What happened?

When a model is configured with custom_llm_provider: "anthropic" and explicit zero costs (input_cost_per_token: 0, output_cost_per_token: 0), LiteLLM ignores the zero-cost override and calculates spend using the internal Anthropic pricing database.

This affects any deployment that proxies Anthropic models through an intermediate service (subscription-based, internal proxy, etc.) where the actual per-token cost is zero.

Relevant code

In litellm/cost_calculator.py, cost_per_token() has a provider dispatch chain:

elif custom_llm_provider == "anthropic":
    return anthropic_cost_per_token(model=model, usage=usage_block)

This runs before the else branch that would check litellm.model_cost for custom pricing entries. So when custom_llm_provider == "anthropic", the function dispatches directly to anthropic_cost_per_token(), which calls get_model_info() against the internal price DB (model_prices_and_context_window.json), finds the real Anthropic pricing (e.g. $15/$75 per MTok for Opus 4.6), and returns that — completely bypassing the configured 0 costs.

Additionally, use_custom_pricing_for_model() correctly detects input_cost_per_token: 0.0 as custom pricing (is not None check), but this signal is never consulted in the Anthropic dispatch path.

LiteLLM Version

v1.82.3-stable (Docker image: ghcr.io/berriai/litellm:v1.82.3-stable)

Configuration

model_list:
  - model_name: my-proxy/claude-opus-4-6
    litellm_params:
      model: my-proxy/claude-opus-4-6
      api_base: https://my-internal-proxy.example.com
      api_key: os.environ/PROXY_API_KEY
      custom_llm_provider: anthropic
      input_cost_per_token: 0.0
      output_cost_per_token: 0.0
      cache_creation_input_token_cost: 0.0
      cache_read_input_token_cost: 0.0
    model_info:
      input_cost_per_token: 0.0
      output_cost_per_token: 0.0
      cache_creation_input_token_cost: 0.0
      cache_read_input_token_cost: 0.0

Evidence

Expected: $0.00 spend (zero-cost subscription model)

Actual: real Anthropic pricing applied

Recent request logs:

The calculated costs match standard Anthropic API pricing, confirming LiteLLM's cost calculator is using the internal price DB rather than the configured 0.0 overrides.

Aggregated daily spend shows ~$700/day for claude-opus-4-6 across our team — all phantom costs from this bug.

Additional context

• custom_llm_provider: "anthropic" is required in our setup because the intermediate proxy expects Anthropic-format requests (with metadata.user_id for per-user routing)
• Setting base_model to a fake name via the /model/update API does not work — the API auto-populates base_model from the model name and ignores the override
• The same issue likely affects all provider-specific dispatch paths (openai, anthropic, bedrock, etc.) where cost_per_token() dispatches to the provider function before checking custom pricing
• Related (but opposite symptom): #23309 — custom pricing not applied for anthropic_messages call type

Suggested fix

In cost_per_token(), check for custom pricing (via use_custom_pricing_for_model() or custom_cost_per_token) before the provider-specific dispatch chain. If custom pricing is configured, use it directly and skip the provider lookup. Something like:

# Check custom pricing FIRST, before provider dispatch
if custom_pricing and _has_custom_cost_overrides(model_info):
    return _apply_custom_pricing(model_info, usage)

# Then fall through to provider-specific paths
elif custom_llm_provider == "anthropic":
    return anthropic_cost_per_token(...)