What It Actually Costs to Run an AI Agent in 2026

A single AI agent task typically costs between about half a cent and thirty cents in model fees — and the number is almost entirely determined by how many times the agent loops, not by which model you picked. That is the part founders get wrong when estimating AI integration cost. They compare per-token prices across vendors, pick the cheap one, and then get a bill driven by something the price list never mentioned: an agent re-sends its entire accumulated context on every single step.

Here is the anchor number, published by a vendor rather than estimated by us. Anthropic's own pricing documentation works through a customer support agent processing 10,000 support tickets at roughly 3,700 tokens per conversation on Claude Haiku 4.5, totalling about $37.00 — that is $0.0037 per conversation. The same page prices a one-hour agent session on Claude Opus 5 consuming 50,000 input and 15,000 output tokens at $0.705, dropping to $0.525 once prompt caching covers 40,000 of those input tokens.

Those two figures are 190× apart, and both are real. The gap is the whole story: LLM cost per user depends on the shape of the work, not the sticker price of the model.

All rates below were read directly from the Claude Platform pricing docs, the OpenAI API pricing page, and the Gemini API pricing page on 15 August 2026. Model pricing moves fast — check the source pages before you build a financial model on them.

What do the major models cost per million tokens right now?

Published standard-tier rates, USD per million tokens, as of 15 August 2026:

* Google lists Gemini 3.7 Flash at $0.75 input / $3.75 output through 31 December 2026, rising to $1.50 / $7.50 on 1 January 2027. If you are modelling next year's budget, use the higher number.

Two things worth pulling out. The spread between the cheapest and most expensive model here is 25× on input and 25× on output — model choice is the single biggest lever you have. And output consistently costs 5–6× more than input across all three vendors, which means verbose agents are expensive agents.

One detail that catches teams out: Google states explicitly that its output price includes thinking tokens. Reasoning the model does internally and never shows you is billed at the output rate. Anthropic also notes that Claude 4.7 and later models use a newer tokenizer producing approximately 30% more tokens for the same text — so an apparently identical workload can cost 30% more purely from a tokenizer change.

What is the LLM cost per user for an AI agent?

There is no universal number, but the arithmetic is simple enough to do yourself in about two minutes. Cost per task is:

(total input tokens × input rate) + (total output tokens × output rate)

The trap is "total". An agent that takes six steps to finish a job makes six model calls, and each call re-sends the system prompt, the tool definitions, the conversation so far, and every tool result returned up to that point. Context accumulates. Step six is far more expensive than step one.

Here is a worked example. This calculation is ours, not a vendor figure — but every rate in it comes from the published pages linked above. Assume a six-step agent averaging 8,000 input tokens and 500 output tokens per step: 48,000 input and 3,000 output tokens per completed task.

Now convert that to cost per user. If a typical user triggers 20 agent tasks a month, LLM cost per user runs from about $0.26 on Luna to $6.30 on Opus 5. For a SaaS product at $49/month per seat, the cheap end is a rounding error and the expensive end is 13% of revenue before you have paid for hosting, support, or salaries.

That is the calculation to run before you pick a model — not after.

What hidden costs sit on top of the token price?

This is where AI integration cost estimates usually break, because none of these appear in the headline per-token table.

Web search is billed per call, separately. Anthropic charges $10 per 1,000 searches on top of token costs. OpenAI charges $10 per 1,000 calls for reasoning models and $25 per 1,000 calls for non-reasoning models. Google gives 5,000 free grounded search requests per month across Gemini 3.x models, then $14 per 1,000 requests.

Do the arithmetic and this stops being a footnote. If our six-step agent above performs two web searches per task, that is $0.02 per task in search fees. On GPT-5.6 Luna, where the entire task costs $0.013 in tokens, the searches cost more than the model does — roughly 1.5× the token bill.

Tool definitions are billed as input tokens, on every call. Anthropic publishes the exact overhead: enabling tool use adds between 286 and 804 tokens of system prompt depending on model and tool-choice setting. The bash tool adds another 325 tokens, the text editor tool 700 tokens, and computer use 735 tokens plus 466–499 tokens of extra system prompt. Small per call. Multiplied by six steps and 10,000 tasks a month, it is real money.

Code execution and sandboxes are metered. Anthropic gives each organization 1,550 free container-hours per month, then charges $0.05 per hour per container, and bills its Managed Agents session runtime at $0.08 per session-hour. OpenAI bills hosted containers per 20-minute session per container, from $0.03 (1 GB) to $1.92 (64 GB).

Retrieval has its own line items. OpenAI's file search costs $2.50 per 1,000 calls plus $0.10 per GB per day of storage after the first free gigabyte. Embeddings are cheap but not free: OpenAI's text-embedding-3-small is $0.02 per million tokens, text-embedding-3-large $0.13, and Gemini's File Search embeddings $0.15. If your agent sits on a knowledge base, read our breakdown of RAG vs fine-tuning before choosing an architecture — the decision changes your cost curve more than your model choice does.

Data residency carries a premium. Both OpenAI and Anthropic apply roughly a 10% uplift (a 1.1× multiplier) for region-pinned inference. If you have EU or US-only data requirements, add 10% to everything above.

How do you actually cut the bill?

Three levers, all documented by the vendors, in rough order of impact.

