How much does an AI agent cost to run per month in 2026?

In October, the agents I run for Fifth Nine itself did 1,140 runs. At Anthropic’s and OpenAI’s list API prices, that work comes to about $2,400. My collections agent, the one that follows up on unpaid invoices every Monday, cost $5 of it. The agent that coordinates all the others cost $1,092.¹

That spread is the part most cost guides miss. A narrow recurring job is cheap. An agent that reads everything is expensive.

The timing matters because your company may already be paying for some of it. Anthropic announced monthly API credits for its Claude Team plan on October 7, and OpenAI added the same kind of credit to ChatGPT Business in late October. Both are worth up to $500 a month for building your own agents.²³

How much does an AI agent cost per month to run?

An AI agent doing one narrow recurring job costs a few dollars to a few hundred a month in model usage. My collections agent costs about $5 at list prices on Claude Opus 5.5; my coordinator costs about $1,100. It depends on how much the agent reads and writes per run, how often it runs, and which model it uses.

Most small business owners are spending far less than either number on AI today. A Goldman Sachs survey of 2,473 alumni of its small-business program, fielded in August 2026, found that 73% of the owners using AI either use only free tools or spend less than $100 a month.⁴ Most of those owners rely on chat tools like ChatGPT, Claude or Gemini, or AI features built into software they already use.

An agent that does a job on a schedule adds a second bill. It runs on the vendor’s API, which charges by the token, a unit of text a little shorter than a word. You pay for what the agent reads, what it writes, and what it saves to re-read on its next run.

What did my AI agents cost to run in October 2026?

From October 1 to 30, the 16 agents that ran in Fifth Nine’s own workspace did 1,140 runs, worth about $2,400 at list API prices. The coordinator was 46% of that. The website engineer did 283 runs for $103. The collections agent did 11 runs for $5, and the receipts clerk did 23 runs for under $5.

The month, agent by agent, comes from the run records my agents’ platform keeps for every run.¹

Agent and its job	Model	Runs	List-price cost	Per run
Coordinator: reads every open task and routes the work	Claude Opus 5.5	246	$1,092	$4.44
Setup agent: writes and tests the other agents’ instructions	Claude Fable 5.1	64	$362	$5.65
Blog writer: the agent that drafted this post	Claude Opus 5.5	29	$116	$4.02
Website engineer: code changes to fifthnine.com	OpenAI GPT-6.1 Sol	283	$103	$0.37
Reviewer: checks other agents’ work before it ships	Claude Opus 5.5	127 (69 priced)	$49	$0.70
Collections: follows up on unpaid invoices every Monday	Claude Opus 5.5	11	$5	$0.46
Receipts clerk: forwards card receipts to the bank	Claude Sonnet 5.5	23	$5	$0.21
Nine more: marketing, proofreading, video, design, engineering lead, filings, security	Claude Opus 5.5	357	$640	$1.79
Total		1,140	$2,372	$2.08

These are list prices, not my invoice. Most of these runs bill against flat-rate subscription plans rather than metered API usage.

Two things in the table surprised me when I first ran the numbers. The collections agent, the one doing the most valuable job, is one of the cheapest. The agent that drafted this post cost more in October than the agent that writes this website’s code.

Why does one AI agent cost 200 times more than another?

An AI agent’s cost mostly follows how much it reads. My coordinator reads every open task, every reply and every other agent’s report before it decides anything. My collections agent reads one bank account and one list of invoices, on a weekly Monday sweep. Model choice comes second: the receipts clerk runs on a cheaper Claude model at 21 cents a run.

The token split shows where the money goes. Across the month, 94% of the tokens my agents used were cache reads: instructions and history the agent had already loaded, read again at a steep discount. On Claude Opus 5.5 a cache read costs $0.20 per million tokens against $4 for fresh input.⁵ Those reads were still 27% of the bill. Saving context to the cache in the first place was 38%, and the agents’ own output was 29%.

The lesson for an owner is that an agent’s cost follows its reading list. An agent with a short set of rules, one system to check and one thing to produce stays cheap even on a top model. An agent you ask to “keep an eye on everything” re-reads everything, every time it wakes up.

Do ChatGPT Business and Claude Team include API credits for agents?

Yes, as of October 2026. ChatGPT Business and Claude Team each give $20 a month in API credit per Standard seat and $100 per Premium seat, pooled for the company and capped at $500 a month. The credits expire at the end of each billing period and can only be spent on building your own apps and agents.

Anthropic announced its credit on October 7, alongside its new small model, Claude Haiku 5.5, and said the credits “are designed to allow our users to experiment with building tools, apps, and agents that call our API.”² OpenAI’s version followed in late October. Simon Willison, who tracks model pricing closely, called it “pretty much exactly the same deal.”³

Plan	Monthly API credit	Doesn’t cover
ChatGPT Business	$20 per paid Standard seat, $100 per Premium seat, up to $500	Codex used through API keys; your ChatGPT usage limits
ChatGPT Pro ($100, $200 or $500)	Same as the plan price	Same as above
Claude Team	$20 per Standard seat, $100 per Premium seat, up to $500	Interactive Claude Code; Claude through Amazon, Google or Microsoft clouds
Claude Max / Ultra	$100 / $200	Same as above

In practice, a company with ten Standard seats on Claude Team gets $200 a month. OpenAI’s own example: five Standard seats and three Premium seats on ChatGPT Business add up to $400.³ On Claude Team, a Standard seat costs $25 a month billed monthly, and it now brings $20 of API credit with it.⁶

The fine print matters more than usual. An owner of the workspace has to claim the credit and link one API account to it, after the plan has been paid for at least seven days. Unused credit doesn’t roll over. OpenAI says the benefit “may change or be removed in the future.”³ Claude Pro and ChatGPT Plus get nothing, and neither vendor’s Enterprise plan is included.

