Claude Opus 5.5: Pricing, Benchmarks and API Changes

Claude Opus 5.5 is Anthropic’s September 22 model release for long-running agentic coding and knowledge work. Its official API ID is claude-opus-5-5, with a 1M-token context window and 128K standard maximum output. Standard API pricing is $4 per million input tokens and $20 per million output tokens; cache reads and writes have separate rates. Anthropic reports about 40% lower cost on its typical workloads than Opus 5, but that is a vendor evaluation of completed tasks, not a fixed discount on every request. Before upgrading, review Opus 5.5’s always-on thinking, tool-choice and computer-use changes.

The release targets coding agents and knowledge work, and is offered through the Claude API and supported cloud platforms. This page separates Anthropic’s benchmark claims from the implementation and billing details that developers can verify in the current docs. Copilot access and AI Credits are a separate product question covered in our Claude Opus 5.5 in GitHub Copilot guide.

Model, access and core specs

Anthropic describes Opus 5.5 as suited to long-running agentic coding and knowledge work. The current model page lists text and image input, text output, a June 2026 knowledge cutoff, 1M context and 128K maximum output. The documented retirement commitment is no earlier than September 22, 2027.

The standard API model has adaptive thinking on by default, with default effort set to medium. The Claude subscription apps have their own plan limits: Anthropic says five-hour usage limits increased for Pro, Max, Team and seat-based Enterprise users. Those subscription limits are separate from the per-token API rates below.

Anthropic API pricing and a request-cost example

Anthropic’s current price table distinguishes cache reads from cache creation. A 5-minute cache write costs 1.25 times the base input price; a 1-hour write costs twice the base price. A cache read is a separate, lower rate:

For a worked example, assume one request is billed as 120,000 uncached input tokens, 50,000 cache-read tokens, 20,000 tokens written to the 5-minute cache, and 10,000 output tokens. At the listed rates, the arithmetic is $0.48 + $0.01 + $0.10 + $0.20 = $0.79. If those 20,000 cache-write tokens use the 1-hour duration instead, the request totals $0.85. This is an illustrative calculation from Anthropic’s rates, not a measured session; agent workloads make many calls, and their total changes with token use, cache reuse and output length.

Fast mode is a separate first-party Claude API research preview. Its input and output rates are $8 and $40 per million tokens, respectively; Anthropic says it can generate output up to 2.5 times faster. The preview is not available on Amazon Bedrock, Google Cloud, Microsoft Foundry or Claude Platform on AWS. It is not a Copilot feature or price. For batch API calls, check the current price table and model-specific constraints rather than applying the Fast rate or cache multipliers by assumption.

Anthropic’s “about 40% less” typical-workload claim combines lower per-token rates with fewer tokens used to complete the vendor’s evaluated tasks. The standard input and output rates themselves are each 20% below Opus 5 ($4/$20 versus $5/$25); cache reads are 60% lower ($0.20 versus $0.50). The realized task-cost change depends on how many calls and tokens your own workflow needs. For Opus 5 launch-era evidence and the older model’s unknowns, see our Opus 5 launch guide and the earlier Opus 5 benchmark and pricing comparison; those pages retain their original model intent.

What Anthropic’s benchmarks say

Anthropic reports the strongest results on agentic coding, computer use and knowledge work, while warning that small benchmark margins do not reliably predict real-world differences. These are the company’s own results and evaluation notes, not an independent DMT test:

Unless noted, the comparison table reports Claude Opus 5.5 at maximum effort; Terminal-Bench 4.0 uses xhigh. The separate cost charts use default medium effort for some results. Anthropic also says its system performed benchmark tasks with production safeguards enabled; some cyber, biology and frontier-model-development tasks were completed by fallback Opus models when safeguards intervened. The result therefore describes a protected deployed system, not an unrestricted model-only run.

Anthropic’s reported task examples, early-customer statements and benchmark scores are useful starting points for evaluation. They do not guarantee a similar improvement in your repository, agent harness or knowledge-work workflow. For a broad cross-vendor comparison of Astra, Fable and Gemini, see our existing frontier comparison; it does not claim to benchmark Opus 5.5.

API changes to check before upgrading from Opus 5

The model ID is simple to change, but the API behavior is not identical. Anthropic documents these upgrade checks:

1. Thinking is always on. Remove thinking.type: disabled and manual budget_tokens settings; Opus 5.5 returns a 400 error for them. Omit the thinking field or use adaptive thinking, then set output_config.effort to control depth. Recalibrate because the default is medium here, compared with high on Opus 5.
2. Forced tool choice changed. The old tool_choice values any and tool return 400 errors. Anthropic’s migration example uses auto, strict tool use (strict: true for each tool), and an instruction in the prompt about when the tool applies. This validates arguments but does not guarantee the model invokes a tool on every turn, so test the control flow.
3. Thinking blocks are tied to the model and conversation. In tool loops, pass the returned blocks back unchanged and do not carry them across an incompatible model or conversation. On the Claude API and cloud platforms, Anthropic enforces the preceding-message, tool and system integrity check by default for accounts created on or after August 31, 2026.
4. Computer-use tools differ by platform. On the Claude API and Google Cloud, remove the older computer_20251124 beta header/tool and declare computer_toolset_20260801 without a name or display dimensions. Update the agent loop to read the action from each tool_use block’s name, handle several blocks per turn, and echo toolset_name in every result. Bedrock continues to accept the older tool.
5. Progress text can arrive in a different block. Text between tool calls is returned in thinking blocks and is empty at the default display setting. Read blocks by type, then render non-empty progress blocks before the next tool call. If the UI needs those updates, the migration guide documents thinking.display: updates (beta with its header) or summarized; pass the blocks back unchanged in the tool loop.

These are integration changes, not just prompt wording. Test tool-use loops, streaming clients, and saved conversation replay on the exact provider surface you use before switching production traffic. The official Opus 5.5 migration guide and current API change notes contain the supported request shapes. No code sample here was executed.

Safeguards and what the launch does not prove

Anthropic says Opus 5.5 launched with safeguards similar to Fable 5.1 for cybersecurity, biology and distillation. Depending on the request, the system can refuse or route work to a different model. The API can return a refusal in an HTTP 200 response, with a refusal stop reason and policy details; clients should not interpret a successful HTTP status as proof that Opus 5.5 completed the task.

Anthropic’s claims about cyber and life-science safeguards, verification programs and retention choices are a broader governance topic. For the existing Fable 5.1 enterprise safeguards and data-retention owner, see our Fable 5.1 safeguards guide. This Opus 5.5 article keeps the relevant fallback behavior in view without replacing that page’s enterprise decision checklist.

Specifications, availability and rates reflect Anthropic’s official documentation checked September 23, 2026. Benchmark values and typical-workload savings are attributed to Anthropic or the named evaluation source. No Anthropic account or API request was tested, and no independent benchmark was run for this article.

Sources

• Anthropic: Introducing Claude Opus 5.5
• Claude Platform Docs: Opus 5.5 model specifications and availability
• Claude Platform Docs: current model, cache, Fast mode and batch pricing
• Claude Platform Docs: Opus 5.5 behavior changes and supported features
• Claude Platform Docs: migrating from Opus 5
• Anthropic: September 1 Fable 5.1 and Mythos 5.1 launch context