Most published prices for building a custom AI agent in 2026 fall between $10,000 and $75,000, and simple single-workflow agents start at a few thousand dollars. The best public data point is Gigabit's October 2026 price index: across 30 AI firms that publish a price for a production agent, the median entry price is $20,000, and the middle half start between $10,000 and $32,000. Running the agent is a separate monthly bill, and that bill got much smaller this year because OpenAI and Anthropic cut model prices hard between June and October.
This guide covers both numbers. We look at what firms actually publish, how cost changes by agent type, what it costs per month to run an agent at current model prices (with the math shown), which costs teams forget, and when a freelancer makes more sense than an agency. Every figure has a source and a date.
How much does an AI agent cost? The short answer
| What you're buying | Typical published price | Source |
|---|---|---|
| Paid discovery or assessment | median $4,500 (middle half $2,130 to $7,204) | Gigabit index, 20 firms |
| Pilot, proof of concept, or MVP sprint | median $12,000 (middle half $7,750 to $18,500) | Gigabit index, 22 firms |
| Production AI agent or feature | median $20,000 (middle half $10,000 to $32,000) | Gigabit index, 30 firms |
| Monthly run or support retainer | median $2,500 a month (middle half $812 to $11,250) | Gigabit index, 13 firms |
These are entry prices, the lowest figure each firm publishes. Gigabit read the websites of 122 AI service firms on October 22, 2026, and half of them publish no price at all. None of the 13 large consultancies in the sample (Accenture, Deloitte, IBM and others) publishes one. A real project with several integrations usually lands above the entry price.
Cost by agent type
The 2026 cost guides all tie the price to how much the agent is allowed to do and how many systems it touches. The model it runs on matters much less. Here is how three of them break it down.
| Agent type | Dreambit | Beevr | Keyhole (single June 2026 guide) |
|---|---|---|---|
| Simple or single-workflow agent / chat over your docs | $3,000 to $15,000 | $5,000 to $25,000 | $15,000 to $40,000 |
| Production agent with RAG and real integrations | $15,000 to $75,000 | $15,000 to $75,000 | $30,000 to $80,000 |
| Autonomous multi-tool agent | $50,000 to $150,000 (listed as a learning, context-aware agent) | $50,000 to $150,000 | $60,000 to $150,000 |
| Multi-agent or enterprise system | $75,000 to $250,000+ | $150,000 to $500,000+ | $120,000 to $400,000+ |
The ranges overlap but the shape is the same. A useful rule from Dreambit: integration and orchestration alone often run $20,000 to $50,000 before the agent does anything clever.
Why estimates spread so far: hours times rate
9io.ai built its estimate bottom-up and got 330 to 980 engineering hours for a first production release. At $95 to $250 an hour that's roughly $32,000 to $245,000. The same hours at $250 cost 2.6 times what they cost at $95, so the rate moves the total as much as the scope does. For comparison, the median published hourly rate in Gigabit's index is $30, but that number comes mostly from offshore staffing firms. Specialist AI consultancies mostly don't publish a rate.
What it costs to run an AI agent each month
Run cost has four parts: model tokens, infrastructure (hosting, vector store, logs, traces), evals and monitoring, and people (tuning prompts, reviewing outputs, fixing what breaks). Published estimates:
- Dreambit: $3,200 to $13,000 a month for a production agent serving real users, with token spend alone ranging from a few hundred dollars to $20,000+.
- Keyhole: from $200 to $800 a month for a simple agent up to $4,000 to $20,000+ for a multi-agent system, plus 15% to 25% of the build price a year for maintenance.
- 9io.ai: $1,700 to $14,300 a month for an agentic workflow with 1,000 users at moderate use, upkeep included.
The token bill, with the math
Model prices changed a lot in 2026. Max Woolf tracked the releases from June to October and found that OpenAI and Anthropic cut prices on comparable models by up to 80%. Vercel's AI gateway data in the same post shows the average cost per token at 43.3% of the April 2026 level. Current list prices per million tokens (read October 27, 2026):
| Model | Input | Output | Cached input (read) |
|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
| GPT-6.1 Sol | $2 | $10 | $0.10 |
| Claude Haiku 5.5 (prompts up to 100k tokens) | $0.10 | $0.50 | $0.01 |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
Sources: Anthropic pricing, OpenAI pricing.
Now an example. Take an agent that handles one support ticket or one document per task. Each task makes 10 model calls. Each call sends 15,000 input tokens, of which 13,000 are the cached conversation and 2,000 are new, and returns 500 output tokens. Add one cache write of 13,000 tokens per task. That's an assumption for illustration, not a benchmark.
| Model | Per task | 1,000 tasks a month | 10,000 tasks a month |
|---|---|---|---|
| Claude Opus 5.5 | about $0.27 | about $271 | about $2,710 |
| Claude Sonnet 5.5 or GPT-6.1 Sol | about $0.14 | about $136 | about $1,355 |
| Claude Haiku 5.5 or GPT-6 Luna | about $0.007 | about $7 | about $74 |
At moderate volume, the model bill is now often smaller than the hosting, monitoring and human review around it. Cached input also matters a lot: in agent loops most of the input is the same conversation sent again, and Woolf estimates cache reads at 96% to 99% of token spend for coding agents. Prompt caching and routing easy steps to a small model are the two biggest levers on the token bill. (We wrote more about routing in how to choose a model for each task.)
