Costs, in plain English
How much does it cost to run an AI agent?
Short answer: inside an app plan you already pay for, your bot runs within that plan's limits. On a pay-per-use AI API, the 11 everyday bots in our blueprints come to between under 1¢ and $5.28 a month, depending on the bot and the model. Every number below is worked out from official prices checked Oct 21, 2026.
Updated · By BotsDoThat
What sets the price
- How often it runs. Once a weekday is 22 runs a month; every day is 30.
- How much it reads and writes each run. AI APIs charge per token, a small chunk of text a few characters long, with one price for what the model reads and another for what it writes.
- Which model it uses. Within one company, the small model can cost up to 100× less than the flagship in our AI price table.
What everyday bots cost a month
Our estimates for each blueprint, at standard API prices, cheapest and priciest model in our table. Click “Try it” to open the calculator with that bot loaded.
| Bot | Runs a month | Tokens per run (in / out) | Cheapest | Priciest | |
|---|---|---|---|---|---|
| Morning Inbox Briefing | 22 | 15,000 / 800 | $0.03 Groq openai/gpt-oss-20b | $4.18 Anthropic Claude Fable 5.1 | Try it → |
| School Email Digest | 4 | 20,000 / 1,000 | $0.01 Groq openai/gpt-oss-20b | $1.00 Anthropic Claude Fable 5.1 | Try it → |
| Web Page Watcher | 30 | 7,500 / 200 | $0.02 Groq openai/gpt-oss-20b | $2.55 Anthropic Claude Fable 5.1 | Try it → |
| Scam Checker | 30 | 3,000 / 300 | $0.01 Groq openai/gpt-oss-20b | $1.35 Anthropic Claude Fable 5.1 | Try it → |
| Weekly Meal Plan Bot | 4 | 6,000 / 2,000 | under 1¢ Groq openai/gpt-oss-20b | $0.64 Anthropic Claude Fable 5.1 | Try it → |
| Family Calendar Guardian | 4 | 8,000 / 600 | under 1¢ Groq openai/gpt-oss-20b | $0.44 Anthropic Claude Fable 5.1 | Try it → |
| Travel Day Guardian | 30 | 10,000 / 300 | $0.03 Groq openai/gpt-oss-20b | $3.45 Anthropic Claude Fable 5.1 | Try it → |
| Follow-Up Finder | 4 | 30,000 / 1,000 | $0.01 Groq openai/gpt-oss-20b | $1.40 Anthropic Claude Fable 5.1 | Try it → |
| Job Match Scout | 22 | 20,000 / 800 | $0.04 Groq openai/gpt-oss-20b | $5.28 Anthropic Claude Fable 5.1 | Try it → |
| Receipt Matcher (small business) | 4 | 15,000 / 1,000 | $0.01 Groq openai/gpt-oss-20b | $0.80 Anthropic Claude Fable 5.1 | Try it → |
| New Inquiry Logger | 40 | 1,500 / 200 | $0.01 Groq openai/gpt-oss-20b | $1.00 Anthropic Claude Fable 5.1 | Try it → |
Token counts are our editor's estimates, not measurements. Not included: caching discounts, tool or web-browsing fees, taxes and free tiers. Agents that browse or use many tools can use several times more tokens.
Worked example: the Morning Inbox Briefing
The Morning Inbox Briefing reads your new email every weekday morning: about 15,000 tokens in and 800 out each run, 22 runs a month (our estimate). The math is runs × (tokens in × input price + tokens out × output price), with prices per 1 million tokens:
- Cheapest: Groq openai/gpt-oss-20b: 22 runs × (15,000 × $0.075 + 800 × $0.30) ÷ 1,000,000 = $0.03 a month
- Cheapest mid-size model: Groq openai/gpt-oss-120b: 22 runs × (15,000 × $0.15 + 800 × $0.60) ÷ 1,000,000 = $0.06 a month
- Priciest: Anthropic Claude Fable 5.1: 22 runs × (15,000 × $10.00 + 800 × $50.00) ÷ 1,000,000 = $4.18 a month
App plan or pay-per-use?
Agent apps like Claude Cowork, ChatGPT dots, Gemini Spark and Grok Bot run inside a plan instead of billing per token. Plan prices we've checked on official pages:
| Plan | Price | Details | Source |
|---|---|---|---|
| Anthropic Claude Free | $0 | Official page | |
| Anthropic Claude Pro | $20 | per month billed monthly ($17/month billed annually, $200 up front) Excludes tax. Includes "Hand off and schedule tasks". Usage limits apply. | Official page |
| Anthropic Claude Max | from $100 | per month 5x or 20x Pro usage; includes $100 to $200 a month in API credits. | Official page |
| Mistral Free | $0 | Limited access. | Official page |
| Mistral Pro | $14.99 | per month Excluding taxes. Students: $5.99/month. | Official page |
| OpenAI ChatGPT Free | $0 | Official page | |
| Google Gemini (free) | $0 | With a Google Account. | Official page |
For a few small bots, pay-per-use is often a few dollars a month or less. A plan you already pay for may be simpler. Compare: decide between a monthly AI plan and paying per use.
