Anthropic Releases Migration Guide for Claude Opus 5.5

Anthropic has released the official migration and prompting guide for the Claude Opus 5.5 model. Opus 5.5 is optimized for long-horizon agentic coding and knowledge work, offering output token generation speeds that are more than 30% faster than Opus 5 while using fewer tokens for the same tasks. Pricing is $4/1M tokens for input and $20/1M tokens for output, which is cheaper than Opus 5 ($5/$25).

Key API Structural Changes

When migrating to Opus 5.5, the following items must be modified in existing code to prevent 400 errors.

• Forced Thinking Activation: The thinking: {"type": "disabled"} option has been removed and is always enabled. Instead, it is controlled via the effort parameter (low, medium, high, xhigh, max), with the default value changed from high to medium.
• Tool Choice Restrictions: any and tool types are rejected; only auto or none are allowed. If forced tool calls are required, they must be replaced with Strict tool use or prompt instructions.
• Sampling Parameter Removal: temperature, top_p, and top_k are rejected if set to values other than the defaults.
• Prefill Removal: Messages ending with an assistant turn are rejected and must be replaced with Structured Outputs or system prompts.
• Computer Use Tool Changes: Only computer_toolset_20260801 is supported on the Claude API and Google Cloud; the previous computer_20251124 is rejected (Amazon Bedrock continues to support the existing tool).

Response Handling and Migration Cautions

• Thinking Block Position: A thinking block may appear first at the start of the response, so position-based (content[0]) reading code must be modified to rely on the type field.
• Tool Loop Maintenance: The thinking block from the immediately preceding assistant response must be returned unchanged; editing it causes errors.
• Thinking Text Display: With the default display: "omitted", the thinking field is empty. If summaries are needed, set display: "summarized" or the beta feature "updates".
• Cross-Model Compatibility: Thinking blocks from Opus 5.5 can be read by Opus 5 and earlier models, but thinking blocks may be ignored when routing to models other than Fable 5.1/Mythos 5.1.

Prompting Guide and Performance Tips

• Effort Adjustment: The medium effort level in Opus 5.5 provides performance equal to or better than the high effort level in Opus 5. It is recommended to start with low or medium and measure to reduce cost and latency.
• Unattended Agent Stability: To prevent early termination, add checklist-based continuation instructions to the system prompt and strengthen the handling logic for stop_reason: "end_turn".
• Safety Guardrail Refusals: When stop_reason: "refusal" occurs, policy categories (cyber, bio, reasoning_extraction, etc.) can be checked via the stop_details object, and server-side fallback functionality can be utilized.