How much does OpenClaw cost to run? Monthly calculator (2026)
OpenClaw is free; the model and server aren't. Work out your monthly bill by model, messages and heartbeat, and see what drives OpenClaw's token use.
OpenClaw itself is free, but running it isn’t: you pay for the AI model it calls and for the machine it runs on. With moderate use (20 requests a day) the model bill ranges from under $10 a month on a budget model to several hundred dollars on a top model, and the heartbeat, OpenClaw’s check-in every 30 minutes, can cost more than everything you ask it to do. The server is the small part: $0 on a computer you already own, or about $6 to $15 a month for a VPS.
We haven’t run OpenClaw ourselves, so this isn’t a bill we measured. The calculator and the figures below combine OpenClaw’s documentation on token use with API and server prices we checked at each provider in October 2026 (see how we work). Change the inputs to match how you’d use it.
Where the money goes
| Cost | What you pay | Typical range, as of October 2026 |
|---|---|---|
| The software | Nothing. MIT license, “No subscription. No hosted tier.” (openclaw.ai) | $0 |
| The machine | Your own computer, a Raspberry Pi, a Mac mini or a VPS that stays on | $0 to about $15/mo for a small VPS (see where to run OpenClaw) |
| The model | Tokens sent to an API such as Claude, OpenAI’s models, Gemini or DeepSeek, or a subscription login, or your own GPU | From a few dollars to hundreds of dollars a month; this page is mostly about this line |
| Extras | Paid APIs you connect, e.g. web search, text-to-speech, image generation | Brave Search, for example, gives $5 a month of free credit, then charges $5 per 1,000 requests (OpenClaw docs) |
The machine needs little. OpenClaw’s FAQ puts the minimum at “1 vCPU, 1 GB RAM, ~500 MB disk” and recommends “1-2 vCPU, 2 GB+ RAM” (FAQ), which the cheapest plans at Hetzner and Hostinger cover. The model is where costs vary by a factor of a hundred.
Why OpenClaw uses so many tokens
A chat app sends the model your question. An agent sends its whole working context, with every call. OpenClaw’s documentation shows what goes in, in its own example of the /context list command (context docs):
| Part of each model call | Size in the docs’ example |
|---|---|
| System prompt: instructions, workspace files, skills list, runtime details | 38,412 characters, about 9,600 tokens |
| Tool definitions (JSON schemas), counted but not shown | 31,988 characters, about 8,000 tokens |
| Largest single tool: the browser | about 2,450 tokens |
The shell (exec) tool |
about 1,560 tokens |
| Before any conversation history | about 17,600 tokens |
The docs stress that these values “vary by model, provider, tool policy, and workspace”. Four things make it grow:
- Workspace files.
AGENTS.md,SOUL.md,USER.md,MEMORY.mdand the rest are injected into every call, up to 20,000 characters per file and 60,000 in total (system prompt). A longMEMORY.mdis paid for on every turn. - Skills and tools. Each installed skill adds “~97 characters + your
name,description, andlocation” to the prompt (skills), and every enabled tool adds its schema. - Conversation history and tool results. The session keeps growing until compaction summarizes it, which itself costs at least one model request (compaction). A single tool result can be up to 16,000 to 64,000 characters, depending on the model’s context window (token use).
- Several calls per request. When the model uses a tool, the result goes back to it in another call, with the whole context again. A request like “find my booking and check me in” can take several calls.
Then there’s the heartbeat. It runs “full agent turns” in your main session, every 30 minutes by default, and the docs warn that “Shorter intervals burn more tokens”. In a long-running main session, a heartbeat can carry the full history: the docs say isolating it cuts a run from “~100K tokens down to ~2-5K” (heartbeat).
Caching softens this. Providers charge much less for input they’ve seen recently: Claude Sonnet 5.5’s cached input costs $0.10 per million tokens against $2 for fresh input, in our price data. OpenClaw keeps its prompt prefix stable so the cache works (prompt caching). But the cache OpenClaw sets up for Anthropic by default lasts 5 minutes, so it has usually expired by the next 30-minute heartbeat.
Calculate your monthly cost
The defaults describe a moderate personal user: 20 requests a day, 3 model calls each, 30,000 input tokens per call (the docs’ 17,600-token example plus some history) and 600 output tokens. The heartbeat runs every 30 minutes around the clock with 30,000 tokens of context and no cache hits. Open “Token assumptions and heartbeat” to change any of it, including a cheaper model just for the heartbeat.
What OpenClaw costs you a month
Pick a model and a machine, then set how much you use it. The table under the result shows the same usage on every model we track.
