What an AI Agent Actually Costs to Run Each Month

An AI agent that triages 500 support emails a month costs between roughly 56 cents and $56 a month in model tokens, depending entirely on which model you point it at. The runtime, where your vendor meters one, adds cents rather than dollars at that volume, and on Google's meter it falls inside the free tier altogether. The figure that moves your bill is not the price per token — it is how many turns the agent takes, because it re-reads its whole context on every one.

All prices below are from each vendor's own published pricing pages, checked October 11, 2026. Both OpenAI and Anthropic have new models at the head of their price tables since this page was last priced in September, and the old ones are still listed underneath, so the floor is lower than it was then: the cheapest row this page quoted in September works out at a little over twice the cheapest row today.

The job being priced

Concrete beats generic, so here is a specific agent. It reads an inbound support email, looks the sender up in the CRM, checks their most recent order status, then writes a suggested reply and applies a category label. Three model turns: read and call a tool, receive the tool result and call a second tool, receive that and produce the answer. Five hundred emails a month.

Reasonable token counts for that shape of job:

• System prompt, routing rules and tool definitions: 1,500 tokens, fixed, sent on every turn
• The email itself: 600 tokens
• CRM lookup result: 400 tokens. Order status result: 300 tokens
• Model output: about 150 tokens for each tool call, 350 for the final drafted reply

These are estimates for illustration, not measurements, and one vendor now says in writing that a token count measured on an earlier model should not be reused to estimate costs on a later one. That caveat has its own section below, and the last section covers how to replace these figures with measured ones.

The part that gets billed again every turn

The API has no memory. Each request carries the entire conversation so far, and you are billed for all of it again. OpenAI's conversation state guide puts it plainly: "While each text generation request is independent and stateless, you can still implement multi-turn conversations by providing additional messages as parameters to your text generation request." The same page warns that even when you hand OpenAI a response ID instead of the messages, "all previous input tokens for responses in the chain are billed as input tokens in the API."

Anthropic's context windows documentation describes the same mechanic: "as the conversation advances through turns, each user message and assistant response accumulates within the context window, and previous turns are preserved completely." Each turn's input phase "contains all previous conversation history plus the current user message," and "everything in the request counts toward the context window: the system prompt, every message in messages (including tool results, images, and documents), and your tool definitions." The same page adds that the output of each turn, including extended thinking, counts too.

Run the arithmetic on the triage agent and the effect is stark.

Today's published token prices

Per million tokens, standard rates, read on October 11, 2026 from OpenAI's API pricing page, Anthropic's pricing documentation, and the Gemini API pricing page. OpenAI's flagship table shows only its first three by default and Anthropic's table opens with these four; on Google's page each model family has its own section, and the Google rows below were re-read on October 28, 2026, when Gemini 4 Argon had taken the top of the page, with 3.8 Flash under it and Flash-Lite and 3.1 Pro Preview under several older Flash models. All three keep earlier models listed as well, and at OpenAI and Anthropic some of those cost more than today's headline row: gpt-5.5-pro at $30.00 input, Claude Opus 4.1 at $15.00.

Four footnotes that matter, and the first one is about this table rather than in it. The three OpenAI figures are its short-context column. The same table carries a second set of columns for long context, where across all three models the input rate doubles and the output rate rises by half — $20.00 and $75.00 on astra, $4.00 and $15.00 on sol, $0.20 and $0.75 on luna. The prompts in this article top out at 3,100 tokens, and the page does not publish the threshold where short becomes long — but an article about context piling up should say that the price of context piling up far enough is a different set of columns. Google's page states that Gemini 3.8 Flash is $0.75 input and $3.75 output "through December 31, 2026", rising to $1.50 and $7.50 "starting January 1, 2027" — a doubling that lands inside most 2027 budget cycles. Gemini 4 Argon doubles sooner: $2.00 and $10.00 "through November 30, 2026", then $4.00 and $20.00 "starting December 1, 2026". Gemini 3.1 Pro Preview charges $4.00 input and $18.00 output once a prompt passes 200,000 tokens. And Claude Haiku 5.5 is $0.10 and $0.50 only for prompts up to 100,000 tokens; above that it is $0.50 and $2.50, a five-fold step that a long agent loop can walk into without anyone noticing.

The monthly bill

At 7,850 input and 650 output tokens per email, 500 emails a month is 3.925 million input tokens and 325,000 output tokens. Multiply through:

Figures are rounded to the cent, which is why the two cheapest rows show components that do not quite sum: the exact total on both is $0.555.

Two things stand out. The top and bottom of that table are a hundred times apart for identical work, which makes model choice the single largest lever on this bill — larger than caching, batching or anything else below. And in the middle, three models from three different vendors come out at exactly the same monthly figure: gpt-6.1-sol, Claude Sonnet 5.5 and Gemini 4 Argon all land on $11.10, with Gemini 3.1 Pro Preview 65 cents above them. On mid-tier models at this volume price is no longer a tiebreaker at all, and the decision moves to output quality and to whichever vendor's tooling fits how you already build. Our comparison of AI agent platforms covers that side of it.

The assumption in that table, and why the vendor tells you not to make it

Every row above applies the same 7,850 input tokens. That is the convenient assumption, and Anthropic's token counting documentation now tells you not to reuse counts across its own generations: "Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer. The same input text produces approximately 30 percent more tokens than on earlier models. The exact increase depends on the content and workload shape. Recount prompts against the model you plan to use rather than reusing counts measured against earlier models."

