How much does OpenClaw cost to run?
OpenClaw itself is free, open-source software (MIT licence). You pay for the AI model it talks to, the computer it runs on and that computer's electricity. Using prices checked in September 2026, my three worked examples below come to about $14 a month for light use on a laptop you already own, $29 for a typical mini PC setup and $271 for heavy use of a top model on a rented server.
The short answer: four costs, and one is much bigger than the rest
Running OpenClaw has four costs, and the AI model is usually the biggest. Setup help is a fifth, optional one.
| Cost | Typical range (September 2026) | How you pay |
|---|---|---|
| Software (OpenClaw or Hermes Agent) | $0 | Free, MIT licence |
| AI model | $0 on free tiers to hundreds of dollars | Per token, or a flat subscription |
| Computer | $0 if you own one; about $6 to $24 a month to rent | One-off purchase, or monthly rent |
| Electricity | About $0.50 to $9 a month for one small computer | Your power bill |
| Setup help (optional) | $0 if you do it yourself; about $200 to $600 from paid services | One-off |
The software: free
OpenClaw costs nothing to download or use. Its code is on GitHub under the MIT licence, held by the OpenClaw Foundation. Hermes Agent, from Nous Research, is also MIT-licensed. Paid "OpenClaw plans" you see advertised, such as Hostinger's, bundle hosting, setup or model credits around the free software. If you're choosing between the two agents, see Hermes vs OpenClaw.
The AI model: where most of the money goes
You pay for the model every time the agent thinks. You can pay per token with an API key or, with some providers, use a flat monthly subscription.
Pay-per-token prices
Prices are in US dollars per million tokens. A token is about four characters of English text, according to OpenClaw's docs. Input is what the agent sends; output is what the model writes back.
| Model | Input | Output | Cached input |
|---|---|---|---|
| Claude Opus 5.5 (Anthropic) | $4.00 | $20.00 | $0.20 |
| Claude Sonnet 5.5 (Anthropic) | $2.00 | $10.00 | $0.20 |
| Claude Haiku 4.5 (Anthropic) | $1.00 | $5.00 | $0.10 |
| GPT-6.1 Sol (OpenAI) | $2.00 | $10.00 | $0.10 |
| GPT-5.4 mini (OpenAI) | $0.75 | $4.50 | $0.075 |
| Gemini 3.8 Flash (Google) | $0.75 | $3.75 | $0.075 |
| DeepSeek Flash, peak hours | $0.30 | $1.20 | $0.006 |
Google's Gemini 3.8 Flash prices double on 1 January 2027, to $1.50 input and $7.50 output. DeepSeek charges half its peak rates outside peak hours.
Agents use far more input than output. OpenClaw rebuilds its instructions on every run, and its standing files (such as AGENTS.md and MEMORY.md) can add up to 60,000 characters, roughly 15,000 tokens, before any conversation or tool results. That is why cached-input prices matter. Anthropic charges 10% of the normal input price to re-read a cached prompt (5% on Opus 5.5) and 1.25 times the input price to write one to its 5-minute cache.
Flat subscriptions: what's allowed
Some subscriptions work with OpenClaw and some are unsettled. As of September 2026:
- ChatGPT (OpenAI): OpenClaw supports signing in with a ChatGPT or Codex account, and its docs say OpenAI explicitly allows this. Plus is $20 a month; Pro is $100, $200 or $500. Each plan has usage limits. When you hit one, you wait for it to reset or, where offered, buy credits.
- Claude (Anthropic): On 4 April 2026 Anthropic stopped Claude subscription limits covering OpenClaw and other third-party tools, TechCrunch reported. A replacement monthly credit was announced, then paused on 15 June. OpenClaw's docs now say its Claude CLI route, which runs your own Claude Code login, draws from your plan's usage limits, and they recommend an API key for predictable spending. Pro is $20 a month ($17 billed annually); Max starts at $100. Expect this route to change again.
