Bug Description

The /v1/messages Anthropic passthrough endpoint sets kwargs["model"] to the raw Bedrock model ID (e.g., global.anthropic.claude-opus-4-6-v1) instead of the provider-prefixed model name (e.g., bedrock/global.anthropic.claude-opus-4-6-v1). This causes duplicate Prometheus metric series on litellm_llm_api_latency_metric — the same underlying model appears under two different model labels.

How to Reproduce

1. Configure LiteLLM with a Bedrock model and the prometheus callback:

model_list:
  - model_name: claude-opus-4-6
    litellm_params:
      model: bedrock/global.anthropic.claude-opus-4-6-v1

litellm_settings:
  callbacks: ["prometheus"]

1. Send a request via the Anthropic passthrough endpoint:

curl -X POST http://localhost:4000/v1/messages \
  -H "x-api-key: sk-..." \
  -H "Content-Type: application/json" \
  -d '{"model": "claude-opus-4-6", "max_tokens": 100, "messages": [{"role": "user", "content": "hello"}]}'

1. Send the same request via the standard endpoint:

curl -X POST http://localhost:4000/chat/completions \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{"model": "claude-opus-4-6", "max_tokens": 100, "messages": [{"role": "user", "content": "hello"}]}'

1. Check Prometheus metrics:

curl http://localhost:4000/metrics | grep litellm_llm_api_latency

Expected: All metric series show model="bedrock/global.anthropic.claude-opus-4-6-v1"

Actual: Two separate series appear:

• model="bedrock/global.anthropic.claude-opus-4-6-v1" (from /chat/completions)
• model="global.anthropic.claude-opus-4-6-v1" (from /v1/messages)

Root Cause

In the Anthropic passthrough logging handler, kwargs["model"] is populated from the Bedrock API response body, which returns the raw model ID without the bedrock/ LiteLLM provider prefix. The standard /chat/completions path preserves the prefix because the model goes through LiteLLM's router resolution.

The PrometheusLogger.async_log_success_event reads the model from kwargs.get("model", "") (prometheus.py line ~795) and uses it directly as the Prometheus label value, so the inconsistency propagates into metrics.

Impact

• Duplicate metric series for every Bedrock model (doubles metric cardinality)
• The two series record different subsets of traffic with different latency characteristics, making dashboards and monitors unreliable when grouping by model label
• The litellm_model and host labels are also N/A on the passthrough series, suggesting the logging payload is incomplete for this code path
• Affects all Prometheus-based dashboards and monitors that group by model label

Environment

• LiteLLM version: v1.80.8.rc.1
• Provider: AWS Bedrock (cross-region inference profiles)
• Callbacks: prometheus, datadog_llm_observability

Workaround

We're currently monkey-patching PrometheusLogger at startup to normalize kwargs["model"] before metric emission — prepending bedrock/ when a bare Bedrock model ID is detected.