Open-source agents: OpenClaw and Hermes Agent
The software is free; you pay for the model and the machine. OpenClaw has “no paid tier, hosted service, or token” (OpenClaw README), and its minimum server is 1 vCPU and 1 GB of RAM (OpenClaw FAQ: providers and hosting). Hermes Agent is free under the MIT license; Nous Research sells optional model and tool credits through Nous Portal, and cloud hosting is optional (Hermes Agent site). Hermes' docs also show how to run it on your own machine with Ollama at no API cost (Hermes docs: Run Hermes locally with Ollama).
Self-hosted agents send their own instructions, memory and tool lists with every message, so they can read more tokens per run than our estimates. Hermes' docs call this a fixed per-message cost and show how to check it with /usage and hermes prompt-size (Hermes docs: Tips and best practices). Full comparison: OpenClaw vs Hermes Agent: a beginner's guide.
Phone and voice bots are priced per minute
Bots that talk, like the Phone Call Bot, are billed by the minute instead of by the token. Examples from official pages:
- xAI Grok Voice speech to speech (grok-voice-3.0): $0.08 a minute, so a 10-minute call is about $0.80 (official page)
- OpenAI GPT-Live 1 voice sessions: $0.05 a minute, so a 10-minute call is about $0.50 (official page)
- Mistral Voxtral 3 voice agent API with Mistral Small 4: $0.04 a minute, so a 10-minute call is about $0.40 (official page)
How to pay less
- Set spend limits and usage alerts on your AI accounts
- Keep your bot's prompts, tools and context short
- Stop always-on bots burning tokens: run them less often
- Run open AI models on your own computer for simple jobs
- Cut your AI bill with prompt caching and batch discounts
- Decide between a monthly AI plan and paying per use
- Build and test your bot on a free tier before you pay
- Use a small AI model for simple jobs, a big one for hard ones
Batch APIs take 50% off at Anthropic, Google, Mistral, OpenAI for jobs that can wait. All the tips: Run your bots for less.
Questions people ask
How much does it cost to run an AI agent?
Inside an app plan, a bot runs within your plan's usage limits. On a pay-per-use AI API, our 11 blueprint estimates range from under 1¢ to $5.28 a month depending on the bot and the model, at standard API prices checked Oct 21, 2026. Agents that browse the web or use many tools can use several times more tokens than these estimates.
What decides the price of an AI agent?
Three things: how often it runs, how much it reads and writes each time (counted in tokens), and which model it uses. Within one company, the small model can cost up to 100× less than the flagship in our price table.
Is a ChatGPT or Claude plan cheaper than paying for the API?
For a few small bots, pay-per-use API prices are often a few dollars a month or less. App plans charge a flat monthly fee with usage limits and include the app itself. If you already pay for a plan that can run scheduled tasks, that may be the simplest option. More: decide between a monthly AI plan and paying per use.
How much does OpenClaw or Hermes Agent cost?
Both are free, open-source software. You pay for the AI model you connect (an API key, or a plan you already have where supported) and for the computer it runs on, unless you run it on a machine you already own. OpenClaw says it has no paid tier or hosted service (OpenClaw README). Nous Research says Hermes Agent is free under the MIT license and that model providers and optional hosted services have their own pricing (Hermes Agent site). Details: OpenClaw vs Hermes Agent.
What's the cheapest way to run Hermes Agent?
Run it on a computer you already own and point it at a local model: Hermes' own guide to running with Ollama is titled “Zero API Cost” and lists 8 GB of RAM as the minimum for small 3B models (Hermes docs: Run Hermes locally with Ollama). The trade-off: OpenClaw's docs warn that smaller models are more open to prompt injection (OpenClaw docs: Prompt injection). The next cheapest is a small pay-per-use model; the cheapest in our table is Groq openai/gpt-oss-20b (AI price table).
Why did my AI agent cost more than I expected?
Usually because it ran more often than needed, read more than needed (long threads, whole web pages, big tool lists), or used a bigger model than the job needed. Always-on agents that check for changes every few minutes add up fast: stop always-on bots burning tokens: run them less often. See also keep your bot's prompts, tools and context short.
How do I avoid a surprise AI bill?
Set a spend limit and usage alerts on your AI account before anything runs on autopilot (set spend limits and usage alerts on your AI accounts), test on a free tier first (build and test your bot on a free tier before you pay), and check the first week's usage.
Next step
Pick a bot, open the calculator with it loaded, and set a spend limit before it runs on autopilot. New to agents? Start with how to build an AI agent without coding, and before you connect your inbox read is it safe to give an AI agent your email and calendar?