- Others: OpenClaw's docs also list subscription-style options from Qwen Cloud, MiniMax and Z.AI (GLM).
A subscription caps your bill but also caps your usage. With pay-per-token, the bill grows with use.
The computer: one you own, or one you rent
Any computer that stays on will do. OpenClaw runs on macOS, Linux and Windows. Its docs put the minimum at 1 CPU core, 1 GB of memory and about 500 MB of disk, with 2 GB of memory or more recommended.
Own it. A computer you already have costs nothing extra except power. OpenClaw's FAQ says you don't need a Mac mini; a small VPS, home server or Raspberry Pi-class box works too. If you're thinking of buying one, read whether a Mac mini is worth it for OpenClaw or try Which computer should run your agent?
Rent it. A virtual private server (VPS) swaps the one-off cost for a monthly bill. OpenClaw's FAQ notes the trade-off: a VPS is often headless (it has no screen), while a computer you own can show a real browser window.
| Option | Price a month | Notes |
|---|---|---|
| DigitalOcean Basic, 1 vCPU, 1 GiB | $6 | Meets OpenClaw's minimum |
| DigitalOcean Basic, 1 vCPU, 2 GiB | $12 | Meets the recommended 2 GB |
| DigitalOcean Basic, 2 vCPUs, 4 GiB | $24 | More headroom; OpenClaw's FAQ says browser automation can be resource-hungry |
| Hetzner CX23 | $6.49 | Germany or Finland; excludes VAT and IPv4; prices rose 15 June 2026 |
| Hetzner CX33 | $9.99 | As above |
| Hostinger managed AI apps (OpenClaw or Hermes) | $5.99 to $10.99 | $5.99 on a 24-month plan, $10.99 month to month; renews at $11.99 |
Electricity: usually a few dollars
A small computer left on all month costs between about 50 cents and $9 in power at the US average price. That average was 18.31 cents per kilowatt-hour in July 2026, according to the US Energy Information Administration (EIA). It ranged from 12.72 cents in Louisiana to 48.00 cents in Hawaii.
The arithmetic: watts × 730 hours ÷ 1,000 = kilowatt-hours a month. Kilowatt-hours × $0.1831 = dollars a month.
| Computer | Power draw | kWh a month | Cost a month |
|---|---|---|---|
| Mac mini (M6), idle | 4 W (Apple) | 2.92 | $0.53 |
| Mac mini (M6), maximum | 70 W (Apple) | 51.10 | $9.36 |
| Intel N100 mini PC, idle | 4.8 W (Notebookcheck) | 3.50 | $0.64 |
| Intel N100 mini PC, average load | 13 W (Notebookcheck) | 9.49 | $1.74 |
| Laptop, switched on | 19 to 32 W (LBNL) | 13.87 to 23.36 | $2.54 to $4.28 |
| Desktop, switched on | 66.1 W (LBNL) | 48.25 | $8.84 |
When the model runs in the cloud, the computer mostly waits for replies, so expect a dedicated agent machine to sit nearer its idle figure than its maximum. The laptop and desktop figures come from a Lawrence Berkeley National Laboratory (LBNL) study of home computers metered in 2012, and the earlier studies it compares, so treat them as rough numbers for older machines. A plug-in power meter will tell you what yours draws.
Setup: your time, or someone else's
Setup costs your time, or money if you pay someone to do it. I found no independent survey of setup prices, so these are examples, not a market average. As of September 2026, US services advertise one-off setup from $199 (OpenClawPro, which also offers a $99-a-month managed plan) and $299 for a one-hour session or around $600 for four hours (OpenClaw Install). In China, the South China Morning Post reported in March 2026 that after users paid to have OpenClaw installed, sellers on the Xianyu marketplace began charging to remove it, one for 299 yuan ($43.55).
Whoever sets it up, keep time for upkeep: updates, and checking what the agent can reach. Is OpenClaw safe? covers the risks.