Token assumptions and heartbeat
| Server | ||
|---|---|---|
| Your messages: input | ||
| Your messages: output | ||
| Heartbeat |
The same usage on every model
| Model | Per message | Model, a month | With server |
|---|
Model prices are list API prices per million tokens (input / output) as last checked 2026-10-18; server prices from our hosting comparison. Months are 30 days. Cache writes (Anthropic charges 1.25 times the input price for a 5-minute cache write), taxes and any free tiers aren't counted. An estimate from your inputs, not a measured bill.
These are list API prices. Your real bill also depends on how long your sessions get before compaction, how much the tools return, and cache writes, which Anthropic charges at 1.25 times the input price for a 5-minute cache.
What the numbers show
At the default usage, the model alone costs this much a month (server not included):
| Model | Price per 1M tokens (input / output) | Model cost a month | Of which heartbeat |
|---|---|---|---|
| GPT-6 Astra (a fresh OpenClaw setup’s default) | $10 / $50 | $741.60 | $439.20 |
| Claude Opus 5.5 (the default for a new Anthropic setup) | $4 / $20 | $290.16 | $175.68 |
| Claude Sonnet 5.5 or GPT-6.1 Sol | $2 / $10 | $145.08 | $87.84 |
| GPT-5.4 mini | $0.75 / $4.50 | $56.54 | $33.05 |
| Gemini 3.8 Flash | $0.75 / $3.75 | $55.62 | $32.94 |
| DeepSeek V4.1 Flash | $0.30 / $1.20 | $21.10 | $13.13 |
| Mistral Small 4 | $0.15 / $0.60 | $15.31 | $6.57 |
| Claude Haiku 5.5 | $0.10 / $0.50 | $7.42 | $4.39 |
Prices are from our AI model price data, checked October 17, 2026. The defaults come from OpenClaw’s docs: “A fresh setup with no primary model uses exact openai/gpt-6-astra”, and a fresh Anthropic setup “defaults to Opus 5.5” (models FAQ, Anthropic provider).
Three things stand out:
- The defaults are the expensive end. If you accept the default model and the default heartbeat, you start at the top of this table. That’s in line with the docs’ advice to use the strongest model available for an agent with tools (see the best model for OpenClaw), but it should be a choice you make knowingly.
- The heartbeat is often the biggest line. At these settings it costs more than all 20 daily requests put together on every model with a cached-input price, because its calls arrive after the cache has expired.
- Per request, it adds up fast. On Claude Sonnet 5.5, one request at the default size costs about 10 cents; on GPT-6 Astra, about 50 cents.
The heartbeat settings change everything
Same usage on Claude Sonnet 5.5, changing only the heartbeat:
| Heartbeat setting | Model cost a month |
|---|---|
| Long main session (~100K tokens per run, the docs’ figure), every 30 min | $346.68 |
| Default in our calculator (30K tokens), every 30 min | $145.08 |
| Isolated session (~5K tokens), every 30 min | $73.08 |
| Isolated, on Claude Haiku 5.5, hourly, 16 active hours a day | $57.50 |
| Heartbeat off | $57.24 |
The docs give the levers: isolatedSession: true “to avoid sending full conversation history”, lightContext: true “to skip workspace bootstrap files for heartbeat runs”, a cheaper heartbeat model, activeHours to keep it quiet at night, and every: "0m" to turn it off (heartbeat).
Eight ways to cut the bill
- Fix the heartbeat first. Isolate it, give it light context and a cheap model, limit it to waking hours, or switch it off if you don’t need check-ins. Scheduled jobs (automations) can do specific tasks at set times instead.
- Choose the model on purpose. Use a strong model for tasks with tools and untrusted input, as the docs advise, and set cheaper fallbacks for “cost/latency-sensitive tasks and lower-stakes chat” (models).
- Keep the cache warm.
cacheRetention: "long"gives a 1-hour cache on Anthropic (1-hour writes cost 2 times input). The docs note that a heartbeat every55mwith a 1-hour cache “can avoid re-caching the full prompt” (prompt caching). - Trim the workspace. Keep
MEMORY.mdandAGENTS.mdshort, since they ride along on every call in your main session. - Turn off tools and skills you don’t use. The browser schema alone is about 2,450 tokens in the docs’ example. Check sizes with
/context detail. - Compact and reset.
/compactsummarizes a long session;/newstarts fresh. A cheaper compaction model can be set withagents.defaults.compaction.model(compaction). - Shrink images.
agents.defaults.imageMaxDimensionPx(default 1200) caps image size; “Lower values reduce vision-token usage” (token use). - Watch it.
/statusshows the last reply’s tokens and estimated cost (for API-key models),/usage fulladds a usage line to each reply, and the dashboard’s Usage tab shows 30 days. The docs call these “local estimates, not an invoice”, so set a spending limit with your API provider too (usage tracking).
