What OpenClaw actually costs.
OpenClaw is free, open-source software — there's no subscription, no seat fee, no bill from a vendor. But "free to download" is not "free to run." Your real cost is whatever you plug into it. Here's the honest math.
01 The software is free. The intelligence is not.
OpenClaw itself costs nothing — you download it and run it yourself. What it does not include is a brain. OpenClaw is a shell that drives a model; you supply the model. That means your cost is one (or both) of two things:
- Bring your own key (BYOK): you connect your own API key from OpenAI, Anthropic, Google, xAI, etc. OpenClaw then spends your API tokens — billed by that provider, not by OpenClaw.
- Run a local model: you host an open-weights model on your own hardware. No per-token bill, but you pay in compute — a capable GPU, electricity, and your time keeping it running.
02 Path A — Bring your own key
This is the common setup. Your cost is just the underlying model's per-token API price. So OpenClaw's "cost" is really a passthrough of whatever you connect:
| Model you plug in | Input / Output (per 1M) |
|---|---|
| Grok 4.1 Fast | $0.20 / $0.50 — cheapest |
| Gemini 3.5 Flash | $1.50 / $9 |
| Claude Sonnet | $3 / $15 |
| GPT-5.2 | $1.75 / $14 |
The lesson: OpenClaw doesn't have a price — the model you choose does. Pick a cheap fast model for everyday work and a frontier model only when the task needs it.
03 Path B — Run a local model
If you host an open-weights model yourself, there's no per-token bill at all. Instead you pay in:
- Hardware — a capable GPU (or a paid cloud GPU by the hour).
- Electricity — running that hardware.
- Your time — setup, updates, and troubleshooting. Often the biggest real cost.
Local makes sense when privacy is paramount (nothing leaves your machine) or when you run so much volume that fixed hardware beats per-token fees. For most people starting out, BYOK with a cheap API model is simpler and cheaper.
04 What a task actually costs
Say OpenClaw handles a multi-step task and makes ~10 model calls totaling 30K input / 15K output tokens:
- On Grok 4.1 Fast: about $0.01.
- On Claude Sonnet: about $0.32.
Same task, 30× cost difference — entirely from the model you picked. That's the only pricing lever OpenClaw gives you, and it's a powerful one.
OpenClaw is free software with a real running cost. Budget for the model — BYOK token spend, or hardware + power if local — and set a cap before any unattended run. There's no subscription to cancel, but there's no free lunch either.
Use the calculator to compare what the same OpenClaw workload would cost across different models you might plug in.
Open the AI Cost Calculator →Set the cap first
Before you run OpenClaw on anything real, go to your API provider's console and set a hard monthly spending limit — start at $10. Then run one task and watch the usage. You'll learn your true per-task cost and sleep better knowing a runaway loop can't drain your card.
Run the numbers →What you can do now
- Separate "free software" from "free to run" — they're not the same
- Know your two cost paths: BYOK tokens, or local hardware + power
- Treat OpenClaw's cost as a passthrough of the model you choose
- Always set a provider spending cap before unattended runs