The cost to build an AI agent in 2026 comes in three parts: the build, the model bill, and the people who keep it working. For a production agent with real tools and evaluation, I'd budget about $44,000 to $106,000 for the build with a senior nearshore team, a few hundred to a couple of thousand dollars a month in tokens, and more for upkeep than for tokens. That last part surprises almost everyone.
Key Takeaways
- Build cost is engineer hours. A focused single-job agent is about 800 hours; a multi-tool agent with proper evals is closer to 1,900.
- Token costs are predictable if you count calls per task. Model choice can swing the bill 40x.
- Prompt caching cuts the input side of the bill hard, because agents resend the same system prompt and tool definitions on every call.
- Ongoing engineering (evals, prompt fixes, tool changes) usually costs more per month than the tokens.
The Cost to Build an AI Agent, Part One: Engineering Hours
The rate below is an assumption: a blended $55 per hour for a senior nearshore team. It's my number for the arithmetic, not a quote. A person is 40 hours a week.
| Agent scope | Team | Weeks | Hours | At $55/h |
|---|---|---|---|---|
| Single job, 2-3 tools (e.g., support triage) | 1 AI engineer + 1 backend | 10 | 800 | $44,000 |
| Multi-step, 6+ tools, eval suite, human handoff | 2 AI engineers + 1 backend | 16 | 1,920 | $105,600 |
Most of those hours aren't prompt writing. They go to the tools the agent calls (your APIs, permissions, rate limits), the evaluation set that tells you whether a change made things better or worse, and the fallback paths for when the model gets it wrong.
For a third-party anchor, Clutch's AI development pricing page (updated September 2026) says reviewed AI projects range "from $10,000 to $49,999," with an average of $120,595 and a typical timeline of 10 months. That page covers AI development in general, not agents, and it's built from reviews clients chose to submit. Use it as a sanity check. One detail I found interesting: Clutch lists both US and Mexico AI firms in the same $50 to $99 per hour bucket, so AI talent isn't discounted as heavily as general development.
Part Two: The Model Bill
Count it per task. Here's a support agent that handles 20,000 tickets a month. Each ticket takes 4 model calls, averaging 3,000 input tokens and 500 output tokens per call. Per ticket that's 12,000 input and 2,000 output. Per month: 240 million input tokens and 40 million output tokens.
Prices per million tokens, from Anthropic's pricing page and OpenAI's pricing page, as of October 2026:
| Model | Input / Output per MTok | Monthly bill (240M in, 40M out) |
|---|---|---|
| Claude Opus 5.5 | $4 / $20 | $960 + $800 = $1,760 |
| Claude Sonnet 5.5 | $2 / $10 | $480 + $400 = $880 |
| Claude Haiku 4.5 | $1 / $5 | $240 + $200 = $440 |
| gpt-6-luna | $0.10 / $0.50 | $24 + $20 = $44 |
That's a 40x spread between the top and bottom rows. The cheap model won't handle the hard tickets, but it can often handle the routing step, which is why most production agents mix models.
Caching changes the math. Agents resend the same system prompt and tool definitions on every call. If 70% of input tokens are cached, on Sonnet 5.5 that's 168M tokens at the $0.20 cache-read price ($33.60) plus 72M at $2 ($144), plus the same $400 of output. Roughly $578 a month instead of $880, ignoring the smaller cache-write charges.
If your agent searches documents, embeddings are almost free. Embedding 10 million tokens with OpenAI's text-embedding-3-small at $0.02 per million costs $0.20. Storage can be free too: pgvector runs inside the Postgres you already have, and Pinecone's Builder tier is $20 a month flat. Our guide on how to hire RAG developers goes further on the retrieval side.
Part Three: Upkeep
Models change. Your tools change. Customers find edge cases nobody wrote an eval for. Somebody has to read failed transcripts every week, add them to the test set, and fix the prompt or the tool.
Plan on a quarter of an engineer for a single-job agent. At 160 hours a month that's 40 hours, or $2,200 a month at my assumed rate. In the example above, that's almost four times the cached Sonnet bill. Teams that skip this line watch their agent get quietly worse for three months and then decide "AI doesn't work for us."
A Concrete Version
A 40-person B2B SaaS company wants an agent to triage support tickets: read the ticket, look up the account, check recent incidents, tag it, draft a reply for simple cases, and route the rest.
Build: the single-job row, 800 hours, $44,000. Ten weeks.
Year-one running cost, 20,000 tickets a month with caching on Sonnet 5.5: about $578 a month, or $6,936 a year. Upkeep at 40 hours a month: $26,400 a year. Vector search on pgvector in their existing database: $0 extra.
Year one total: $44,000 + $6,936 + $26,400 = $77,336. If the agent fully resolves even a third of tickets, that's one or two support hires they don't make, which is the comparison that matters to their CFO.
The Honest Counterpoint
A lot of "agents" should be scripts. If the steps are always the same (read form, call API, write row), plain code is cheaper, faster, and never hallucinates. Use a model where the input is messy language and the decision really varies.
Multi-step reliability is also still hard. An agent that's right 95% of the time on each of five steps is right about 77% of the time end to end (0.95 to the fifth power). That's why the bigger build in the table spends so many hours on evals and human handoff. If you can't tolerate a wrong answer reaching a customer, budget for a human in the loop and price that too.
And sometimes you should buy. If an off-the-shelf support tool already has an agent that fits 80% of your flow, the build math above rarely wins on year one.
Frequently Asked Questions
What is the cost to build an AI agent for a small business?
A narrow agent for a small business, like booking or FAQ answers, can sit well under the first row of the table. Our post on the cost to build an AI chatbot for a small business covers that end of the market.
What is the monthly cost to run an AI agent?
Count calls per task, tokens per call, and tasks per month, then multiply by the official price sheet. For the support example above, it ranges from $44 to $1,760 a month depending on the model, before caching.
Should I hire an AI engineer or an agency to build an AI agent?
If the agent touches your core product and will need weekly upkeep, you want an engineer on your team, even a contracted one. See how to hire an AI agent developer for what to test in the interview.
The Bottom Line
Price the build in hours, price the tokens per task, and don't forget the person who keeps it honest. If you want senior AI engineers from LATAM to build or run it, look at our AI engineers or request a shortlist. You'll see vetted profiles within 72 hours. If you're still scoping, how much it costs to build an MVP uses the same formula for the product around the agent.
Roberto Espinoza is CEO of Ruzora, which helps US startups hire pre-vetted senior LATAM engineers, with a vetted shortlist in 72 hours. See available engineers.