The rule of thumb everyone reaches for has the same limit. Google's token documentation says "a token is equivalent to about 4 characters" and "100 tokens is equal to about 60-80 English words" — but it says it "for Gemini models," not for tokenizers in general.

Caching and batching

The 1,500-token system-and-tools block is identical on all three turns, which is exactly what prompt caching is for. Anthropic's pricing page prints cache write and cache hit rates for every model it lists, and adds that all other models use the standard 0.1x multiplier. Most of the older rows do work out at a flat 10% of input; the newer ones do not: on Sonnet 5.5 a five-minute cache write is $2.50 against a $2.00 input rate, and a cache hit is $0.10, which is 5% of input rather than the flat 10% that used to apply across the range. Apply that to the fixed prefix inside each email's loop and the Sonnet 5.5 bill drops from $11.10 to about $8.63 a month.

That is a real 22% saving and it is also smaller than it looks, because a five-minute cache does not survive the gap between one email and the next. Caching helps within a loop, not across a month's worth of them.

Batching helps more, if your job can wait. Anthropic's Batch API is a flat 50% off both input and output, which halves Sonnet 5.5 to about $5.55. That works for a triage queue processed hourly or overnight. It does not work for anything a customer is waiting on.

The runtime charge, where there is one

All three will now run the agent loop for you if you let them, and two of the three publish an hourly rate for doing it. Anthropic's Managed Agents adds $0.08 per session-hour on top of tokens, metered only while a session is actually running rather than idle; its Agent SDK overview describes sessions running either in an Anthropic-managed cloud sandbox or in a self-hosted one on your own infrastructure. OpenAI's Agents API runs the loop on a managed harness of its own, and its pricing page publishes no hourly runtime rate for it. Google publishes the most detailed meter of the three — and the page that documents it has moved, with the old Vertex AI pricing URL now redirecting to Gemini platform pricing.

Agent hosting there is billed on three SKUs. Agent Compute is $0.085 per vCPU-hour after a free tier of the first 50 hours a month per account. Agent Memory is $0.009 per GiB-hour after the first 100 gibibyte-hours a month, the free tier applying per account for each resource. Agent Storage is $0.000410959 per gibibyte-hour after the first free gibibyte-month, which the page's own Memory Bank section restates as $0.30 per GiB-month. One- and three-year Flexible Savings Plans bring the compute rate down to $0.0765 and $0.068. The page adds: "All usage is rounded to the nearest second. For Runtime, idle time spent waiting for the next prompt between turns is not billed."

Hosted tools are the charge that does sneak up. OpenAI and Anthropic both price their main web search tool at $10.00 per 1,000 — calls on one side, searches on the other, with Anthropic counting one use per search regardless of how many results come back. OpenAI charges $25.00 per 1,000 instead for web search preview on non-reasoning models, and that tier does not bill the search content tokens. Anthropic's documentation adds that search results "are counted as input tokens," in that turn and in every later turn of the conversation, so a search does not just cost $0.01, it inflates the context you re-read for the rest of the loop. OpenAI separately prices its hosted shell and code interpreter per 20-minute session per container, from $0.03 for 1 GB to $1.92 for 64 GB.

Compare it against the thing it replaces

Three minutes of human handling per email, across 500 emails, is 25 hours. US private establishments with 1 to 49 workers paid $37.64 an hour in total compensation in June 2026, per BLS Employer Costs for Employee Compensation, Table 6 — so 25 hours is about $941 a month of labor against 56 cents to $56 of tokens.

That gap is why agents get built, and it is also why the token bill is almost never the number that decides the project. Review time, error handling and the cost of a wrong answer reaching a customer dominate. We work through that calculation properly in how to calculate automation ROI.

Measure your own, because these numbers are not yours

Every figure above rests on assumed token counts. Real prompts differ enormously — a verbose system prompt, a long tool schema, a customer who pastes an entire email thread, or a model that decides on six turns instead of three will all move your bill by multiples, not percentages.

Three ways to get real numbers, in increasing order of usefulness:

1. Rule of thumb. Google's 4-characters-per-token figure, quoted above, is good enough for a first sanity check and nothing more.
2. Count before you send. Anthropic offers a token counting endpoint that is, in its words, "free to use but subject to requests per minute rate limits," with published per-tier ceilings of 5,000, 10,000 and 20,000 requests a minute; its FAQ notes that the count "provides an estimate without using caching logic." Google's API has a count_tokens method that "returns the total number of tokens in the input only." Count against the exact model ID you intend to ship, for the reason in the tokenizer section above.
3. Log what you were actually billed. Responses carry a usage object, and Google's documentation names its fields: input, output, thinking, cached content, tool use and total — and thinking tokens are the line to watch, because they are billed as output and never appear in anything you wrote. Run 20 real emails through the agent, sum the usage across every turn of each one, and take both the average and the worst case. Multiply the average by your monthly volume for the budget, and the worst case by your volume for the number you give your finance person.

The third method is the only one that catches the failure that actually blows budgets: an agent that loops more times than you designed it to. Token prices are published and stable enough to plan around — all four pricing pages read for this article were identical on October 9 and on October 11, while the models at the head of two of them are not the ones this page quoted in September. Turn counts are not published anywhere, and they are where the money goes.