Under a video about an AI-run business, one of the top comments asks the question the demo skipped: does this run inside a $20-a-month subscription, or does it need API spend on top, and how much? It's the right question, and it has a better answer than a guess. The running cost of an AI job is arithmetic you can do before you commit: four numbers, some multiplication, and a clear look at what the job replaces.
The two meters
AI vendors bill two ways. A subscription is a flat monthly rate, $20 on the standard tiers at ChatGPT and Claude, with usage caps and a seat at the table: you, interacting, steering, reviewing. API billing is a meter. A call to a model gets charged per token, the unit models read and write, at rates published on a pricing page. It bills like electricity: the more the meter spins, the more you pay.
The subscription buys capacity for work you do in person. The meter serves work that runs without a person watching: a script on a schedule, an agent working through a task list, a pipeline that fires at 2 a.m. Either meter can drive the same job.
The four numbers that price any job
Four numbers set the bill for any scheduled job:
- Input tokens per run: everything the model reads, the instructions plus the material (comments to summarize, a document to check).
- Output tokens per run: everything the model writes back.
- Runs per month: how often the schedule fires.
- The model's rate: published per million tokens (MTok), priced separately for input and output.
A token is roughly three-quarters of an English word; the standard estimate is about four characters per token, though tokenization varies by model and formatting.
Worked example: a job that runs every morning, reads the 50 newest comments on a YouTube channel (about 4,000 input tokens including instructions), and writes back a ranked summary of themes (about 600 output tokens). Thirty runs a month. Rates below are the published prices as of October 20, 2026; they move, so check the page before estimating.
| Model | Input / output per 1M tokens | Cost per run | Cost per month |
|---|---|---|---|
| gpt-6-luna (budget) | $0.10 / $0.50 | about $0.0007 | about $0.02 |
| Claude Sonnet 5.5 (mid-tier) | $2.00 / $10.00 | about $0.014 | about $0.40 |
| Claude Opus 5.5 (premium) | $4.00 / $20.00 | about $0.028 | about $0.84 |
For this shape of job, the API bill is the small number. Forty cents a month on a mid-tier model. Add a second checking pass, where the model re-reads its draft against the source, and output roughly doubles; the bill is still under a dollar. The $200 tier exists for people whose working days look nothing like this job: flagship models, interactively, all day.
Where bills climb
Small bills stay small while the job stays simple. Four things multiply the meter.
Agent loops re-read everything. An agent that plans, acts, checks, and retries makes several calls per task, and each call can carry the working context again. Suppose a task becomes ten calls, each carrying 10,000 input tokens, and the agent writes 2,000 output tokens across the whole task. At the mid-tier rates above, input costs 20 cents and output costs 2 cents: about 22 cents per task, or about $6.60 a month at a task a day. On the premium model, about 44 cents per task, or about $13.20 a month. A long loop on a premium model charges subscription-sized numbers, which is where the flat plans start earning their price back.
Checking passes stack. A draft-and-verify pipeline pays twice by design, once to produce and once to check. The check earns its cost when a wrong answer costs more than another pass. Price it either way.
Reasoning tokens are output tokens. Models that think before answering bill their thinking as output, and thinking can run longer than the answer. A model at $10 per million output tokens that spends 2,000 tokens on thought per run adds two cents a run. At a thousand runs a month, that's $20.
Retries count too. Failed runs may still get billed, and so do the experiments you run while building. Estimates that assume zero waste run low.
One more: rates change. The model you estimated with may be replaced within months by a cheaper one or a pricier one, and published prices move with them.
The other side of the ledger
The meter prices the model's part of the job. The other side is human time, and for simple jobs it's where the money is. The daily comment summary replaces maybe 20 minutes of someone's morning: about 10 hours a month, which at a $25 hourly rate is roughly $250 of time (hours per month times what an hour of that work costs). Against a bill under a dollar, the comparison is lopsided.
The hours that don't show up on any vendor's bill are the ones around the job: writing the instructions, reviewing the output, fixing the step that quietly broke. We budget for those separately, because a cheap meter doesn't make an unmaintained workflow cheap.
Which meter wins
The shape of the work decides:
- Conversational work, you in the loop daily: subscription. A flat rate covers a month of interacting, and a cap is more predictable than a meter spinning while you think.
- Scheduled volume on a capable-enough model: API. The meter prices per unit of work.
- Heavy interactive use of flagship models: the big tiers. The $200 plans exist for people who hit the $20 caps regularly, not for jobs on a schedule.
We run a mix on this site, split on the same line: subscription seats for the work people do in person, metered API calls for jobs that fire on a schedule. The two don't connect to each other; the files carry the state between them.
When you need the pieces:
- How Many AI Subscriptions Do You Need?: the flat-rate side of the decision, and when a second plan earns its place.
- What Is Prompt Chaining? Building Multi-Step AI Workflows: the structure that keeps multi-step jobs checkable before they're billable.
- What Is an AI-Native Company? How We Built One: the operation this costing method comes from, with the pipeline and the numbers visible.
FAQ
How much does it cost to run an AI agent per month?
Depends on the loop. A single daily job (summarize, classify, draft) on a mid-tier model typically prices in cents to a few dollars a month. Multi-step agent loops on premium models, where each step re-reads context, can run from tens of dollars upward at daily frequency. The estimate uses the same arithmetic either way: input and output tokens per step, times steps per run, times runs per month, times the model's published rate.
Is an API cheaper than a subscription?
For scheduled volume on a capable-enough model, usually yes, by a wide margin: the worked example above bills under a dollar where a plan costs $20. For conversational daily use, the subscription wins, since a flat rate covers a month of interacting and APIs would meter every message. Mixed operations commonly run one of each.
What costs besides tokens should I include?
Tool calls often carry their own fees (web search and code execution bill per use on most platforms), and a complete operation adds hosting for the script, storage for whatever the job produces, and any other vendors in the chain. Tokens are usually the smallest line; the platform fees around them are what turn a 40-cent job into a few dollars.
How do I estimate tokens before building anything?
Count the characters the model will read and divide by four. Add the instructions to the material for input, estimate the length of the output you want, then multiply both by the runs per month and the model's published rate. The estimate is short enough to do before building the job; compare it with the usage dashboard once the job has run for a week or two.
Why is my API bill higher than my estimate?
Four usual suspects: retries and failed runs that still got billed, reasoning tokens billed as output, agent steps that re-read the whole context on every call, and experiments you forgot to turn off. The usage dashboard on the vendor's platform shows per-call costs, which is a fast way to find which one it is.
Do I need a $200 plan to run agents?
No. The big plans buy interactive capacity on flagship models for a person at the keyboard; agent mode inside ChatGPT is a plan feature, not an API line item. Custom agents running on a schedule bill through the meter instead, and the table above shows a scheduled job on a mid-tier model costing less per month than a single day of the big plan.