Skip to main content

​
Anthropic: Claude Opus 4.5 to 4.6

Use this guide when you are moving from anthropic/claude-opus-4.5 to anthropic/claude-opus-4.6. Claude Opus 4.6 is not just a model-ID swap. Anthropic’s migration guidance includes request-shape changes, removed beta flags, and stricter behavior around prefills.

​
What changed

  • assistant prefills are no longer supported on Claude 4.6 models and now return 400
  • structured output moved from output_format to output_config.format
  • adaptive thinking is now handled through output_config.effort
  • older beta headers should be removed
  • Opus 4.6 specifically no longer needs the interleaved-thinking-2025-05-14 beta header
  • Opus 4.6 adds a 1M-token context window in beta and improves long-running coding and agentic tasks

​
What to change in your integration

​
1. Remove assistant prefills

If you were forcing JSON or a response prefix by partially pre-seeding the assistant turn, that pattern now breaks on 4.6. Move that control into:
  • clearer system instructions
  • examples in the user or system prompt
  • structured outputs through output_config.format

​
2. Migrate structured outputs

If you still send output_format, update the request body to use output_config.format instead.

​
3. Review thinking configuration

If you were depending on older thinking betas, migrate to the current effort setting inside output_config and retest latency-sensitive routes.

​
4. Remove stale beta flags

Audit hard-coded beta headers in SDK wrappers, middleware, and gateway presets. Leaving obsolete flags in place creates noisy failures during rollout.

​
What to test

​
Prompt behavior

  • long coding and refactor prompts
  • multi-step research or document workflows
  • prompts that previously depended on assistant prefills
  • long-context tasks if you plan to use the 1M-token beta window

​
Contracts and tools

  • schema pass rate after moving to output_config.format
  • tool-call argument quality on multi-step agent tasks
  • retry behavior when the model refuses or reformats output
  • regression handling in code-review and debugging workflows

​
Operations

  • median and p95 latency with the new thinking settings
  • token consumption on longer runs
  • timeout rate on large-context tasks
  • success rate versus Opus 4.5 on your production eval set

​
Safe rollout

  1. Remove prefills and migrate structured output config before any traffic shift.
  2. Shadow Opus 4.6 on production prompts that matter to you.
  3. Canary traffic with Opus 4.5 still configured as a fallback.
  4. Promote only after schema pass rate, latency, and task completion stay stable.

​
Sources

Last modified on March 11, 2026