AI agent development cost: what it takes to build and run one

Three things people call an AI agent

“AI agent” covers everything from one model call inside a workflow to software that plans, uses your systems and acts on its own. Quotes differ by an order of magnitude mostly because they describe different things. Before comparing numbers, decide which of these you mean:

Anthropic’s guidance on building agents draws the same line: “Workflows are systems where LLMs and tools are orchestrated through predefined code paths,” while agents are systems where “LLMs dynamically direct their own processes and tool usage.” It also warns that “the autonomous nature of agents means higher costs, and the potential for compounding errors.” The further down the table, the more it costs to build, to run and to keep reliable.

What it takes to build each

Build effort depends less on the model than on everything around it: access to your systems, the evaluation set, guardrails, the review step for people and the interface. These planning ranges, in person-weeks for a small senior team, are assumptions for this guide rather than quotes:

To turn person-weeks into money, multiply by your team’s weekly rate; our MVP cost guide explains how to calibrate one with public benchmarks. As a sense of scale, the same 6 person-weeks is very different money from a small team in a lower-cost region than from a large consultancy onshore.

In our AI agent and automation work we usually recommend starting with the first type, or with a knowledge assistant, and adding autonomy only once an evaluation set shows the simpler version works.

What it costs to run every month

A production agent has six running costs. Only the first is on the model provider’s price page:

1. Model usage, charged per million tokens of input and output.
2. Embeddings and search for knowledge assistants: turning documents into vectors and storing them.
3. Hosting and orchestration: the service or workflow tool that runs the steps.
4. Tracing and evaluation tooling to see what the agent did and score changes.
5. Human review of the cases the agent hands over. This is a cost, and the reason the agent is safe.
6. Upkeep: re-running evaluations when models change, updating prompts and tools when your systems do.

Model prices, as listed today

Model prices are per million tokens (a token is roughly three-quarters of an English word), with output priced higher than input. List prices on the providers’ own pages:

Three things change the effective price. Batch processing, for work that can wait (OpenAI says each batch completes within 24 hours), is 50% off at all three providers. Prompt caching cuts the price of input that repeats, such as long instructions sent with every request, by up to 90–95%, with a small premium to write the cache. And tokens are not the same unit everywhere: Anthropic notes that its newer models’ tokenizer “produces approximately 30% more tokens for the same text”, so compare models on your own prompts, not on the price list alone. Very long prompts can also cost more: OpenAI’s prices rise above 272,000 input tokens, and Claude Haiku 5.5’s above 100,000.

Worked example: the monthly model bill

The agent (hypothetical): a support triage workflow that reads each incoming ticket, retrieves similar past tickets and help-centre articles, then classifies, routes and drafts a first reply for a person to approve.

Assumptions, per ticket:

• Instructions, the ticket and retrieved context: about 4,000 input tokens.
• Classification, routing and a draft reply: about 600 output tokens.
• One model call per ticket, and 10,000 tickets a month.

Input tokens a month

10,000 × 4,000 = 40 million

Output tokens a month

10,000 × 600 = 6 million

Small model (Claude Haiku 5.5 or GPT-6 Luna)

40 × $0.10 + 6 × $0.50 = $7

Mid-tier model (Claude Sonnet 5.5, GPT-6.1 Sol or Gemini 4 Argon)

40 × $2 + 6 × $10 = $140

Flagship model (GPT-6 Astra)

40 × $10 + 6 × $50 = $700

Model bill a month

$7 to $700, depending on the model

Caching narrows it further. If 2,500 of the 4,000 input tokens are the same instructions and examples on every call, reading them from Claude Sonnet 5.5’s cache costs $0.10 per million instead of $2, which brings the mid-tier bill from $140 to roughly $100 a month once cache writes are included.

Two lessons follow. First, model usage for a workflow like this is often modest next to the people and engineering around it. Second, the choice of model moves it by an order of magnitude, which is why the evaluation set matters: it tells you whether the cheaper model is good enough, instead of guessing.

The rest of the monthly bill

The largest line is usually human review, and it is set by your confidence threshold: the more cases the agent hands to a person, the safer and the more expensive it is. Tuning that threshold against the evaluation set is one of the most valuable things you can do after launch.

The costs quotes leave out

• Building the evaluation set. Collecting and labelling real examples takes time from the people who know the process.
• Access and permissions. Connecting to systems safely, with the agent seeing only what the user may see, is often the slowest part.
• Data preparation. Documents in scanned PDFs, duplicates and outdated versions make retrieval worse until cleaned up.
• Model changes. Providers retire and replace models; each change needs a re-run of the evaluation set and sometimes prompt changes.
• Security and legal review of what data goes to which provider, under which terms.
• Change management. Teams need to know when to trust the agent, how to correct it and who owns it.

How to keep the cost predictable

• Start with a workflow. Use a fixed sequence with one model step before an open-ended agent; add autonomy only where evaluation shows it pays.
• Route by difficulty. Send simple cases to a small, cheap model and only hard ones to a large model.
• Cache and batch. Reuse the fixed part of prompts, and run work that can wait through batch processing at a discount.
• Cap every run. Limit steps, tool calls and tokens per task, so a confused agent stops instead of looping.
• Change models on evidence. Adopt a new model when it scores better on your evaluation set at an acceptable cost, not because it is new.

If you are at the point of choosing a first use case, our guide to running an AI proof of concept that reaches production covers the four weeks before any of these costs become real. For workflows built in automation tools, see n8n vs Zapier.

Sources

Every source below was read on 26 October 2026. Prices and product details change; check the current page before relying on a figure.

1. Anthropic: Building effective agents
2. OpenAI: A practical guide to building agents
3. OpenAI: API pricing
4. OpenAI: Batch API
5. OpenAI: Prompt caching
6. Anthropic: Pricing (Claude Platform)
7. Google: Gemini Developer API pricing
8. pgvector: pgvector: open-source vector similarity search for Postgres
9. Pinecone: Pricing
10. Langfuse: Pricing
11. Langfuse: Self-host pricing
12. LangChain: LangSmith pricing