Skip to main content

Summary

Claude 4.6 and later models, and Claude Mythos Preview, reject prompts that end with an Assistant message. In a Braintrust playground, you may see: This model does not support assistant message prefill. The conversation must end with a user message.

Affected models

Anthropic’s current model list includes:
  • Claude Opus 4.6, 4.7, 4.8, 5, and 5.5.
  • Claude Sonnet 4.6, 5, and 5.5.
  • Claude Haiku 5.5.
  • Claude Fable 5, 5.1, and 5.2.
  • Claude Mythos Preview, 5, 5.1, and 5.2.
The restriction also applies to later Claude models, even if they are not listed here.

Fix

  1. Open the failing playground prompt and check the final message in Messages.
  2. If that message is intended as an instruction to Claude, change its role from Assistant to User. If it represents a previous assistant response, add a User message after it instead.
  3. Run the prompt again.
Only a final Assistant message is treated as the start of Claude’s response, called a prefill. Earlier Assistant turns can remain in the conversation. The restriction also applies when Claude is accessed through an OpenAI-compatible gateway. See Anthropic’s API error reference.