Three worked budgets
These examples show the arithmetic with September 2026 prices. The usage figures (how many model calls, and how big each one is) are my assumptions for illustration, not measurements. Your own use will differ, so check your provider's dashboard after the first week.
Light: a laptop you already own and a cheap model
- Usage: 20 model calls a day × 30 days = 600 calls. Each sends 20,000 tokens and gets 500 back: 12 million input tokens and 0.3 million output tokens.
- Model, Gemini 3.8 Flash: 12 × $0.75 = $9.00, plus 0.3 × $3.75 = $1.13. Total $10.13.
- Computer: $0, already owned.
- Electricity: laptop at 32 W = $4.28.
- Total: $10.13 + $4.28 = $14.41 a month.
From 1 January 2027 the same Gemini use would cost $20.25 ($18.00 + $2.25). DeepSeek Flash at peak rates would cost $3.96 ($3.60 + $0.36).
Typical: a mini PC on a shelf and a mid-priced model
- Usage: 50 calls a day × 30 days = 1,500 calls, at 20,000 tokens in and 500 out: 30 million input and 0.75 million output.
- Model, Claude Sonnet 5.5, assuming 80% of input is read from cache: 24 million × $0.20 = $4.80 for cache reads; 6 million × $2.50 = $15.00 for cache writes; 0.75 million × $10 = $7.50 for output. Total $27.30.
- Computer: a mini PC you own, or its purchase price spread over the years you keep it.
- Electricity: N100 mini PC at 13 W = $1.74.
- Total: $27.30 + $1.74 = $29.04 a month.
Without caching, the same model use would cost $67.50 (30 × $2 + 0.75 × $10). Treating every uncached token as a cache write slightly overstates the cached figure. With ChatGPT Plus instead of an API key, the model line becomes a flat $20, within Plus's usage limits.
Heavy: a rented server and a top model all day
- Usage: 150 calls a day × 30 days = 4,500 calls, at 30,000 tokens in and 1,000 out: 135 million input and 4.5 million output.
- Model, Claude Opus 5.5, assuming 80% cached: 108 million × $0.20 = $21.60; 27 million × $5.00 = $135.00; 4.5 million × $20 = $90.00. Total $246.60.
- Computer: DigitalOcean 2 vCPUs, 4 GiB = $24.00.
- Electricity: included in the rent.
- Total: $246.60 + $24.00 = $270.60 a month.
Without caching, the model alone would cost $630 (135 × $4 + 4.5 × $20). A ChatGPT Pro plan at $100 or $200 would cap the model line instead, if its limits fit your use.
What Hermes Agent costs
Hermes Agent has the same cost structure: the software is free, and you pay for a model, a computer and power. Nous Research recommends its Nous Portal subscription. As of September 2026 the Portal has a free tier (free models only) and paid tiers of $20, $100 and $200 a month, which include $22, $110 and $220 of monthly credits. The Portal also lets you connect an eligible ChatGPT plan without a paid Nous subscription. Hermes can also sign in with ChatGPT/Codex or GitHub Copilot, or use pay-per-token API keys from Anthropic, OpenAI, Google, DeepSeek, OpenRouter and many others.
How to keep model costs down
The cheapest savings come from sending fewer, smaller and better-cached requests.
- Use a cheaper model for routine work. OpenClaw can run a fast chat agent on one model and a separate agent on a bigger model, each with its own default.
- Let caching work. OpenClaw can prune a session once the cache expires, and a heartbeat set just under the cache lifetime (its docs use 55 minutes for a 1-hour cache) avoids re-caching the whole prompt.
- Keep standing instructions short. Files such as AGENTS.md, SOUL.md and MEMORY.md go into requests, up to 20,000 characters each and 60,000 in total by default.
- Leave huge context windows off. OpenClaw limits GPT-5.5 and GPT-5.6 to 272,000 tokens by default because OpenAI charges long-context rates on the whole request above that.
- Shrink screenshots. Lowering
imageMaxDimensionPx(default 1,200 pixels) cuts image tokens. - Watch the meter.
