What Does an AI Agent Cost to Run? Tokens, Tools and the Hidden Line Items
An AI agent's running cost is model tokens across every step, plus retries, tool and search fees, hosting, monitoring and human review. Budget per task, not per token.
6 min read

Contents
Running an AI agent costs the model tokens for every step it takes, plus retries, tool and search fees, hosting, monitoring and the time people spend reviewing its work. The model bill is usually the smallest part you can predict. Price it per task the agent completes, with written assumptions, and check it against real logs after the first month.
Picture an operations lead who has just had a support-triage agent approved. The pilot cost a few dollars a day, so the budget line says "API costs: small". Three months later finance asks why the invoice tripled, why there is a search bill nobody planned for, and why two people still spend their mornings checking the agent's replies. None of it was hidden on purpose. It just wasn't written down.
What goes into the cost of running an AI agent?
The running cost of an AI agent is everything you pay each month to have it complete tasks: model calls, retries, tools, infrastructure and human oversight. An AI agent is software that uses a model to decide its own next step and call tools until a task is done, which is why its cost grows with the number of steps, not the number of users. Our post on what an AI agent is covers that difference in detail.
The line items
Model calls: input and output tokens for every step, priced per million tokens.
Retries and failed runs: calls repeated after a timeout, a malformed answer or a failed check.
Tool and search fees: web search, grounding, enrichment or other paid APIs the agent calls.
Hosting: the server, queue or workflow platform the agent runs on.
Monitoring and evaluation: logs, traces and the test runs you repeat after every prompt or model change.
Human review: the minutes people spend approving, correcting or redoing the agent's work.
How are model calls priced in late 2026?
Model calls are priced per million tokens, with separate rates for input and output, and lower rates for cached input and batch jobs. A token is a small piece of text, roughly part of a word, and output tokens cost several times more than input tokens on every model below.
Standard rates per million tokens, as listed on each vendor's own pricing page on 8 November 2026:
Claude Sonnet 5.5 (Anthropic): Input: $2. Output: $10. Cache hits: $0.10.
Claude Haiku 5.5 (prompts up to 100,000 tokens): Input: $0.10. Output: $0.50. Cache hits: $0.01. Prompts over 100,000 tokens cost $0.50 input and $2.50 output.
Claude Opus 5.5: Input: $4. Output: $20. Cache hits: $0.20.
GPT-6.1 Sol (OpenAI): Input: $2. Output: $10. Cached input: $0.10. Prompts over 272K input tokens: $4 input, $15 output.
GPT-6.1 Terra: Input: $1. Output: $5. Cached input: $0.05.
Gemini 4 Argon (Google): Input: $2 through 30 November 2026, then $4 from 1 December. Output: $10, then $20. Cached input: $0.10, then $0.20.
Two details change budgets more than the headline rate. First, batch processing is half price on all three platforms, for work that can wait. Second, the Gemini 4 Argon price doubles on 1 December 2026, so a budget built on November test runs will be wrong in December. For choosing between these models by task rather than price, see our guide to choosing a model for business workflows.
Why does an agent use more tokens than a chatbot?
An agent uses more tokens because each step resends the instructions, the tool definitions and everything that has happened so far, and a single task can take several steps. A chatbot answers once; an agent reads, plans, calls a tool, reads the result and calls the model again.

