Claude Opus 5.5 always has adaptive thinking enabled. When migrating, use effort to control reasoning depth and check compatibility with forced tool calls, thinking-block handoff, and computer-use tools.
Suitable tasks: Migrating integrations from Claude Opus 5 or older Claude APIs to Opus 5.5; checking tool calls and thinking-block handling in long-running agents; confirming model IDs and feature support across cloud platforms.
Not suitable for: Inferring how the model will perform in a specific business context; writing a complete prompt based on this page; or implementing every API flow in place of the relevant feature documentation.
Applicable model version: Claude Opus 5.5, with the Claude API model ID claude-opus-5-5.
Applicable clients, agents, or APIs: Claude API, AWS Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Platform support varies by feature; check the limitations below.
Recommended reasoning level and parameters: Omit thinking or use {"type":"adaptive"}; control the reasoning level through output_config.effort. The default is medium. Set effort explicitly, then remeasure latency, cost, and quality for the target task.
This source is an API configuration and migration guide; it does not provide a complete, copyable prompt. Its core configuration guidance is:
Set the requested model to claude-opus-5-5; see “Original evidence and data” for cloud-platform model identifiers.
Do not send thinking: {"type":"disabled"} or thinking: {"type":"enabled", "budget_tokens": N} with a manually set token budget; either request returns HTTP 400. Omit thinking or set it to {"type":"adaptive"}.
Do not use tool_choice in the any or specific tool form. Use auto. For structured JSON, enable strict: true on the tool schema or use structured outputs.
Thinking blocks in conversation history are tied to the model that generated them. In tool loops, identify content by each block’s type and pass thinking blocks back unchanged; do not rely on a fixed content-block position.
If using the older computer_20251124 tool with Claude API or Google Cloud, migrate to computer_toolset_20260801, remove the old beta header, and update the agent loop to handle tool_use blocks of type member, batched actions, and toolset_name in results.
If the application displays progress text between tool calls, check the thinking-block display setting. With the default thinking.display: "omitted", text in those blocks is not visible.
Update the model ID to claude-opus-5-5.
Find and remove disabled thinking and manual budget_tokens thinking configuration. Set output_config.effort for the task, and begin evaluation from the medium baseline.
Check tool_choice in requests. Change any or a specific tool selection to auto, and explain in the prompt when tools should be used. If a strict schema is needed, use strict: true or structured outputs.
Check response handling and history serialization: parse blocks by type, pass thinking blocks back unchanged in subsequent tool requests, and decide whether progress text between tool calls needs to be shown; set thinking.display accordingly.
If using computer use, confirm the tool type for the deployment platform: Claude API and Google Cloud use computer_toolset_20260801; Bedrock still supports computer_20251124. Integrations already using the new toolset or browser use do not need to migrate for this item.
Rerun integration tests and an effort sweep for the target platform and real workload. Do not carry over Opus 5’s default level or tool assumptions.
Anthropic positions Opus 5.5 as a model for long-running agentic coding and knowledge work. Its Claude API ID is claude-opus-5-5.
Thinking is always enabled in Opus 5.5. Disabling thinking or specifying a manual thinking-token budget returns a 400 invalid_request_error. Anthropic recommends omitting the thinking field or using adaptive thinking, and controlling reasoning with effort.
tool_choice: {"type":"any"} and tool_choice: {"type":"tool", "name":"..."} are unsupported; auto (the default) and none are supported. For schema-valid JSON, the documentation also points to auto with strict tool use, or structured outputs.
Thinking blocks record which model generated them. Opus 5.5 can read thinking blocks from Opus 5 and earlier Opus, Sonnet, and Haiku models; when switching to other models, reasoning content from the previous model may be lost. Blocks that cannot be read are discarded before the model sees the request and are not billed. Eligible accounts also, by default, check whether system prompts, tools, or messages before a block have changed. See the original source for account creation dates and beta behavior.
The old computer_20251124 tool is not accepted on Claude API or Google Cloud; it remains available on AWS Bedrock. The official migration target is computer_toolset_20260801.
Officially listed support includes per-message effort (beta), system messages in conversations, task budgets, prompt caching with a minimum of 512 cacheable tokens, batch processing, Files API, PDF, vision, and server-side and client-side tools. Exceptions for computer use vary by deployment platform.
Fast mode (research preview) is available only on Claude API. It uses speed: "fast" and the fast-mode-2026-02-01 beta header.
Inline tool definitions use the inline-tools-2026-09-15 beta header; on-demand compaction uses the compact-2026-09-04 beta header. Each has its own platform and model availability; check the linked feature documentation before integrating either.
Model IDs: Claude API: claude-opus-5-5; Amazon Bedrock: anthropic.claude-opus-5-5; Claude Platform on AWS, Google Cloud, and Microsoft Foundry: claude-opus-5-5.
The documentation lists prices of $4 per million input tokens and $20 per million output tokens; batch pricing is $2 / $10. Prices and feature support may change, so check the current platform pricing page.
This is Anthropic’s official compatibility and migration guide, not an independent performance evaluation. It does not publish a reproducible task set, complete request examples, or a results table.
The documentation does not give a publication date; the collection date above is the date this information was checked. Recheck prices and platform support at deployment time.
Tool support differs across cloud platforms. In particular, do not generalize the computer-use tool limitations for Claude API / Google Cloud to Bedrock.
The default effort level differs from Opus 5, and thinking usage may also change at the same level. Calibrate with your own evals; this information alone does not establish that quality will be higher on any task.
Anthropic also describes differences in safety refusals and chart recognition on this page; these do not constitute independent evidence of performance for this configuration entry.
Prepare an existing Opus 5 request on Claude API and on each target cloud platform. Check the model ID, thinking, tool_choice, round-trip serialization of thinking blocks, and computer-use tool type. After applying the migration guidance on this page, run request-level integration tests for successful paths and expected rejection paths. Record the platform, SDK, request parameters, response status, stop reason, latency, and token usage. Compare at least medium with the actual candidate effort levels, using the same task set and tool environment.
Claude Opus 5.5