/status,/usage fulland/usage costshow token use and estimated cost in chat. - Consider a flat plan if your use is steady and heavy.
Questions
Is OpenClaw free?
Yes. OpenClaw is open-source software under the MIT licence, so it costs nothing to download or use. You still pay for the AI model it uses, the computer it runs on and that computer's electricity.
Can I run OpenClaw for nothing?
Nearly. On a computer you already own, Google's Gemini API has a free tier, though it has limits and Google may use free-tier content to improve its products. You still pay for electricity. OpenClaw's FAQ says small local models are usually a poor fit: they truncate context and raise prompt-injection risk.
Can I use my ChatGPT subscription with OpenClaw?
Yes. OpenClaw supports signing in with ChatGPT or Codex, and its docs say OpenAI explicitly allows it. Usage counts against your plan's limits: Plus is $20 a month, Pro $100 to $500, as of September 2026.
Can I use my Claude Pro or Max subscription with OpenClaw?
Only through OpenClaw's Claude CLI route, which runs your own Claude Code login, and Anthropic's rules have changed several times in 2026. OpenClaw's docs say this currently draws from your plan's usage limits and recommend an Anthropic API key for predictable spending.
Do I need to buy a Mac mini to run OpenClaw?
No. OpenClaw's FAQ says a small VPS, home server or Raspberry Pi-class box works too. You need a Mac only for macOS-only features such as iMessage. See Mac mini for OpenClaw.
How much does Hermes Agent cost?
The software is free (MIT licence). Nous Portal, the subscription Nous Research recommends, costs $0, $20, $100 or $200 a month as of September 2026. You can also bring a ChatGPT plan or pay-per-token API keys instead.
Sources
- OpenClaw on GitHub: LICENSE (MIT) (checked 29 September 2026)
- Nous Research on GitHub: hermes-agent (MIT License) (checked 29 September 2026)
- OpenClaw Docs: Token use and costs (checked 29 September 2026)
- OpenClaw Docs: FAQ: providers, hardware, and hosting (checked 29 September 2026)
- OpenClaw Docs: Anthropic (checked 29 September 2026)
- OpenClaw Docs: OpenAI (checked 29 September 2026)
- Anthropic: Pricing (API) (checked 29 September 2026)
- Anthropic: Claude plans and pricing (checked 29 September 2026)
- Anthropic Help: Use the Claude Agent SDK with your Claude plan (15 June 2026 update)
- TechCrunch: Anthropic says Claude Code subscribers will need to pay extra for OpenClaw usage (4 April 2026)
- OpenAI: API pricing (checked 29 September 2026)
- OpenAI Help: What is ChatGPT Plus? (checked 29 September 2026)
- OpenAI Help: About ChatGPT Pro tiers (checked 29 September 2026)
- Google AI for Developers: Gemini Developer API pricing (checked 29 September 2026)
- DeepSeek API Docs: Models and pricing (checked 29 September 2026)
- Hermes Agent Docs: LLM and model providers (checked 29 September 2026)
- Nous Research: Nous Portal subscription tiers (checked 29 September 2026)
- DigitalOcean: Droplet pricing (checked 29 September 2026)
- Hetzner Docs: Price adjustment 15 June 2026 (last changed 8 July 2026)
- Hostinger: Managed AI automation apps (checked 29 September 2026)
- US EIA: Electric Power Monthly, Table 5.6.A, July 2026 data (released 24 September 2026)
- Apple Support: Mac mini power consumption and thermal output (checked 29 September 2026)
- Notebookcheck: Minisforum UN100L review (26 January 2024)
- Lawrence Berkeley National Laboratory: Computer usage and national energy consumption (1 December 2014)
- OpenClawPro: OpenClaw setup service pricing (checked 29 September 2026)
- OpenClaw Install: Pricing (checked 29 September 2026)
- South China Morning Post: China's OpenClaw users paid to install viral AI agent. Now they spend to remove it (12 March 2026)