What an AI agent actually costs to run per month

Answer: An AI agent’s monthly run cost is call volume times tokens per call times the model’s per-token price, plus infrastructure and monitoring. At current published API rates (roughly $1–$10 per million input tokens and $5–$50 per million output tokens depending on model tier), a workflow running a few thousand calls a month typically lands in the low hundreds to low thousands of dollars — but almost nobody publishes the actual arithmetic, which is exactly why it’s worth showing.

Why is this number so hard to find?

Because it depends on 4 variables that vary by an order of magnitude each: which model tier you use, how many tokens a typical call actually burns, how many calls the workflow runs per month, and whether you’re paying full retail per-token rates or capturing discounts through caching and batching. Vendors selling agent platforms have an incentive to keep this vague, because a specific number invites a specific comparison against what the workflow is worth. We publish the model instead of a headline figure, because the model is what actually transfers to your workflow.

What does per-token pricing actually look like right now?

Published rates from the major model providers currently sit in roughly this range per million tokens: smaller, faster models around $1 input / $5 output; mid-tier models around $2–$5 input / $10–$30 output; and top-tier reasoning models around $5–$10 input / $25–$50 output. Most providers also offer meaningful discounts for cached input (frequently 90% off the base input rate for repeated context) and for batch processing (roughly 50% off for requests that don’t need an immediate response). A workflow that reuses the same system prompt and reference documents across thousands of calls should be pricing against the cached rate, not the sticker rate — the difference is not marginal.

How do tokens per call actually add up?

A typical agentic call is not just the user’s message. It includes the system prompt, any retrieved context or documents, the tool definitions the agent has access to, and the conversation history if the interaction spans multiple turns. A modestly complex customer-support agent with retrieval and 2–3 tool calls can easily burn 3,000–8,000 input tokens and 300–800 output tokens per resolved ticket. Multiply that by a model’s per-token rate and you get cost per resolution, which is the number that actually matters — not cost per token, which nobody who isn’t an infrastructure engineer should be optimizing for directly.

What does the arithmetic look like end to end?

Take a workflow handling 5,000 calls a month, averaging 5,000 input and 500 output tokens per call, on a mid-tier model priced around $3 input / $15 output per million tokens: that’s 25 million input tokens ($75) and 2.5 million output tokens ($37.50) — about $112 a month in raw inference cost before caching discounts, which on a workflow with repeated context can cut the input line meaningfully. Add monitoring, logging, and the evaluation-set re-runs described in how we evaluate agents before they meet customers, and the realistic all-in number for a workflow at that volume typically lands in the low hundreds to low thousands monthly, not the tens of thousands some platforms imply.

What actually drives the cost up, and how do you catch it?

Runaway cost almost always comes from one of 3 places: a model tier chosen for headline capability rather than the workflow’s actual difficulty, unbounded conversation history that keeps growing the input token count on every turn, or a retrieval step pulling in far more context than the question needs. Every automation and agent we ship carries a cost ceiling and a monitored run counter for exactly this reason — the payback and the run cost should both be a report your finance team can read, not a number that shows up as a surprise on next month’s API bill.

If you want the actual arithmetic run against your specific workflow instead of the general model above, that’s the first thing we do in an agentic engineering build.