Is $500 a month in API credits enough to run an AI agent?

$500 in monthly API credits easily covers one or two narrow jobs. My collections agent would use about 1% of $500. A daily summary or a receipts job on a small model would use less. An agent that coordinates everything on a top model is a different story: my coordinator’s October usage would have spent $500 in about two weeks.

The small models are what make the credits go far. Claude Haiku 5.5 and OpenAI’s GPT-6 Luna both list at $0.10 per million input tokens and $0.50 per million output tokens, against $4 and $20 for Claude Opus 5.5.⁵ That’s 40 times cheaper per token. Anthropic pitches Haiku 5.5 for “quick and repetitive workloads (like summaries, compactions, database queries, and classification requests).”² A lot of recurring office work looks like that list.

The catch is quality. Most of my agents run on Opus because the jobs involve judgment and money, and a wrong answer costs more than the model. I’d start a new job on the strong model, measure the cost per run for a week, and only then try it on a smaller one.

What does an AI agent cost besides the model bill?

The model bill is usually the smallest line. The bigger costs are building the agent (connecting it to your systems, writing its rules, testing it on real work), your time approving what it prepares, and someone keeping it working when a vendor changes a model or a tool. The new credits cover none of that.

Some of the platform pieces also bill separately. On Anthropic’s platform, hosted agents cost $0.08 per session-hour of active runtime, and web searches cost $10 per 1,000, on top of the tokens.⁶

If you have one clear job and someone on staff who likes this kind of work, the credits and a few weekends may be all you need. Start with a job where a mistake is cheap and visible, like a weekly report. Leave anything that sends money or messages to customers until you trust it, and keep it behind your approval after that. I wrote about where that line sits in AI agent security.

How do I control an AI agent’s monthly cost?

Set a hard limit before the first run, keep each agent’s instructions and reading list short, and put simple jobs on small models. Turn off auto-reload, so requests stop when the credit runs out instead of charging your card. Check the cost per run after the first week, and be ready to move a job to another vendor’s model.

Both vendors make the hard stop easy. On Anthropic’s platform, if your organization has no other credits, “API requests stop until your next monthly credits arrive,” and you can set spend limits per project or teammate.² On OpenAI’s, requests fail once no credits remain, unless you’ve added paid credits or auto-recharge.³

Don’t let the credits pick your vendor, either. Sometimes Claude doesn’t do a good job on a task, and I move that agent to an OpenAI Codex model. In October my website engineer ran on OpenAI’s GPT-6.1 Sol. My agents can move between Anthropic’s Claude, OpenAI’s Codex and open-source agent software, and the setups Fifth Nine builds for clients work the same way. Use the free credit on whichever plan you already pay for, and build so the agent can leave.

For what my collections agent actually does each Monday, see Should you let QuickBooks AI send your invoice reminders?. If you’re still deciding whether you need an agent at all, start with AI agent vs ChatGPT.

If you want to know what one of your recurring jobs would cost to hand to an agent before anyone builds it, Fifth Nine’s AI Audit prices out the job, the model and the approvals it needs.

---

Sources

How this was written: drafted with AI from my notes and Fifth Nine’s operating records, then edited and approved by me.

Footnotes

1. From the run records of Fifth Nine’s agent workspace, October 1–30, 2026: every run’s input, output, cache-read and cache-write tokens, priced at Anthropic’s and OpenAI’s list API prices on October 31, 2026, with Anthropic’s 5-minute cache-write rate. At Anthropic’s 1-hour cache rate the month comes to about $2,900. 54 of the reviewer’s runs, from October 20 to 22, didn’t record which model they used, so their tokens are left out; it has run on Claude Opus 5.5 since October 23. Three agents didn’t run in October. These are list-price equivalents, not what Fifth Nine paid. ↩ ↩²

2. Anthropic, “Introducing Claude Haiku 5.5”, October 7, 2026, and Anthropic’s help article “Monthly API credits for Max, Ultra, and Team plans”, checked October 31, 2026. The date is from Simon Willison’s post on the launch. ↩ ↩² ↩³ ↩⁴

3. OpenAI Help Center, “API credits benefit for eligible ChatGPT subscriptions”, checked October 31, 2026; Simon Willison, “API credits benefit for eligible ChatGPT subscriptions”, October 29, 2026. Business workspaces billed annually get the year’s credit at once, capped at $6,000. ↩ ↩² ↩³ ↩⁴ ↩⁵

4. Goldman Sachs 10,000 Small Businesses, “The AI Economy Is a Small Business Economy”, October 2026, a survey of program alumni fielded August 10–24, 2026. I compared it with other 2026 surveys in How many small businesses use AI?. ↩

5. Anthropic, Claude pricing and pricing documentation; OpenAI, API pricing. Both checked October 31, 2026. Haiku 5.5’s prices apply to prompts up to 100,000 tokens; longer prompts cost five times as much. ↩ ↩²

6. Anthropic, Claude pricing, checked October 31, 2026. A Claude Team Standard seat is $20 a month billed annually or $25 billed monthly. ↩ ↩²