Token & Cost count does not function with a custom provider

[KiaArmaniHees]

Describe the bug

When using a custom provider (Azure OpenAI via LiteLLM), tokens aren't being counted and cost isn't shown.

Setup

Please complete the following information along with version numbers, if applicable.

• Windows 11 23H2
• Powershell 5.1.22621.2706

To Reproduce

Steps to reproduce the behavior:

1. Create a crush.json with the following content:

{
  "$schema": "https://charm.land/crush.json",
  "providers": {
    "your-custom-provider": {
      "type": "openai",
      "base_url": "https://litellm.local.host/v1",
      "api_key": "REDACTED",
      "models": [
        {
          "id": "gpt-5",
          "name": "GPT-5",
          "context_window": 400000,
          "default_max_tokens": 16384,
          "cost_per_1m_in": 1.25,
          "cost_per_1m_out": 10.0
        },
        {
          "id": "gpt-5-mini",
          "name": "GPT-5-Mini",
          "context_window": 400000,
          "default_max_tokens": 8192,
          "cost_per_1m_in": 0.25,
          "cost_per_1m_out": 2.0
        }
      ]
    }
  },
  "models": {
    "large": {
      "model": "gpt-5",
      "provider": "your-custom-provider"
    }
  }
}

2. Launch crush
3. Make any request
4. Token count and cost remain at 0 and $0.00 respectively.

Expected behavior

Token and cost count should match the tokens used count and the calculated cost based on cost_per_1m_in and cost_per_1m_out.

Screenshots

Additional context

This issue does not occur with a built-in provider (i.e. using OpenAI directly using an API Key).

---

Maintainers’ note: Formatted JSON for readability.

[meowgorithm]

Appreciate the heads up—we're looking into this.

[csacca]

I just noticed this same issue in 0.11.1 with a LiteLLM endpoint.

[kujtimiihoxha]

this is on my list to do, it could be that litellm does not send usage as expected.

[csacca]

Is there a way to see if LiteLLM is not reporting token usage in the crush logs?

[kujtimiihoxha]

@csacca yes if you run crush -d you should get the streaming logs and see all streaming chunks when you run crush logs

[csacca]

Ok, I captured a simple debug log with the chunks:

crush.log

It looks to me like LiteLLM is reporting token usage. Comparing with OpenAI GPT-5 chunks, nothing stands out to me as super different regarding token reporting.

[kujtimiihoxha]

This should be fixed by #1171 and we will merge and release it soon