## Core issue

This is a **routing bug**, not just a tokenizer fallback issue.

LiteLLM supports Anthropic-compatible custom backends using `custom_llm_provider: anthropic` with a custom `api_base`, for example `api_base: http://<self-hosted-backend>/v1`.

In this setup, Anthropic Messages API routes should respect the configured `api_base`.

The normal `/v1/messages` route does this correctly.

However, `/v1/messages/count_tokens` does not. It ignores the configured `api_base` and directly calls the hardcoded Anthropic SaaS endpoint: `https://api.anthropic.com/v1/messages/count_tokens`.

This means one Anthropic sub-route silently bypasses the user-configured backend.

## Why this matters

This breaks valid LiteLLM deployment patterns:

* **self-hosted Anthropic-compatible backends** * **vLLM** serving Anthropic-compatible `/v1/messages` * **air-gapped or firewall-restricted environments** * **Claude Code integrations** routed through LiteLLM to an internal backend

Expected behavior:

`/v1/messages/count_tokens` → LiteLLM → configured `api_base` → `<api_base>/messages/count_tokens`

Actual behavior:

`/v1/messages/count_tokens` → LiteLLM → `https://api.anthropic.com/v1/messages/c…