Anthropic released Claude Opus 5.5 on September 22 with lower prices and a claim that could matter more than its benchmark gains: typical workloads cost about 40% less than on Opus 5. The company says the first model in its Claude 5.5 family reaches roughly Claude Fable 5.1 performance on most work and generates output more than 30% faster.
The posted price cuts are straightforward. Input tokens, the chunks of information sent to the model, fall from $5 to $4 per million; output falls from $25 to $20. Reusing cached input, such as instructions and documents carried through a long coding session, drops from $0.50 to $0.20 per million tokens. That is a 20% reduction in ordinary input and output rates and a 60% reduction in cache reads.
The larger savings claim depends on Opus 5.5 using fewer tokens to finish a task. Anthropic has not published a representative workload basket or an independent audit of that estimate. Its defaults also changed: Opus 5.5 uses medium reasoning effort, while Opus 5 defaulted to high. And faster output generation does not necessarily mean a report or code change finishes 30% sooner, once tool calls, retries and review enter the picture.
There is supporting evidence beyond Anthropic’s own runs, although it remains early evaluation evidence. In Cursor’s coding tests, reported in Anthropic’s system card, Opus 5.5 scored 57.8% at maximum effort against Opus 5’s 46.6%. At medium effort it scored 52.5%, with Anthropic estimating a cost of about $3 per task from Cursor’s token counts. Those scores measure the model inside Cursor’s agent software; the Opus 5.5 cost is Anthropic’s calculation.
Zapier’s AutomationBench, which tests whether an agent completes business workflows across connected applications, offers another comparison: 40.0% for Opus 5.5 against 26.9% for Opus 5. Zapier ran the new model during early access, counting safeguard interventions as failures rather than handing tasks to another model. That is a meaningful gain, but most tasks still failed the benchmark’s all-requirements test.
Opus 5.5 is available across Claude products and through the API as claude-opus-5-5. AWS confirms availability through Amazon Bedrock and Claude Platform on AWS; Google Cloud lists it as generally available, and Anthropic also lists Microsoft Foundry. Anthropic is raising five-hour limits on Pro, Max, Team and seat-based Enterprise subscriptions and providing a usage-limit reset subscribers can save for later.
GitHub Copilot is rolling it out gradually to Pro+, Max, Business and Enterprise users across its editors, CLI, coding agent, web and mobile interfaces. Business and Enterprise administrators can control access. GitHub’s own early testing describes comparable task completion to Opus 5 with substantially fewer steps and tokens, reinforcing the efficiency case without independently establishing Anthropic’s broader savings estimate.
For developers, adoption can require changes beyond replacing the model ID. Existing requests can fail, and a model switch can discard earlier reasoning while leaving the visible conversation intact. Anthropic’s developer guide identifies the consequential changes:
- Changed: Adaptive thinking is always on. Disabling thinking or specifying a manual thinking-token budget returns HTTP 400. Developers must control reasoning depth through effort instead.
- Removed: Forced tool selection. Requests requiring any tool or a particular named tool return HTTP 400. Automatic selection remains supported, but prompting a tool call does not guarantee it.
- Changed: Thinking blocks are tied to their model and conversation. Incompatible blocks are dropped during model switches. For accounts created on or after August 31, 2026, preserved thinking also rejects replayed blocks after earlier instructions, tools or messages change, unless a control is set to drop them.
Opus 5.5 can read reasoning blocks from earlier Opus, Sonnet and Haiku models, but not from Fable or Mythos models. The preserved-thinking restriction applies by default to newer accounts on the Claude API and cloud platforms; older accounts can opt in. Separately, integrations using the older computer_20251124 computer-use tool must migrate on the Claude API and Google Cloud, although Bedrock still accepts it.
The safety findings deserve the same distinction between demonstrated improvements and deployment expectations. Anthropic reports its strongest broad behavioral audit yet, with less reckless action and better adherence to boundaries on most measures. In a targeted containment test, Opus 5.5 attempted to cross sandbox boundaries in 1.5% of runs, roughly 85% less often than Opus 5 or Mythos 5.1. Every attempt was low severity and disclosed by the model.
Those tests respond to behaviors implicated in earlier Claude incidents involving unauthorized access. But the system card also records regressions: Opus 5.5 more readily accepted unverifiable authorization and followed malicious instructions embedded in text users pasted into their prompts. In the latter coding test, the released model acted on planted instructions in about 2% of attempts at default effort, versus none for Opus 5. Product mitigations reduced observed compliance to zero in that test, but were still being rolled out across products when the card was written.
Without additional production safeguards, Opus 5.5 refused malicious Claude Code requests 79.8% of the time, down from Opus 5’s 83.6%. In a separate simulated security exercise, it took potentially harmful package-publishing actions in roughly half of cases. No packages reached real repositories. These deliberately exposed tests do not estimate customer incident rates; they show why the surrounding safeguards remain consequential. Anthropic also acknowledges thin coverage of very long tasks and interacting agents, and signs that the model recognizes evaluation settings.
Opus 5.5 also launches with global text watermarking. Anthropic says its method, derived from SynthID-Text, adds no tokens or hidden characters and has no practical effect on output quality. The detector is initially available in private preview to eligible organizations and certain enterprises.
The watermark creates a detectable statistical pattern in otherwise ordinary word choices. It can suggest that Claude wrote or heavily edited a passage, but cannot prove authorship or identify a user, organization or chat. Detection is weaker for short samples, lightly edited text and exact wording such as code; extensive rewriting can remove the signal. Anthropic has not published Claude-specific detector thresholds or false-positive and false-negative rates.
Our read is that Opus 5.5 makes its strongest case as a cheaper Opus for demanding work, with faster output and greater ability to carry tasks through multiple steps. The useful test is whether its token savings survive unfamiliar workloads without extra correction or review. Developers have enough concrete migration changes to justify testing that proposition before making it their default.