5’s launch shows that “pacing” does not mean stopping releases: Anthropic describes this model as its first since that proposal, while retaining restricted access for sensitive biology and cybersecurity uses through verification programs. Because both the launch page and the pacing essay come from Anthropic’s leadership, the essay supplies context for the company’s stated policy—not independent confirmation that its release tempo or safeguards satisfy that policy.
The essay’s proposal for embedded evaluators, common standards, and global coordination provides policy context for Anthropic’s latest release.
For buyers, the practical question is not whether Opus 5.5 is cheaper per token; the published schedule says that it is. The unresolved question is how often the lower prices and reported token efficiency translate into lower end-to-end costs for a specific workflow. A defensible evaluation would run the same representative tasks against the previous model, record total tokens and cache use, count tool calls and failed attempts, inspect the outputs, and include the cost of human review. Until comparable results are available, Anthropic’s 40% figure is best treated as a company estimate supported by selected tests—not a guaranteed saving or proof of a general capability lead.