Prices and the cost shown
markitai reports what each conversion cost from the number of tokens a model used and a price list it ships. A request whose model has no reviewed price is counted, but its cost is unknown, not $0: the cost shown is the known part (the priced requests), and the total says it is incomplete (! Cost is incomplete: …).
| Model | Price |
|---|---|
| Listed OpenAI, Anthropic, Gemini and DeepSeek models, at the provider's own address | From the list |
| OpenRouter | The charge OpenRouter reports |
A server of your own (ollama/, lm_studio/, hosted_vllm/, ...) | $0, at any address |
A gateway or proxy (openai/ with an api_base, litellm_proxy/), Bedrock, Vertex AI, other providers | Unknown |
Subscriptions (claude-agent/, copilot/, chatgpt/) | Unknown: billed by your plan |
To get a cost:
- Use a listed model, such as the defaults.
- Remove a custom
api_base(a proxy or relay): its prices are not the provider's. - Pin an exact model name in
litellm_params.modelwhen the provider answers with a different model than you asked for. - Update markitai if a newer release lists the model.
markitai has no setting for your own prices. Costs appear in the summary, --json, batch reports, the web workspace and MCP results.
Spending limit
llm.max_cost_per_document_usd (0: no limit) stops a document's further requests once its known cost passes the cap. With a cap set, a model whose price is unknown, a subscription included, is refused before anything is sent. Requests already running still finish, so a document can end slightly above the cap. llm.max_requests_per_document (50) limits every model.