Where the extra tokens come from
Repeated instructions: the system prompt and tool definitions are sent with every call, often thousands of tokens before the task itself.
Growing history: each step adds the previous tool results to the context, so later calls in a task are larger than earlier ones.
Tool results: a search or a database lookup can return far more text than the agent needs. OpenAI's pricing page notes that for web search, "search content tokens" retrieved from the index are billed at the model's own rates, on top of $10 per 1,000 calls.
Thinking tokens: reasoning models generate tokens you never see. Google's pricing page says of Gemini 4 Argon that "every token the model processes is counted and billed, including thinking tokens", and lists its output price as including them.
Retries: a call that fails validation is paid for twice.
What does a worked example look like?
Here is one support-triage agent, priced on the Claude Sonnet 5.5 rates above. Every number is an assumption for illustration, not a measurement; replace each with your own.
Volume: 2,000 tickets a month.
Steps: 4 model calls per ticket.
Input: 6,000 tokens per call, of which 4,000 are the same system prompt and tool definitions every time.
Output: 500 tokens per call.
Model cost without caching: 24,000 input tokens per ticket cost $0.048 and 2,000 output tokens cost $0.020, so $0.068 a ticket, or $136 a month.
With prompt caching: the 16,000 repeated tokens per ticket drop to the cache-hit rate, which brings the ticket to about $0.038, or roughly $75 a month, plus a small amount for cache writes.
Retries: if one call in ten is repeated, add about 10%, to roughly $83.
Human review: if staff check one ticket in ten for 3 minutes each, that is 10 hours a month. At an assumed internal cost of $40 an hour, it is $400.
On these assumptions the people reviewing the agent cost several times more than the model. The same workload on Claude Haiku 5.5 would cost about $7 a month in model calls before caching, and on Gemini 4 Argon it would go from $136 in November to $272 in December without caching. Hosting and monitoring depend on your stack and should be added as their own lines.
How do you bring the cost down without hurting quality?
You lower the cost by sending fewer and smaller calls, routing simple steps to cheaper models and reducing the review load, while an evaluation set confirms that quality holds. Cut one thing at a time and measure it.

The checklist
Cache the stable prefix. Keep the system prompt and tool definitions identical at the start of every call so they are billed at cache rates. Anthropic's page prices cache hits on Opus 5.5 and Sonnet 5.5 at 0.05 times the base input price.
Trim tool results. Return only the fields the agent needs, not whole records or pages.
Cap the steps. Set a maximum number of calls per task and send anything that hits it to a person.
Route by difficulty. Let a small model classify or extract, and call a larger one only for the steps that need it.
Batch what can wait. Overnight summaries and back-office jobs can use the half-price batch rates.
Watch the context length. Long prompts can move a request into a higher price tier, as with GPT-6.1 Sol over 272K tokens and Claude Haiku 5.5 over 100,000.
Reduce review with evidence. Lower the share of tasks a person checks only when your logs show the agent getting them right. Our checklist for testing an AI agent before production covers how to earn that.
How we approach it at Vizio AI
We cost an agent per completed task before we build it. With the client we write down the expected volume, the steps per task, the model for each step and the share of work a person will review, and we price that against the vendors' current pages, with the date next to every rate.
During the pilot we log tokens, retries, tool calls and review time for every task, then compare the real cost per task with the estimate and adjust the design: caching, smaller models for simple steps, tighter tool results, and lower review rates where the evaluation set supports it. Every agent gets a spending alert and a step cap from day one. This is part of our DX Studio work.
Sources
Anthropic, Claude Platform Docs, Pricing (checked 8 Nov 2026): https://platform.claude.com/docs/en/about-claude/pricing
OpenAI, API Docs, Pricing (checked 8 Nov 2026): https://developers.openai.com/api/docs/pricing
Google AI for Developers, Gemini Developer API pricing (checked 8 Nov 2026): https://ai.google.dev/gemini-api/docs/pricing
Vizio AI, DX Studio: https://www.vizio.ai/studios/dx-studio
FAQs.
Read More

Attribution That Survives Your Tool Stack: Keeping UTMs From Ad to Booking to CRM

Clean Your CRM Data Before You Add AI: Dedupe, Grain and Ownership

Semantic Layer First: Getting Your BI Ready for AI Agents

How F&B CMOs Use Generative AI In Business For Location Decisions

How Real Estate Firms Use Business Analytics For City Parks To Drive Value

Data Storytelling and Visualization

When Does a Business Actually Need Digital Transformation?

Predictive Analysis for Data Visualization

Data Analyst vs. Data Quality Analyst

Analytics for Product Recommendations


