How much will your AI agent cost each month?
For Hermes Agent and OpenClaw. Pick what your agent does and which model it uses. You see the AI bill, the background work most estimates leave out, and the server, side by side, with every assumption written down.
Model prices checked on the 8th of October 2026 from each provider's own pricing page.
Both agents use the same formula here. Your model, your messages and your background jobs decide the bill.
What does your agent do?
Price checked on the 8th of October 2026.
Telegram, WhatsApp, Slack, Discord and the rest add almost nothing on their own. Each message and each person does.
Background work
A check-in that re-reads the whole conversation carries about 100,000 tokens. A fresh short session carries 2,000 to 5,000.
AI usage a month, paid to Anthropic at their own price
About €25 to €138 at the European Central Bank rate of the 7th of October 2026.
Heavy-use bill split:
- Your messages94%
- Scheduled jobs6%
- Check-ins0%
This is an estimate, shown as a range, with every assumption listed below the calculator. Your provider's bill is the final word.
Smaller model: GPT-6 Luna comes to $1.49 to $8.00 for the same use. Good for everyday chat and check-ins.
Free models on OpenRouter: your load of up to 79 calls a day fits the 1,000-a-day limit after a one-time $10 credit purchase. Nous Portal's free plan also offers free models.
Sign in with a plan you already pay for: a ChatGPT or Grok subscription signs in on both agents and runs the models that plan includes. On Hermes Agent, Nous Portal works too (Hermes Plus is $20 a month with $22 in credits).
Plus the server it runs on
Your agent needs a computer that stays on. Two ways to get one.
Run it yourself
A small 2 vCPU, 4 GB plan. Add VAT and any IPv4 fee.
- Install and update the agent
- SSL, firewall, backups and monitoring
- Your own time when it stops replying
Let us run it: Starter
€14.99 a month
€12.49 a month billed yearly, plus VAT.
- Your own dedicated server in the EU
- Set up, updated, backed up and watched for you
- Same plans and prices for both agents
Let us run the server. Try it 7 days free.
A real dedicated Starter server, a Google account to start, no card. Your agent's own /usage command shows your real token count, so after a week you know your number.
oneclickclaw.io/ai-agent-cost-calculator?agent=hermes&preset=personal&model=anthropic.claude-sonnet-5.5&people=1&msgs=30&channels=1&checkins=0&jobs=1&size=lightThe link carries only the numbers above. Nothing is saved on our side.How the estimate works
Every time your agent answers, it sends the model its instructions, its tool list, its memory files and the recent conversation, then gets a reply back. You pay for both directions.
monthly cost = calls × (uncached input × input price + cached input × cache price + reply × output price)
Calls come from three places: your messages, scheduled jobs and check-ins. The low figure uses light settings and the high figure uses heavy ones. Where a provider lists a cache-write price, uncached input is billed at that price. Both settings are listed below, with where each number comes from.
Assumptions per model call
| Setting | Light | Heavy |
|---|---|---|
| Instructions, tools and memory files (tokens) | 15,000 | 30,000 |
| Recent conversation carried (tokens) | 5,000 | 15,000 |
| Model calls per message | 1.5 | 2.5 |
| Reply length (tokens) | 300 | 600 |
| Share of input read from cache | 70% | 50% |
| Check-in that re-reads the conversation (tokens) | 60,000 | 150,000 |
| Check-in in a fresh short session (tokens) | 3,000 | 8,000 |
| Cache share on check-ins | 0% | 0% |
| Calls per scheduled job | 2 | 4 |
| Context per scheduled job (tokens) | 10,000 | 20,000 |
Where these numbers come from
- OpenClaw caps the files it injects into every call at 20,000 characters each and 60,000 characters in total. OpenClaw docs, token use
- A check-in that runs in the main conversation carries about 100,000 tokens; one in a fresh short session carries about 2,000 to 5,000. OpenClaw docs, heartbeat
- The default prompt cache lives 5 minutes, so calls spaced 30 minutes or more pay the full input price. OpenClaw docs, prompt caching
- Hermes Agent check-ins are set by the user and return to the same conversation and prompt cache. Hermes Agent docs, heartbeat
- Hermes Agent compresses long chats at half the context window, keeps a 10,000 to 25,000 token tail and caches Claude prompts; its docs put the saving at about 75% of input cost on multi-turn chats. Hermes Agent docs, context compression and caching
- On our own OpenClaw test install (May 2026) a one-word reply sent 50,968 input tokens with a heavily customised workspace. The heavy setting sits below that. (OneClickClaw test install)
The cost most estimates leave out
An agent that checks in on its own every 30 minutes makes 1,440 model calls a month before you type a word. If each check-in re-reads a long conversation, that is around 144 million tokens a month.
Three settings keep it small, on either agent:
- Give check-ins a fresh short session.
- Space them to your real needs, or pause them at night.
- Send check-ins and scheduled jobs to a small model and keep the big one for conversations.
Quick answers
Are Hermes Agent and OpenClaw free? Both are open source with no licence fee. You pay for the AI model and for the server.
What is the cheapest way to run one? A small model, a free OpenRouter model, or a subscription you already pay for, such as ChatGPT Plus or SuperGrok.
Can I cap the bill? Yes. Set a monthly spend limit on the API key you connect. Your provider enforces it.
Why is my number a range? The same message costs more when the agent carries a long conversation, calls more tools or writes longer replies. The low figure assumes light use and the high figure heavy use.
Related reading
Sources checked on the 8th of October 2026: OpenAI, Anthropic, Google, DeepSeek and xAI pricing pages, OpenRouter limits, Nous Portal plans, and our own plan prices. Currency: European Central Bank rate, 1 EUR = 1.1177 USD (the 7th of October 2026). These are estimates. Your provider's bill is the final word.