Issue evidence · AI API · BerriAI/litellm

Anthropic /v1/messages does not translate stop_sequences for OpenAI-format models

GitHub issue: https://github.com/BerriAI/litellm/issues/33075

Not reproduced

A /v1/messages request with stop_sequences sent to an OpenAI-format provider forwards stop_sequences unchanged and the upstream rejects it.

What Badgr ran

Ran LiteLLM 1.104.2 as a proxy with an azure_ai deployment pointed at a local mock upstream that records request bodies, and sent a /v1/messages request with stop_sequences ["STOP"].

What came back

  • The upstream received a chat/completions request with "stop": ["STOP"] and no stop_sequences key, and the proxy returned a normal message.
  • An openai/ deployment was sent through the Responses API path instead, which has no stop parameter, and my mock did not answer that shape, so it says nothing about stop handling.

The reporter was on an earlier version against real Azure AI Foundry; this shows 1.104.2 translates the field, not that older versions did.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence