Anthropic released Claude Opus 5.5 on September 22, 2026. It lists at $4 per million input tokens and $20 per million output tokens, and Anthropic says it costs about 40% less to run than Claude Opus 5 on typical workloads. For teams already on Opus 5, the price is the easy part. The harder part is four breaking API changes that can return errors in code that works today.
Claude Opus 5.5 is the first model in Anthropic's Claude 5.5 family, followed by Claude Sonnet 5.5 on September 28 and Claude Haiku 5.5 on October 7. In its announcement, Anthropic says it performs at the level of Claude Fable 5.1 on most work, and the models overview now recommends starting with Opus 5.5 for most workloads, reserving Fable 5.1 for demanding reasoning and long-horizon agentic work. The same overview lists a 1M-token context window, 128K maximum output tokens, a June 2026 knowledge cutoff, and a retirement date no sooner than September 22, 2027. The API identifier is claude-opus-5-5, and the model is offered on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.
Anthropic's 40% figure combines two effects: lower per-token prices and fewer tokens per task. Only the first is fixed.
| Per million tokens (USD) | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache writes | $5 | $6.25 |
| Cache reads | $0.20 | $0.50 |
Source: Anthropic announcement pricing table, checked October 10, 2026.
The line that matters most for agents is cache reads, down 60%. An agent that re-reads a large cached context on every turn spends most of its tokens there, so the effective saving on agentic work can exceed the 20% cut on input and output. Workloads that do not cache benefit less.
The token-efficiency half of the claim deserves a check before it goes into a budget. The what's new page notes that the default effort level dropped from high on Opus 5 to medium on Opus 5.5, and that at a given effort level the model tends to think more per turn. Two deployments can therefore see opposite results: one that relied on the default gets a cheaper medium-effort model, while one that pinned high or max effort may see more thinking tokens, not fewer.
Anthropic's announcement says a fast mode, available in Claude Code and on the Claude Platform, runs at up to 2.5 times standard speed for $8 input and $40 output per million tokens. The what's new page is more specific for developers: it labels fast mode a research preview, available on the first-party Claude API only (requested with speed: "fast" and a beta header), and not on Amazon Bedrock, Claude Platform on AWS, Google Cloud or Microsoft Foundry.
Anthropic's what's new page lists four changes that return errors in requests Opus 5 accepted. Each is worth checking before switching the model ID.
thinking: {"type": "disabled"}, or a manual budget_tokens setting, returns a 400 error. Omit the field or send {"type": "adaptive"}, and use the effort parameter to control depth and cost. Code should select content blocks by type, not by position, because a response can start with thinking blocks.tool_choice of any or a named tool returns a 400 error; only auto and none work. Pipelines that force a tool call to get structured output need to move to strict tool use or structured outputs, or instruct the model in the prompt.computer_20251124 returns a 400 error there; integrations must move to computer_toolset_20260801. Anthropic says the older tool still works on Amazon Bedrock.A fifth change produces no error but can look like an outage to users: short notes the model writes between tool calls now arrive as thinking blocks, which are empty at the default display setting. An application that streams those notes as progress updates goes quiet until it changes the display setting. The migration guide has before-and-after request examples for each case.
Anthropic's announcement compares Opus 5.5 with GPT-6 Astra, GPT-5.6 Sol and its own models on several coding evaluations, including FrontierCode, Terminal-Bench 4.0 and CursorBench. OpenAI's current middle tier, GPT-6.1 Sol, was released on September 29, 2026, a week after Opus 5.5, so it is not in Anthropic's comparison. Anthropic reports similar or better scores at a fraction of Astra's cost per task at default effort. It also describes an internal C-to-Rust port of the HAProxy load balancer in which Opus 5.5 finished faster and more cheaply than Fable 5.1.
These are Anthropic's measurements, run with Anthropic's chosen settings, and this article does not reproduce them. The more useful signal for a buyer is the shape of the claim. Anthropic is competing on cost per completed task rather than peak score, so the evaluation that matters is your own: success rate and total spend per task on a representative sample of your work.
On safety, Anthropic says external evaluators including METR tested the model before release. The Claude Opus 5.5 system card documents those evaluations. The what's new page also notes that Opus 5.5 adds a biology safety classifier alongside the existing cybersecurity one, so some life-sciences requests may be declined that Opus 5 handled.
Teams on Claude Opus 5 get lower list prices on every token type; treat the move as a migration with a test pass, not a one-line change.
Teams weighing Opus 5.5 against Sonnet 5.5 should note that Sonnet 5.5 costs half as much on input and output and, since October 7, 2026, half as much on cache reads ($0.10 against $0.20). Anthropic positions Opus 5.5 for long, multi-step work. A reasonable starting point is to route long-running agent jobs to Opus 5.5 and interactive or high-volume work to Sonnet 5.5, then compare cost per completed task.
Teams comparing across vendors can read the GPT-6.1 Sol guide and the GPT-6 Astra overview. For a structural comparison of how Anthropic and OpenAI price long contexts and caching, see Claude Fable 5.1 vs GPT-6 Astra pricing structures.
claude-opus-5-5.thinking: disabled and manual thinking budgets; set output_config.effort explicitly.tool_choice with strict tool use or structured outputs.computer_20251124 integrations to computer_toolset_20260801 (Claude API and Google Cloud).When was Claude Opus 5.5 released?
September 22, 2026.
How much does Claude Opus 5.5 cost?
$4 per million input tokens, $20 output, $5 cache writes and $0.20 cache reads, per Anthropic's announcement (checked October 10, 2026). Fast mode, a research preview on the first-party Claude API, is $8 input and $40 output.
Can I turn off thinking on Claude Opus 5.5?
No. Anthropic's documentation says requests that disable thinking return a 400 error; use the effort parameter to reduce thinking instead.
What is the context window?
1M tokens, with up to 128K output tokens, per Anthropic's models overview.
Discover more content: