## Problem

When running Anthropic models across multiple backends (direct API, AWS Bedrock, GCP Vertex AI, etc.), there is currently no built-in mechanism for health-aware routing with automatic failover. If one backend becomes degraded or rate-limited, requests must be manually re-routed.

## Use case

I operate a LiteLLM proxy that serves a team with Anthropic models. Some users route through direct API, others through Bedrock. When the direct API has a regional outage, I would like the proxy to automatically fail over to Bedrock without manual intervention.

## Prior work

I had a working draft in PR #33340 (now closed) that implemented a `RouterAnthropicPassthrough` class with: - Per-backend health tracking (consecutive failure counters, latency windows) - Configurable fallback order - No dependency on shared state across requests (stateless health probes)

The implementation was contained and had zero new dependencies beyond what LiteLLM already vendors.

## Question

Is health-aware routing for Anthropic backends something the project would consider? If there is interest, I am happy to re-open and polish the PR to meet the team's standards.