The current facts
Anthropic lists Claude Opus 5.5 as the latest Opus model, with the API model ID claude-opus-5-5. The official model page lists a 1M-token context window, a 128K maximum output, adaptive thinking that is always on, a default medium effort and a June 2026 knowledge cutoff.
Anthropic lists $4 per million input tokens and $20 per million output tokens. A 5-minute cache write is $5/MTok, a 1-hour cache write is $8/MTok and cache reads are $0.20/MTok. These are official split rates, not a blended reseller price; calculate a real workload before claiming a saving.
What is actually new for developers
Opus 5.5 is positioned for long-running agentic coding and knowledge work rather than only short chat. The headline capability is not just a larger context: the model is designed to keep tool-driven work coherent across long tasks, while adaptive thinking controls how much reasoning is used for the request.
The model has four important behavior changes compared with Opus 5. First, thinking cannot be disabled. Second, forced tool use returns an error. Third, thinking blocks are tied to the model and conversation. Fourth, the older computer_20251124 computer-use tool is not accepted on the Claude API and Google Cloud. Text between tool calls can also arrive in thinking blocks whose text is empty at the default display setting, so streaming clients should select content by block type instead of assuming every block contains visible text.
How it compares with the current Claude lineup
Anthropic's model table positions Opus 5.5 between Fable 5.1 and Sonnet 5: Fable is more expensive and slower for demanding reasoning, Sonnet is faster and cheaper for general work, and Opus 5.5 is the middle choice for long-running coding and knowledge workflows. The listed rates are $10/$50 for Fable 5.1, $4/$20 for Opus 5.5 and $2/$10 for Sonnet 5, input/output per million tokens.
That table is a routing hint, not a quality guarantee for your application. Run a small representative evaluation with your own codebase, tool schema and output acceptance tests. Compare completed tasks, retries, tool-call count, latency and total token cost rather than only the model name.
Migration from Claude Opus 5
The basic migration is a model ID change from claude-opus-5 to claude-opus-5-5. Do not stop at a string replacement: review thinking configuration, forced tool-use paths, block parsing and computer-use integrations before deploying.
A safe rollout starts with a canary key and a small spend cap. Send one non-sensitive JSON request, then one streaming request with a tool schema if your application uses tools. Verify that your parser handles thinking, text and tool-use blocks, that the terminal usage is recorded and that your retry logic cannot duplicate an already completed request.
A fair cost example
Suppose one task uses 200,000 input tokens and 20,000 output tokens without cache. At the official rates, the estimate is 0.2 × $4 + 0.02 × $20 = $1.20. If 150,000 input tokens are cache hits, the estimate changes materially because cache reads are billed at $0.20/MTok; the write cost and cache lifetime must also be included.
A prepaid gateway may publish one blended charge per million charged tokens instead. Never put that number next to $4 input or $20 output and call the smaller number cheaper. Define the token mix, cache operations, minimum charge, retry policy and currency first, then calculate the same workload under both billing formulas.
Is Claude Opus 5.5 available through Model.sale?
Model.sale only exposes a model to customer traffic after a live JSON/SSE compatibility check, terminal usage verification, an active public price and a margin review. An official Anthropic release does not automatically make a model live in this catalog.
Check the current Models page and GET /v1/models immediately before integration. If claude-opus-5-5 is not listed as live, do not send production traffic to it through an alias or silently substitute another Claude model. Use the explicit unavailable state and wait for a completed validation report.
Production checklist
1. Pin the exact model ID. 2. Confirm the account and endpoint support the model. 3. Test JSON and SSE with a synthetic prompt. 4. Parse thinking, text and tool-use blocks by type. 5. Test a long context and a cancellation. 6. Verify input, output, cache and reasoning usage. 7. Compare the final ledger charge with the expected formula. 8. Roll out behind a small key limit and monitor stream failures and P95 time to first token.
Keep the announcement date and official source URL in your internal price record. Re-check pricing, model lifecycle and computer-use compatibility after Anthropic changes the model family; these details are operational inputs, not permanent marketing copy.