Latest
Deal Heidi cuts headcount by 18% and co‑founders leave after $475 million raiseDeal What buyers look for when they acquire an AI companyGlobal How to tell whether an AI agent is paying for itselfGlobal Apple trims daily AI spending limits for engineering teamsGlobal Box adds Nvidia Nemotron to speed up agent content retrievalGlobal Claude adds personal finance tool that reads your bank dataGlobal Perplexity launches always‑on AI teammates for work
United States edition
Friday, 6 November 2026
Infercline.
Advertise
The Economics of AI

How much does an AI agent cost to run?

Agents are priced by the token, the outcome, the conversation or the seat. Here is how each model works, what vendors charged at the time of writing, and the costs that sit outside the bill.

The Infercline Desk · Friday, 6 November 2026 Sign in to save
A radio sitting on top of a table next to a yellow fire hydrant
Photo by Simone Dinoia on Unsplash

There is no single price for an AI agent. What you pay depends on how the vendor charges, how much work the agent does and how much of your team's time goes into setting it up and checking it. Before comparing quotes, it helps to know the four ways agents are usually sold: by the token, by the outcome, by the conversation or action, and by the seat.

Token pricing is the raw cost of the AI model underneath. A token is a small piece of text. Google's developer documentation says a token is about four characters, and that 100 tokens is roughly 60 to 80 English words. Model makers charge per million tokens, with one price for what you send in (the input) and a higher price for what the model writes back (the output). An agent that reads long documents or works through many steps uses far more tokens than a single question and answer.

Prices vary widely between models. At the time of writing (November 2026), Anthropic's pricing page listed Claude Haiku 5.5 at 0.10 US dollars per million input tokens and 0.50 dollars per million output tokens for shorter prompts, Claude Opus 5.5 at 4 dollars and 20 dollars, and its most expensive model, Claude Fable 5.1, at 10 dollars and 50 dollars. OpenAI's pricing page showed a similar spread, from gpt-6-luna at 0.10 and 0.50 dollars to gpt-6-astra at 10 and 50 dollars. Google listed Gemini 3.1 Flash-Lite at 0.25 dollars input and 1.50 dollars output, with a free tier, and Gemini 3.1 Pro Preview at 2 and 12 dollars for prompts up to 200,000 tokens.

There are discounts if you can wait or repeat yourself. Anthropic, OpenAI and Google all offer a batch option, where work runs later rather than instantly, at a 50 per cent discount according to their own pages. Anthropic and OpenAI also discount repeated material: Anthropic says reading from its prompt cache costs a tenth of the normal input price on most models, and OpenAI lists cached input at a tenth of the input price. For an agent that uses the same long instructions on every task, this can make a large difference.

Outcome pricing charges only when the agent finishes a job. Intercom's pricing page said its Fin agent cost 0.99 dollars per outcome at the time of writing, with no seats needed. Zendesk's pricing page said its plans include a small allowance of automated resolutions, meaning requests the AI resolves without passing to a person, with extra resolutions at 1.50 dollars each if committed in advance or 2 dollars pay as you go. Outcome pricing is easy to budget, but read how the vendor defines an outcome or a resolution.

Conversation and action pricing sits in between. Salesforce's Agentforce pricing page listed 2 dollars per conversation for agents that talk to outside customers, and Flex Credits at 500 dollars per 100,000 credits, with a standard action using 20 credits. That works out at about 10 cents per action. Microsoft's Copilot Studio page listed capacity packs of 25,000 Copilot Credits for 200 dollars per pack per month, with a pay-as-you-go option billed through Azure. Both pages say prices may change.

Seat pricing is the familiar per-user licence. Salesforce, for example, listed an Agentforce user licence at 5 dollars per user per month, which still needs Flex Credits for the work itself, and higher-priced add-ons. Seat prices look predictable, but the usage charges on top are where bills grow.

The model bill is rarely the biggest cost. Someone has to describe the job, connect the agent to your systems, test it on real cases and decide what it may and may not do. Someone has to check its work, especially at the start, and handle the cases it passes back. If the agent touches customer data, someone has to check that you meet data protection rules in each country you operate in. These hours are real and belong in the budget. Gartner, in a June 2025 forecast that over 40 per cent of agentic AI projects would be cancelled by the end of 2027, named escalating costs as one of the main reasons.

Currency is a further factor outside the United States. Most of these prices are set in US dollars. Businesses paying in pounds, rand, naira, Singapore or Australian dollars, dirham or riyal carry exchange rate risk, and local taxes may apply on top.

Why it matters to your business

The pricing model shapes your risk. Token pricing rewards careful design and punishes an agent that loops or reads more than it needs. Outcome pricing moves some risk to the vendor but can get expensive at high volume. Seat pricing is easy to approve but hides usage charges. Knowing which you are buying lets you compare a quote against the cost of the person-hours it is meant to replace, and spot a bill that is growing faster than the work.

What to do on Monday

  1. Pick one job you are thinking of handing to an agent and count how many times it happens in a normal month. You need that volume to compare any quote.
  2. Ask each vendor which pricing model applies, what counts as a billable unit, and what happens to the bill if volume doubles.
  3. Ask whether batch processing or caching discounts apply to your use, and whether the quote already includes them.
  4. Add the hours for set-up, testing and checking to the budget, using your own team's rates.
  5. Set a monthly spending cap or alert with the vendor before going live, and review the first three bills line by line.

More on this topic: AI agent costs

Sources