Clarification on /cost calculation when using free custom providers (NVIDIA NIM)

[methedjangoguy]

Hi team! Quick question regarding token cost tracking in OpenClaude:

Setup: I am using OpenClaude with an NVIDIA NIM model (nvidia/nemotron-3.5-lightning-30b-a3b).

Issue: Both OpenClaude and the NVIDIA NIM model tier I'm using are free, but running /cost reports an accumulated total (e.g., ~$0.18).

Question: Does /cost calculate an estimated market value for the tokens consumed (e.g., using built-in rates from Claude/OpenAI models), or does it attempt to track actual API balance deductions?

Want to double-check so I can accurately monitor usage. Appreciate any context on how cost estimates are computed for custom OpenAI-compatible endpoints! Will it cost any money?

[Maxaubert]

Short answer: /cost is a local estimate, not your real balance, and it won't cost you any money. It never reads anything from NVIDIA NIM. It multiplies the token counts from each response by a price table built into OpenClaude.

Why you see ~$0.18: the built-in table only knows Claude models. Any model id it doesn't recognize, like your nemotron one, falls back to the default "unknown model" price of $5 input / $25 output per million tokens:
https://github.com/Gitlawb/openclaude/blob/9a2910da236395d259c72c3b4ddf92278625a6c1/src/utils/modelCost.ts#L91
https://github.com/Gitlawb/openclaude/blob/9a2910da236395d259c72c3b4ddf92278625a6c1/src/utils/modelCost.ts#L188-L196

When that fallback is used, /cost also appends "(costs may be inaccurate due to usage of unknown models)" to the total. If you see that note, the number is just a placeholder estimate.

To make /cost show $0.00 for a free endpoint, add an exact-model override to your user settings (~/.openclaude/settings.json) or .openclaude/settings.local.json. Shared project settings are ignored for this key:

{
  "modelPricing": {
    "nvidia/nemotron-3.5-lightning-30b-a3b": {
      "inputTokens": 0,
      "outputTokens": 0,
      "promptCacheReadTokens": 0,
      "promptCacheWriteTokens": 0,
      "webSearchRequests": 0
    }
  }
}

The key has to match the model id sent to the API exactly (case-sensitive, no prefixes or globs). If you're unsure of it, run with OPENCLAUDE_LOG_TOKEN_USAGE=verbose and copy the model field from the log line. The four token fields are required (USD per million tokens). webSearchRequests is optional and defaults to $0.01 per request, so keep it at 0 for a fully free endpoint. Details in the docs:
https://github.com/Gitlawb/openclaude/blob/9a2910da236395d259c72c3b4ddf92278625a6c1/docs/advanced-setup.md#exact-model-pricing-overrides-settingsjson