What does it cost to run an AI agent in production, per month?

How much does an AI agent cost to run, once it's live? Based on the same market ranges we've verified for the build side — API usage at $100–10k a month and hosting at $200–5k a month, published across vendor guides, not our own pricing — a single production agent's direct infrastructure bill runs roughly $300 to $15,000+ a month before the number that actually blows past most budgets: agentic workloads don't consume tokens the way a chatbot does, and the gap between a pilot's monthly bill and production's is rarely linear.

The reason is architectural, not incidental, and it is documented in the API rather than forecast by an analyst. A model API is stateless: every step of an agentic task is a separate request that carries everything before it. Anthropic's own documentation states that "as the conversation advances through turns, each user message and assistant response accumulates within the context window, and previous turns are preserved completely," and that "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" (Anthropic, Context windows). Tool use adds its own overhead on top: the documented sources of extra tokens are the tool schemas you send, every tool-call and tool-result block in the exchange, and an additional system prompt the API injects whenever any tool is present (Anthropic, Tool use with Claude). A chatbot answers a question once, on one request. An agent that verifies a claim, calls a tool, and re-reads its own output before responding issues a request per step and re-pays for a longer prefix each time — which is exactly why a monthly estimate built by multiplying "expected requests" by "cost per chatbot call" undercounts from the first line.

That is also why falling per-token prices have not turned into falling monthly bills, and why the two numbers get confused so often. Price per token is the number that keeps dropping and the number vendors quote. Tokens per completed task is the number that is actually driving your invoice, and it moves in the opposite direction every time you add a step: another tool call, another verification pass, another retry on an ambiguous input. Cost per completed task is the product of the two, and it can rise while the price per token falls. Any monthly estimate you are handed should be expressed in that unit. A quote in cents per thousand tokens is a quote in the one unit that does not tell you what next month costs.

The scale this reaches once an organization has more than a handful of agents live is large, and worth stating with its measurement status attached. CloudZero's State of AI Costs report, a survey of 500 engineering professionals at manager level and above, measured average enterprise AI spend at $62,964 a month in 2024 and said its findings suggest that would “rise to $85,521 in 2025 (a 36% increase)”, with the proportion of organizations planning to invest over $100,000 per month “set to more than double, jumping from 20% in 2024 to 45% in 2025” (CloudZero, State of AI Costs, 2025). The 2025 figure is that report's projection rather than a later measurement of it, and no subsequent edition has confirmed the outturn — so treat the 2024 number as the observation and the 2025 one as the direction of travel. Most of that isn't one runaway agent; it's the same per-task multiplier compounding across a growing agent portfolio, each one added on the assumption it would cost roughly what the last one did.

The budgeting gap underneath that growth is a measurement gap first. Most teams track AI spend exactly one way: the monthly invoice from the model provider, plus the hosting bill. An account-level number tells you the bill went up. It cannot tell you which capability did it, whether the rise came from more traffic or from more steps per task, or whether a prompt change three weeks ago quietly doubled the retry rate on one workflow while everything else held steady. Teams with mature cost tracking are not the ones spending badly — they are the only ones who find out early enough to do something about it. Everyone else meets the overrun when finance asks about it, by which point the month is already spent. What you stop that night, once the ceiling is already crossed and the feature is still answering, is when the monthly cap is already spent.

The failure mode we see in scoping calls follows a specific pattern: a pilot's monthly estimate gets built from a demo's traffic — a curated set of clean requests where the agent mostly finishes in one or two model calls. Production traffic isn't curated. An ambiguous request triggers a clarifying tool call; a low-confidence output triggers a self-check pass before the agent will commit to an answer; a failed API call triggers a retry. None of those are bugs — they're the same reasoning steps that make the agent trustworthy enough to ship — but each one multiplies the token cost of that one task, and a budget built on the demo's call count has no line item for any of it.

The lever that actually controls this isn't a cheaper model — it's the confidence threshold that decides which outputs get automated approval and which get routed to a human review queue. We built exactly that split into a generative catalog-enrichment pipeline for a marketplace operator: low-confidence listings queue for human review, everything above threshold is approved automatically, and the whole pipeline runs against continuous evaluation with drift alerts rather than a one-time launch check. Move that threshold too loose and the token cost of reasoning and re-checking climbs along with the error rate; move it too tight and the human-review queue — the actually expensive line item — grows instead. Neither number is right by default. It has to be measured against your own traffic and revisited as that traffic shifts, not set once at launch and left alone.

That measurement is the actual product of an evaluation and monitoring practice, not a compliance checkbox bolted onto the build. AI evaluation and observability done right gives you the two numbers that predict next month's bill before it arrives: retry rate and escalation rate, tracked per capability, not averaged across the whole agent. A rising retry rate on one capability is an early warning that's cheaper to catch before the invoice than after it — the same measure-then-gate discipline that keeps accuracy from drifting also keeps the monthly bill from drifting, because both are driven by the same underlying number.

None of this argues against production agents — it argues against budgeting them like a chatbot subscription. Ask a vendor for cost per completed task, not cost per model call, and ask them to show the retry and escalation rates from a comparable production deployment, not a demo. That's the same three-line discipline behind our own AI development pricing: build cost, monthly run cost, and the multiplier that separates a pilot's bill from production's — quoted separately, because folding them into one number is how a $300-a-month estimate turns into a $15,000 one with nobody able to say exactly when it happened.