How much will AI agents cost?

Short answer: an AI agent’s running cost is metered, not mysterious. Monthly cost = tasks per month × model calls per task × tokens per call × the model’s price per token. For a typical finance workflow that lands anywhere from a few dollars to a few thousand dollars a month, and which end you land on depends mostly on one decision: which model you use. Caching and batching then cut the bill by a third or more.

Why is AI cost so hard to predict?

It isn’t, once you know four numbers. Most budgeting trouble comes from treating AI like a software licence: one price per seat, paid once a year. Model usage works more like a utility. You pay for what you consume, measured in tokens, and consumption scales with how much work the agent does.

The four numbers that set the bill:

• Volume. How many tasks the agent handles each month.
• Calls per task. How many times the agent calls the model to finish one task.
• Tokens per call. How much text goes in and comes out each time.
• Price. The model’s rate per million input tokens and per million output tokens.

Know those four and you can forecast spend within a useful range before you build anything.

Token: the unit AI models read and write. One token is roughly three quarters of an English word, so 1,000 tokens is about 750 words. Providers price usage per million tokens, and output tokens cost more than input tokens.

How many tokens does an agent use per task?

More than a chatbot, because an agent works in steps. A chatbot answers a question with one model call. An agent that triages an invoice might make five: read the invoice, look up the purchase order, compare the two, decide whether it matches, and draft a note for the approver. Each step is a model call, and each call usually resends context from the steps before it.

So the multiplier that catches budgets out is not the price per token. It is calls per task. Budget 1 call per task for simple work, and 4 to 8 for agents that use tools and check their own work.

A worked example: an invoice triage agent

Fictional numbers, chosen to show the arithmetic. An accounts payable team wants an agent to triage 4,000 invoices a month.

• Calls per invoice: 5
• Input per call: 3,000 tokens, which covers the invoice, the PO and instructions, about four pages of text
• Output per call: 400 tokens
• Model: Claude Sonnet 5.5, listed at $2 per million input tokens and $10 per million output tokens

Cost per call: 3,000 × $2 / 1,000,000 = $0.006 for input, plus 400 × $10 / 1,000,000 = $0.004 for output, which is $0.01 per call. Cost per invoice: 5 calls × $0.01 = $0.05. Cost per month: 4,000 invoices × $0.05 = $200.

That is the baseline. Now the levers.

What changes the bill the most?

1. Model choice

The same workload costs very different amounts on different models. Here is the invoice agent priced across three tiers at list prices, with no caching or batching:

That is 100 times between the frontier and fast tiers for identical volume, and the other providers show the same spread. OpenAI lists GPT-6 Astra at $10 in and $50 out, and GPT-6 Luna at $0.10 in and $0.50 out. Google lists the Gemini 3.1 Pro preview at $2 in and $12 out, and Gemini 3.5 Flash-Lite at $0.30 in and $2.50 out.

The practical move is model routing. Send the simple steps, like extracting fields or classifying a document, to a fast model, and keep the judgement steps on a stronger one. In our example, if three of the five calls are simple enough for Haiku 5.5, the monthly bill drops from $200 to about $86, and quality holds where it matters.

2. Prompt caching

Agents resend the same instructions and reference material on every call, and providers charge far less for input they have recently seen and cached. Claude Sonnet 5.5 lists cache hits at $0.10 per million tokens, against $2 for fresh input. If 60% of the invoice agent’s input is served from cache, the month costs about $132 instead of $200, roughly a third less, with no change to the work.

3. Batch processing

Work that can wait a few hours, such as overnight reconciliations, month-end reviews and document backlogs, can go through a provider’s batch API. Anthropic, OpenAI and Google all list batch processing at 50% off standard input and output prices. Batch the invoice agent’s overnight run and $200 becomes $100.

Stack the levers and a workload that would cost $1,000 a month on a frontier model can run for well under $100, with nothing given up on the decisions that need the stronger model.

How should you set an AI budget?

Budget it like a utility, not a licence.

• Forecast each workload from the four numbers, and add monthly growth as adoption spreads.
• Set a monthly budget with alerts at 50%, 80% and 100% of plan.
• Tag usage by workload, so you can see which agent drives the bill.
• Review monthly against actuals. Most drift comes from volume growth or longer prompts, and both show up in the data.
• Keep people in the loop for key decisions. The cheapest agent is the one that doesn’t create rework. We recommend a human review for decisions that carry financial or customer risk, and capping automation at around 80% for most tasks.

Plan your own numbers

The AI Budget Planner does this arithmetic for your workloads. It prices them on current list prices from Anthropic, OpenAI and Google, projects twelve months, shows what each lever saves and tracks actual spend against plan. It’s free, and your inputs never leave your browser.

Frequently asked questions

How much does it cost to run an AI agent per month?

It depends on volume, calls per task, tokens per call and the model. As a worked example, an agent triaging 4,000 invoices a month with five calls each costs about $200 a month on a balanced model at list prices. The same work costs about $10 on a fast model and about $1,000 on a frontier model.

Why do AI agents cost more than chatbots?

Agents complete tasks in several steps, and each step is a separate model call that often resends earlier context. A five-step agent uses roughly five times the tokens of a single chatbot answer for the same request.

What is prompt caching?

Prompt caching lets a provider reuse input it processed recently, such as standing instructions or reference documents, and charge a much lower rate for it. For agents that resend the same context on every call, it can cut input costs substantially.

When should I use a batch API?

Use it for work that doesn’t need an answer within seconds: overnight reconciliations, document backlogs and month-end reviews. Anthropic, OpenAI and Google list batch processing at 50% off standard prices.

Do list prices reflect what enterprises actually pay?

They are the starting point. Enterprise agreements, committed-use discounts, taxes, long-prompt surcharges and cache storage fees can all move the final number. Use list prices to plan, then compare them with your invoices each month.

Sources. List prices checked October 18, 2026: Anthropic, Claude pricing; OpenAI, API pricing; Google, Gemini API pricing.