[Bug]: Cost tracking is not working with Vercel Models

[NoAchache]

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

I am using vercel through LiteLLM. All users are supposed to have a rate limit of 10$/month but the cost tracking is not working (c.f. example screenshot, 16M token for 0$). The bug happened with anthropic-claude-opus-4.5 but i feel like it happens for all models

Steps to Reproduce

1. LiteLLM version: v1.81.0-stable (self-deployed on GCP cloudrun)
2. spec in config.yaml:
   - model_name: vercel/anthropic-claude-opus-4.5 litellm_params: model: vercel_ai_gateway/anthropic/claude-opus-4.5 api_key: os.environ/VERCEL_AI_GATEWAY_API_KEY
3. Call the model through the LiteLLM server

Relevant log output

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v.1.81.0-stable

Twitter / LinkedIn details

No response

[NoAchache]

On other models, it partially tracks usage. E.g. with gemini 2.5 pro, it says I used 883k output tokens although 12M are logged in Vercel

[NoAchache]

After further analysis, for the gemini problem, the issue is that the reasoning token are not counted.

[shin-bot-litellm]

Investigating this. The pricing data exists for vercel_ai_gateway/anthropic/claude-opus-4.5 in our model cost database:

"vercel_ai_gateway/anthropic/claude-opus-4.5": {
    "input_cost_per_token": 5e-06,
    "output_cost_per_token": 2.5e-05,
    ...
}

The issue might be:

1. Usage data not being returned from Vercel AI Gateway
2. Model name mismatch during cost lookup

@NoAchache Can you help debug by adding --detailed_debug flag to your LiteLLM proxy and sharing:

1. The actual model name being used in the request (check logs for model=)
2. Whether the response includes usage data (prompt_tokens, completion_tokens)

Also, can you try adding explicit model_info to your config to force cost calculation:

- model_name: vercel/anthropic-claude-opus-4.5
  litellm_params:
    model: vercel_ai_gateway/anthropic/claude-opus-4.5
    api_key: os.environ/VERCEL_AI_GATEWAY_API_KEY
  model_info:
    input_cost_per_token: 0.000005
    output_cost_per_token: 0.000025

This will confirm if the issue is with model lookup vs usage data.

[NoAchache]

• The issue for anthropic is that the model was not referenced in model_prices_and_context_window.json
• I think the issue for gemini models is that they don't have "reasoning: true" in model_prices_and_context_window.json. Can we add a more resilient billing systems to check if there are reasoning token in ALL models?