1. Prompt caching — the biggest single win for agents. Anthropic prices a cache read at 0.1× the base input rate, with a 5-minute cache write at 1.25× and a 1-hour write at 2×. OpenAI's cached input runs about 10% of standard input (GPT-5.6 Terra: $2.00 standard, $0.20 cached). Because caching pays for itself after a single read on the 5-minute tier, and agents re-send near-identical context on every loop, this is close to free money.

Rerun our six-step example on Claude Sonnet 5 with 6,000 of each step's 8,000 input tokens served from cache, plus one cache write: cost per task falls from $0.126 to roughly $0.076 — about a 40% reduction with no change to model or behaviour. Anthropic's own published example shows the same effect, cutting a one-hour Opus 5 session from $0.705 to $0.525.

2. Batch processing — 50% off, if you can wait. All three vendors discount asynchronous batch work by roughly half: Anthropic states a flat 50% discount on both input and output, Google advertises a 50% cost reduction, and OpenAI's batch table is about half the standard rate across the board. This only works for jobs that do not need an immediate answer — overnight enrichment, bulk classification, scheduled reports. It stacks with prompt caching.

3. Route by task difficulty. Most agent steps are not hard. Classifying an incoming message, extracting a field, deciding which tool to call — these run fine on the cheapest tier. Reserve the expensive model for the one or two steps that genuinely need it. Anthropic's own guidance is to use Haiku for simple tasks, Sonnet for most production workloads, and Opus for the most complex reasoning. A mixed pipeline routinely lands between the Haiku and Sonnet rows in our table above while performing like the Sonnet row.

So what does it cost to build one, not just run one?

Running cost is arithmetic — you can calculate it to the cent from published rates. Build cost is not, and we are not going to invent a range for it.

What we can tell you honestly is what drives it. The variables that move an AI integration cost estimate most are: how many systems the agent must connect to, whether those systems have decent APIs, how much of your data needs cleaning before retrieval works, what your compliance posture requires, and how much human-approval workflow sits around the agent's actions. An agent that reads from one clean API and drafts a reply is a fundamentally different project from one that touches four legacy systems under audit requirements.

If you want a grounded number rather than a range pulled from the internet, our pricing page explains how we scope engagements, and you can talk to us directly about your specific setup. It is also worth reading what AI agents actually do first — a surprising number of projects that get scoped as "agents" turn out to need something simpler and much cheaper.

For what real implementations look like in practice, we have shipped an AI customer support agent, a RAG-powered knowledge base, and an AI document processor — the full set is on our portfolio, and the AI integration service page covers how we approach these builds.

What should you measure once it is live?

Four numbers, tracked from day one:

• Cost per completed task, not cost per API call. A task that takes twelve steps instead of six costs twice as much and looks identical on a per-call dashboard.
• Cache hit rate. If it is low, your prompt ordering is probably wrong — stable content must come first for caching to work.
• Average steps per task. This is the number that quietly doubles. It is also the earliest warning that your agent is looping badly.
• Cost per active user per month. The only figure that tells you whether unit economics work.

Both OpenAI and Anthropic expose token usage in API responses and usage dashboards, so none of this requires custom instrumentation to start.

Frequently asked questions

How much does it cost to run an AI agent per month? For 10,000 tasks a month at six steps each, our calculation from published August 2026 rates gives roughly $132 on GPT-5.6 Luna, $630 on Claude Haiku 4.5, $1,260 on Claude Sonnet 5, and $3,150 on Claude Opus 5 — before tool fees. Prompt caching typically removes another 30–40%.

What is a realistic LLM cost per user? It depends entirely on usage intensity. At 20 agent tasks per user per month, our figures put it between about $0.26 and $6.30 per user. Anthropic's published support-agent example works out to $0.0037 per conversation on Claude Haiku 4.5, which is a reasonable floor for simple, single-turn work.

Is the cheapest model always the right choice? No. A cheaper model that needs three attempts to get a task right can cost more than an expensive model that succeeds first time, and it burns user patience while doing it. Compare cost per successfully completed task, not cost per token.

Do AI agents cost more than chatbots? Yes, substantially — usually several times more per interaction. A chatbot makes one model call per user message. An agent makes several, each carrying the full accumulated context plus tool results. The multiplier is the number of loop iterations.

Does prompt caching work for every agent? It works best when a large, stable block of content — system prompt, tool definitions, a document — is reused across calls. Agents fit that pattern almost perfectly. It helps far less if every request is genuinely unique.

How accurate are the prices in this article? Every rate was read from the vendors' official pricing pages on 15 August 2026 and is quoted as published. The per-task and per-user totals are our own arithmetic applied to those rates, clearly marked as such. Model pricing changes frequently — Google has already announced a Gemini 3.7 Flash increase for January 2027 — so verify against the source pages before committing to a budget.

The short version

Model choice sets the order of magnitude. Loop count sets the actual number. Tool fees are the part nobody budgets for and can exceed your entire token bill on cheap models. Prompt caching and batch processing will take 40–50% off the total for roughly a day of engineering work.

If you are trying to work out whether an AI agent makes financial sense for your product, get in touch — we would rather run the numbers with you honestly than sell you an agent you do not need.

Rates cited from Claude Platform pricing, OpenAI API pricing, and Gemini API pricing, all retrieved 15 August 2026. Per-task and per-user totals are Evolvera's own calculations from those published rates.