8 views
-/https://github.com/berriai/litellm/issues/33294
GitHub · issue

#33294 [Bug]: Anthropic /v1/messages treats Fable 5 server-side fallbacks as LiteLLM fallbacks

  • State: open
  • Author: @wwwillchen
  • Labels: proxy, llm translation

## Check for existing issues

- [x] I searched existing issues and did not find a duplicate.

#24004 is related, but covers mid-stream SSE errors rather than Fable's HTTP-200 refusal response.

## What happened?

Anthropic supports server-side fallback when Fable 5 returns `stop_reason: "refusal"`:

```http anthropic-beta: server-side-fallback-2026-06-01 ```

```json { "fallbacks": [{ "model": "claude-opus-4-8" }] } ```

Through LiteLLM's `/v1/messages` endpoint, `fallbacks` appears to be interpreted as LiteLLM router configuration instead of being forwarded to Anthropic.

With `claude-opus-4-8`, LiteLLM returns a model-authorization error. With the configured LiteLLM alias `anthropic/claude-opus-4-8`, the request succeeds but the original Fable refusal is returned without running the fallback.

I expected LiteLLM to pass Anthropic's native fallback parameter through when the server-side fallback beta header is present.

Anthropic documentation: https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback

## Steps to Reproduce

1. Configure both models:

```yaml model_list: - model_name: anthropic/claude-fable-5 litellm_params: model: anthropic/claude-…

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (2 events)
#0 of 0 · 31d18h18m42s ago — entered · #import:https:::github.com:berriai:litellm post #2633
The left issue is harder because it requires context-sensitive request translation, preserving provider-native behavior while avoiding interference from LiteLLM routing semantics, plus compatibility and regression testing across proxy paths. The right issue is comparatively localized to sanitizing unsupported parameters during Anthropic payload construction.
#0 of 0 · 31d18h44s ago — current · #import:https:::github.com:berriai:litellm post #2942
The right issue requires changes across proxy request translation, provider-specific parameter handling, header-gated behavior, and regression coverage, with compatibility risks around existing routing semantics. The left issue is a localized configuration correction with a straightforward validation path.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search