Can you use a ChatGPT or Claude subscription instead?
A subscription you already pay for can replace per-token billing, within its usage limits. The rules differ by company and have changed during 2026.
OpenAI. OpenClaw supports signing in with a Codex login or “Sign in with ChatGPT (Beta)”, which uses your plan’s allowance instead of API billing (OpenClaw docs). OpenAI’s own directory lists OpenClaw among “Open-source integrations” (OpenAI), and its Help Center says requests from other apps “count toward the ChatGPT Work and Codex usage included in your plan”, with a weekly cap per app (OpenAI Help Center). When you hit the cap, OpenClaw marks the login blocked until the reset. OpenAI’s pages disagree on whether Go plans qualify, or only Plus and Pro.
Anthropic. This one is murkier:
- April 4, 2026: Claude subscriptions stopped covering third-party tools such as OpenClaw, which then had to use paid extra usage or an API key (The Register).
- June 15: Anthropic paused a planned separate credit for such tools.
- October 7: its Help Center said you “can still use the Claude Agent SDK,
claude -p, and third-party apps with your subscription limits” (Anthropic Help Center). - At the same time: Anthropic’s Claude Code terms say it doesn’t permit third-party developers “to route requests through Free, Pro, or Max plan credentials on behalf of their users” (Claude Code legal).
OpenClaw reuses your own Claude Code login rather than offering its own. Its docs still say “For Anthropic in production, API key auth is still the safer recommended path”, and warn that “Anthropic token auth expires and can be revoked” (OpenClaw OAuth docs, Anthropic provider). Separately, Claude Max plans now include monthly API credits: $100 on Max 5x and $200 on Max 20x (Anthropic).
A subscription caps your spending, but an always-on agent with a heartbeat can use up a plan’s allowance quickly, and a blocked login stops the agent until the reset.
What others have reported
No maintainer publishes a typical monthly cost, and the figures below come from other people’s setups, not ours:
- One night of heartbeats on Claude Opus. Notebookcheck relayed user reports in February 2026: a 30-minute heartbeat sending about 120,000 tokens of context cost about $0.75 a run, “$18.75 gone in a single night” (Notebookcheck). These reports predate today’s
isolatedSessionandlightContextoptions. - 180 million tokens in about a week. That’s what MacStories’ Federico Viticci said he used on the Anthropic API with Claude Opus 4.5, in January 2026, without giving a dollar figure (MacStories). Larger dollar amounts that circulate on blogs aren’t his.
- Scheduled jobs. A hosting vendor, Phala, estimated a job running every 30 minutes at $6.95 a month in the main session vs $4.38 isolated, on one OpenAI coding model (Phala). Phala sells hosted OpenClaw.
- The $1.3 million month. Creator Peter Steinberger’s widely shared OpenAI bill was for about 100 Codex instances building OpenClaw, paid by OpenAI (Tom’s Hardware). It says nothing about what a user pays.
Running the model yourself instead
A local model replaces the token bill with hardware or GPU rental. On a rented GPU, an RTX 4090 costs $0.34 an hour in RunPod’s Community Cloud, or about $248 a month if it never stops, so it only beats an API at high volume. OpenClaw’s docs also warn that smaller local models are “more vulnerable to prompt injection”. Our guide to OpenClaw with local models runs the numbers.
FAQ
Is OpenClaw free?
Yes, the software is free and open source (MIT), with no paid tier. Running it costs money: the model’s API or subscription, and a computer or server that stays on.
How much does OpenClaw cost per month?
At 20 requests a day with the default 30-minute heartbeat, our calculator gives about $7 a month on Claude Haiku 5.5, $21 on DeepSeek V4.1 Flash, $145 on Claude Sonnet 5.5 and $742 on GPT-6 Astra (OpenClaw’s default for a fresh setup), plus $0 to about $15 for a server. Your usage will differ.
Why is my OpenClaw bill so high?
Usually it’s the heartbeat or a long session. Each heartbeat is a full agent turn, every 30 minutes by default, and long sessions resend their history with every call. Isolate the heartbeat, give it a cheap model and light context, and compact or reset long sessions. Then check /context detail and /usage full.
Does the OpenClaw heartbeat cost money?
Yes. Each heartbeat is a model call with your session’s context, even when it finds nothing to report. Set every: "0m" to turn it off, or use isolatedSession, lightContext and a cheap heartbeat model to make it cheaper.
Can I run OpenClaw with no API costs?
Yes, with a local model on your own hardware, or within a subscription’s limits via a supported login. Both have trade-offs: local models need a capable GPU and are weaker against prompt injection, and subscription allowances have caps and changing rules.