The part every other write-up gets wrong, because it treats the model bill as the whole bill.
On an API key OpenClaw bills every token at the model's published rate — the rates live in our rate tables and move when the provider reprices. Nothing is metered by OpenClaw itself, so there is no allowance to run out of and no window to wait for. That is the simple part.
The part that surprises people is that the model is one tap of five. An assistant living in your messaging apps quietly spends money in places a coding tool never touches, and each one reads a different key with a different bill attached:
- Replies and tool calls — the main model, and the only tap most people budget for.
- Incoming media — a voice note or a photo sent to the assistant is transcribed or described through a provider API before the reply is even written. Every message someone sends you can cost money.
- Memory search — semantic recall runs on an embedding provider. Set it to a local one and this tap closes entirely; leave it remote and it bills quietly in the background.
- Web search and fetch — Brave, Exa, Firecrawl, Tavily, Perplexity and others, each with its own key and its own invoice. Brave ships a renewing monthly free credit; DuckDuckGo and a self-hosted SearXNG need no key at all.
- Generating things — images, video and spoken replies each route to whichever provider you authenticated.
None of this is hidden: the vendor documents it plainly in its API usage and costs reference. It is simply never counted, because counting it is inconvenient for anyone selling the idea that a free assistant is free to run. The practical advice is unglamorous — close the taps you do not use, prefer the key-free search backends, and keep memory embeddings local unless you have a reason not to.
What OpenClaw does give you is visibility. Its status and usage commands estimate what the last reply cost, the control panel totals a date range by provider and model, and quota windows are read back for Anthropic, OpenAI, GitHub Copilot, DeepSeek, MiniMax, Xiaomi, z.ai and ClawRouter. Two honest limits: those totals describe your local session history rather than a provider invoice, and on the reused-Claude route no cost appears at all, because Anthropic does not expose a per-message estimate.