How much does it cost to run an AI agent? Real usage data
What 134 AI agents spent in October 2026: requests a day, tokens per request, what a month costs on each model and how far $15 of AI usage goes.
In this note
Most AI agents cost a few dollars a month in AI usage, as long as they run on a low-cost model. Across the 134 Heyron agents in regular use in October 2026, the median one used about $1.50 of AI a month, and nearly three in four used less than $5.
The model changes the bill more than anything else. On Claude Sonnet 5.5, the same median month would cost about $35 at list prices, and on Claude Opus 5.5 about $69.
The AI isn't the whole cost. An agent also needs a computer that stays on, which runs about $5 to $20 a month if you host it yourself, plus the hours to keep it working. On Heyron, that computer, $15 of AI usage and the Solo Software Studio community are all part of the $49 membership.
This page shows how many requests agents made, how big each request was, what a month costs on each model and where the money went. The numbers come from our hosting records as anonymous totals, and we refresh them every month.
How much does an AI agent cost per month?
For most agents, very little. The median active agent used about $1.50 of AI a month, a tenth of the $15 included with Heyron. A quarter used less than 21¢. At the busy end, a quarter used more than $5, and 12 agents were on pace for $15 or more. Nearly all of them ran OpenClaw.
| AI usage a month | Active agents |
|---|---|
| Under 50¢ | 47 |
| 50¢ to $1 | 16 |
| $1 to $2 | 12 |
| $2 to $5 | 23 |
| $5 to $10 | 18 |
| $10 to $15 | 6 |
| $15 or more | 12 |
The spread is wide because a few agents do most of the work. In October, the 10 busiest agents accounted for nearly half of all the AI spend, and they pull the average up to about $5.30 a month. If you're estimating your own bill, start from the median.
Three things decide where your agent lands: how often it works, how much it sends each time, and which model it uses. The next sections take them one at a time.
How many requests does an agent make a day?
Every time an agent answers you or takes a step, such as searching the web, reading a file or using an app, it sends one request to the AI model. So one message from you can turn into several requests.
The median active agent made about 11 requests a day, around 320 a month. A quarter made 3 or fewer a day, and a quarter made more than 37. Fourteen agents made 100 or more.
| Requests a day | Active agents |
|---|---|
| Under 5 | 41 |
| 5 to 10 | 23 |
| 10 to 25 | 24 |
| 25 to 50 | 23 |
| 50 to 100 | 9 |
| 100 or more | 14 |
Requests add up fastest when an agent works on its own. A scheduled job that runs every 15 minutes runs about 2,900 times a month, and each run is at least one request. A message from Telegram or WhatsApp counts the same as one typed in the dashboard.
Why is every request so big?
A typical request carried about 39,000 tokens, roughly 29,000 words, to get back a reply of about 350 tokens. A quarter of agents sent less than 28,000 tokens a request and a quarter sent more than 60,000. On OpenClaw, the median was about 38,000.
That's because an agent re-reads everything on every request: its instructions, the tools it can use, what it remembers about you and the conversation so far. Across all active agents, input made up more than 99% of the tokens. Output costs 4 to 11 times as much per token on the models below, yet input is still more than 90% of what a request costs.
So the cheapest habit is a short conversation. A chat that has run for days re-sends all of it with every message. Start a new chat for each new task; long chats and starting fresh shows how without losing your work.
What does a month cost on each model?
To compare models, we priced every active agent's October requests on each one, kept its web searches and app actions as they were, and took the middle agent. Default is Heyron's own low-cost model, chosen and kept current for you.
| Model | A month |
|---|---|
| Default | $1.30 |
| DeepSeek V4.1 Flash | $1.10 |
| Claude Haiku 5.5 | $2.10 |
| GPT-6.1 Terra | $18 |
| Claude Sonnet 5.5 | $35 |
| Claude Opus 5.5 | $69 |
| Claude Fable 5.1 | $170 |
| Model | One typical request | Typical month | Busy month | Inside $15 |
|---|---|---|---|---|
| Default (Heyron) | 0.23¢ | $1.30 | $4.90 | 93% |
| DeepSeek V4.1 Flash | 0.19¢ | $1.10 | $3.30 | 93% |
| Claude Haiku 5.5 or GPT-6 Luna | 0.41¢ | $2.10 | $6.30 | 89% |
| MiniMax M3.1 Flash | 0.61¢ | $3 | $9.30 | 81% |
| Gemini 3.5 Flash-Lite | 1.3¢ | $6.10 | $19 | 69% |
| GPT-6.1 Terra | 4.1¢ | $18 | $61 | 47% |
| Claude Sonnet 5.5 or GPT-6.1 Sol | 8.2¢ | $35 | $120 | 38% |
| Claude Opus 5.5 | 16¢ | $69 | $240 | 27% |
| Claude Fable 5.1 or GPT-6 Astra | 41¢ | $170 | $600 | 13% |
Real bills on the bigger models are often lower than this. Most models charge less for input the agent re-sends within a few minutes, which is called prompt caching, and agents re-send a lot.
In practice, 98% of active agents ran mostly on Default or another low-cost model, at about 0.23¢ a request. Fewer than 5 ran mostly on mid-priced or premium models, and together they made up 14% of all the spend. That split makes sense. Low-cost models handle routine work well, and a stronger model earns its price on writing people will read, research you'll act on and code. Our model guide says which model to use for which job, and the OpenClaw cost calculator turns these prices into a monthly bill for the number of messages you send.
