Anthropic has released Claude Opus 5.5 with a one-million-token context window, lower published token rates and adaptive thinking that cannot be disabled. Existing agent workflows may need code and acceptance-test changes before migration.
RELATED BUYER PROFILEReview Claude Opus 5.5 pricing, migration limits and test-pending evidencePricing · access · strengths · limitations →RELATED DECISION GUIDERun a same-work migration pilotWorkflow · evidence · risk · governance →CONTINUE THE DECISIONAdd Claude Opus 5.5 to a current shortlistEvidence · workflow · next action →CONTINUE THE DECISIONSet tool, data and approval boundariesEvidence · workflow · next action →CONTINUE THE DECISIONMeasure complete agent-task costEvidence · workflow · next action →CONTINUE THE DECISIONOpen the developer decision deskEvidence · workflow · next action →What you need to know
- The Claude API model ID is claude-opus-5-5, with a one-million-token context window, 128,000-token output limit and June 2026 knowledge cutoff
- Published standard API pricing is $4 per million input tokens and $20 per million output tokens; Fast mode is $8 and $40
- Adaptive thinking is always enabled, forced tool use can return an error, thinking blocks are model-specific and the older computer-use tool version is not accepted on two routes
The verified migration boundary
Short answer: lower rates do not make this a drop-in upgrade
Anthropic released Claude Opus 5.5 on 22 September 2026 for long-running agentic coding and professional knowledge work. The model is available through the Claude API, AWS, Google Cloud and Microsoft Foundry, with product access across Claude plans. The public model identifier, price and limits are verified. Anthropic's benchmark, speed, cost and safety comparisons are vendor evidence, not an independent HubAI result.
The published standard rate is $4 input and $20 output
Claude Opus 5.5 lists $4 per million input tokens and $20 per million output tokens. Cache reads cost $0.20, five-minute cache writes $5 and one-hour cache writes $8 per million tokens. Anthropic lists a 50% Batch discount. Fast mode is priced at $8 input and $40 output and is described as up to 2.5 times faster; that speed figure is provider-reported and must be measured on the buyer's route.
Adaptive thinking is always on
Opus 5.5 always uses adaptive thinking. The API defaults to medium effort and does not accept a disabled-thinking configuration. Workflows that assumed no reasoning tokens, fixed visible text or a simpler latency profile need to be re-measured. Thinking blocks are tied to the model and conversation, so a multi-model system cannot blindly replay them across model boundaries.
Forced tools and computer use need migration checks
Anthropic says forced tool use returns an error with Opus 5.5. It also says the older computer_20251124 tool is not accepted through the Claude API or Google Cloud, although AWS Bedrock continues to accept it. Text that appears between tool calls may now be placed in thinking blocks and remain empty in the displayed response by default. These are implementation changes, not footnotes: they can affect orchestration, logging and acceptance tests.
One million tokens still need retrieval discipline
The model page lists a one-million-token context window and 128,000-token maximum output. Long contexts can increase latency, token exposure and the chance that irrelevant instructions compete with the current task. Buyers should compare retrieval and structured compaction with whole-corpus prompting, and measure cache reads, cache writes, retries and reviewer time rather than assuming the largest window is the best workflow.
Availability is broad; controls still differ by route
Anthropic lists Claude API, AWS, Google Cloud and Microsoft Foundry routes and describes availability across Claude plans. It also says zero data retention is available. Those statements do not make every account, region or enterprise control identical. Buyers must verify the exact contract, retention, residency, administrator policy and logging for the route that will handle company data.
The launch evidence is a pilot input
Anthropic reports lower typical workload cost than Opus 5, faster output, stronger coding and professional-task results, and external evaluation by METR and Frontier Design. These sources are useful for selecting difficult pilot tasks. They do not certify a buyer's repository, security boundary or regulated workflow, and HubAI does not reproduce the claims as independent measurements.
HubAI buyer verdict
Claude Opus 5.5 belongs on the shortlist for long agent workflows where a lower Opus rate could materially affect complete cost. It should not replace a production model until orchestration, thinking, tool-choice, computer-use and logging behaviour pass a same-work test. HubAI assigns no numerical score before that independent evaluation.
Evidence limits and editorial disclosure
Independent editorial coverage; not sponsored. Pricing, specifications, availability, benchmark and safety statements are attributed to Anthropic and its official documentation, checked on 23 September 2026. HubAI has not independently tested Claude Opus 5.5. The visible UK Google Trends first 25 contained no direct Opus 5.5 query during this review, so no search-volume claim is made. The original cover is a conceptual workflow illustration, not a product interface, Anthropic logo or benchmark chart.
HUBAI VIEWDo not treat a lower token rate as a lower task cost. Re-run long agent workflows with the new thinking, tool-choice and computer-use behaviour before changing production traffic.
Buyer decision signal: New frontier model · pricing and migration changes
What to verify next
1Pin claude-opus-5-5 in a test environment before changing production traffic
2Remove any assumption that thinking can be disabled
3Replace forced tool-choice logic and verify the supported computer-use tool version
4Check multi-model handoffs for incompatible thinking blocks
5Measure accepted tasks, tool failures, latency, cache use, review time and complete cost
6Verify the contracted cloud route, region, retention, logs and administrator controls
7Keep approval gates, rollback and the previous model available during migration
Read the evidence
Capabilities, availability and prices can change. HubAI keeps analysis separate from the underlying official material.
01Anthropic: Claude Opus 5.5 — 22 September 2026Open source ↗02Anthropic developer documentation: Claude Opus 5.5 overviewOpen source ↗03Anthropic migration guide: What's new in Opus 5.5Open source ↗04Anthropic model pricingOpen source ↗
Models
Models
Models
Models