Where does the money go?
Requests to the AI model were 90% of the spend, and web searches were nearly all the rest. Just over half of active agents searched the web in October, and each search cost about as much as two requests on Default. Actions in connected apps, like adding a calendar event or labelling an email, cost almost nothing.
| What the agent did | Share of the spend | Price |
|---|---|---|
| Requests to the AI model | 90% | about 0.23¢ each on Default |
| Web searches | 10% | 0.5¢ each |
| App actions and paid tools | under 1% | about 0.01¢ per app action |
How far does $15 a month go?
On Default, $15 covers about 6,500 requests, a little over 200 a day. On Claude Sonnet 5.5 it covers about 180, and on Claude Opus 5.5 about 90, at list prices.
At its October pace, the median active agent used a tenth of its allowance, and nearly three in four used less than a third. Eight agents used all of it before October was out. When that happens on Heyron, the agent uses any extra credit you've added, then pauses AI replies until the allowance resets on the 1st. Its files and memory stay as they are. You get an email at 80%, and you can add a $10, $25 or $50 top-up in Agent settings that carries over until it's used. AI usage, models and top-ups has the details.
How to keep your agent's AI bill down
- Leave routine work on Default. For a hard job, start a new chat on a stronger model, then go back to Default once the hard part is done.
- Start a new chat for each new task, so your agent isn't re-sending a long conversation with every request.
- Keep scheduled jobs on Default, and run them only as often as you need. An hourly job makes at least 720 requests a month.
- Look at AI usage in Agent settings after your first week. You'll see where the month is heading long before the 80% email.
What you pay for on Heyron
If the AI is a few dollars a month for most agents, the rest of the cost is the computer it runs on and the time it takes to set up and look after.
Heyron is $49 a month. Your agent gets its own computer, with 2 vCPU and 4 GB of RAM, kept on 24/7, with software updates one button away. Every month includes $15 of AI usage, so there's no API key to set up and no separate bill from a model provider. At October's pace, that covers the median active agent about ten times over on Default.
Members also join Solo Software Studio, a members-only community on Skool for solo founders learning to build and launch apps that make money, with the Build in Public Playbook and other people working through it. See what's included, or read what managed OpenClaw hosting includes.
Would you rather run it yourself? The OpenClaw cost calculator shows what the AI and a server would cost you.
How we measured this
- The period is October 2026, from our hosting records. Most agents' records start on October 16, when we moved every agent to new hosting, so their October was 16 days. Agents created later in October count from their first day.
- We counted members' agents that were still on Heyron when we pulled the numbers on November 12, 2026. The AI chat on Heyron's Help page runs separately and isn't included.
- An agent counts as active if it made at least 20 AI requests in October. That leaves out agents nobody was using, which still make a few automatic requests each week, and gave us 134 active agents. All but a few ran OpenClaw; fewer than 5 ran Hermes, too few to report on their own.
- Monthly figures divide each agent's spend and requests by its days in October, then multiply by 30.
- Spend is what each agent's usage counted against its allowance, after any caching discounts. The model comparison uses OpenRouter's list prices from November 5, 2026, before caching.
- Every number is an anonymous aggregate. We used only counts and costs, never what agents or members wrote, and any group of fewer than 5 agents is reported as "fewer than 5".
- We refresh this page every month. November will be the first full month in our records.
Frequently asked questions
How much does it cost to run an AI agent per month?
On a low-cost model, a few dollars of AI usage for most people. The median active Heyron agent in October 2026 used about $1.50 a month, and nearly three in four used less than $5. The same median month would cost about $35 on Claude Sonnet 5.5 at list prices. You also need a computer to run it on. Hosting one yourself costs about $5 to $20 a month; on Heyron it's part of the $49 membership, which includes $15 of AI usage.
How much does one AI agent request cost?
A typical request sends about 39,000 tokens and gets back about 350. That's about 0.2¢ on a low-cost model like DeepSeek V4.1 Flash or Heyron's Default, 8¢ on Claude Sonnet 5.5 and 16¢ on Claude Opus 5.5, at list prices before caching.
Why do AI agents use so many tokens?
Each request re-sends the agent's instructions, its tools, what it remembers and the conversation so far, so input makes up more than 99% of the tokens. Long chats make every request bigger, which is why starting a new chat for each task saves money.
Is $15 a month of AI usage enough for an AI agent?
For most people on a low-cost model, yes. In October 2026 the median active Heyron agent used about a tenth of its $15, and 93% of agents on low-cost models were on pace to stay inside it. At list prices on Claude Sonnet 5.5, only 38% would have.
What's the cheapest way to run an AI agent?
Host it yourself on a small server and use a low-cost model. That's about $5 to $20 a month for the server, a dollar or two of AI for a typical agent, and 1 to 4 hours a month of upkeep. If you'd rather not run a server, Heyron is $49 a month with the computer, $15 of AI usage and the Solo Software Studio community included.
Do I need an API key to run an agent on Heyron?
No. AI usage is included in every membership, so OpenClaw and Hermes run without an account at a model provider. You pick a model per chat from the model picker under the message box, with Default at the top.