Token prices can move again in either direction. Plan the budget with your own traffic numbers and recheck the price pages before launch.
Costs teams forget
- Evals and testing. An agent can give a wrong answer and also take a wrong action. 9io.ai notes that approvals, guardrails and evals carry the most hours in agents that write to other systems. See how to test AI agents.
- Integrations you haven't seen yet. The other system's API is the biggest unknown before discovery, which is why integration estimates have the widest ranges.
- Human review. Someone has to look at edge cases, at least in the first months.
- Overruns. Beevr cites industry surveys that put agentic AI project overruns at 35% to 50% above the first estimate. The usual reason: the budget covered the demo, and the production work showed up later.
- Paid discovery. Gigabit's median is $4,500. A good discovery step is cheaper than building the wrong agent.
Agency or freelancer for an AI agent?
| Agency | Freelancer | |
|---|---|---|
| Typical entry price | Gigabit median $20,000 for a production agent; enterprise projects from $150,000 | Usually lower, priced on one person's hours |
| Who you talk to | Account manager, then the team | The person who writes the code |
| Capacity | Several people in parallel | One person or a small group |
| Risk | Lower if someone gets sick | Higher: the project stops if the freelancer does |
| Good fit | Multi-agent systems, regulated industries, many integrations at once | One workflow, a pilot, an agent MVP, fast iteration with your team |
For agents, ask whether the builder will work with your team, on your data and tools, until the agent is used every day. That matters more than whether they call themselves an agency or a freelancer. Large AI labs send forward deployed engineers to customers for exactly this reason. What that role looks like is covered in what a forward deployed engineer is.
How to keep the cost down
- Start with one workflow that has a clear owner and a measurable result.
- Pay for discovery first. Map the systems the agent must touch before anyone quotes a build.
- Let the agent suggest before it acts. Write access adds permissions, approval paths and tests. Start read-only or with human approval.
- Route by difficulty. A small model for classification and extraction, a strong model only where errors cost money.
- Cache the stable parts of the prompt and keep context tight.
- Budget the run cost with your own volume using the token table above, then add hosting, monitoring and review time.
FAQ
How much does it cost to build an AI agent in 2026?
Across 30 AI firms that publish a price, the median entry price for a production AI agent is $20,000, and the middle half start between $10,000 and $32,000 (Gigabit, October 2026). Simple single-workflow agents start around $3,000 to $15,000, and multi-agent enterprise systems run $75,000 to $250,000 and more.
How much does it cost to run an AI agent per month?
Published estimates range from about $200 a month for a simple agent to $3,200 to $13,000 for a production agent with real users. At October 2026 prices, the model bill for 1,000 tasks a month can be under $10 on a small model or a few hundred dollars on a top model, so hosting, monitoring and human review are often the bigger part.
Did AI agents get cheaper in 2026?
Running them did. OpenAI and Anthropic cut prices on comparable models by up to 80% between June and October 2026, and Vercel's gateway data shows the average cost per token at 43.3% of the April level. Building them did not get much cheaper, because integrations, evals and review still take engineering hours.
What is the hourly rate for an AI agent developer?
The median published rate in Gigabit's index is $30 an hour, mostly from offshore staffing firms. 9io.ai uses $95 to $250 an hour as the 2026 market range for its estimates. Most specialist AI consultancies don't publish a rate.
Should I hire an agency or a freelancer to build an AI agent?
A freelancer fits one workflow, a pilot or an agent MVP, and you talk directly to the person building it. An agency fits multi-agent systems, regulated industries and projects that need several engineers at once.
Why do AI agent cost estimates vary so much?
Because the agent's job varies. The number of systems it touches, how much it may do without approval, the eval effort and the hourly rate all change the total. The same hours cost 2.6 times more at $250 an hour than at $95.
Want to price your agent?
I build AI agents, AI integrations, MVPs, and mobile and web apps. I work as a freelancer, closer to a forward deployed engineer than to an agency: a workshop with your team first, then a written scope and two-week sprints with something working at the end of each. My MVPs start at $5,000, with no upper limit, because the price depends on scope. I bill hourly on a B2B invoice. More on how I build agents is on the AI agent development page. Tell me what the agent should do and I'll reply within 24 hours.
Sources
Read October 27, 2026, unless noted.
- Gigabit, AI services price index, October 2026 (122 firms, data read October 22, 2026, CC BY 4.0)
- Dreambit, AI agent development: cost and when it's worth it
- Beevr, How much does an AI agent cost to build in 2026
- Keyhole Software, Agentic AI development cost in 2026 (45+ sources, June to August 2026)
- 9io.ai, How much does it cost to build an AI agent or app in 2026
- Max Woolf, OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why? (October 25, 2026)
- Anthropic, Claude pricing; OpenAI, API pricing
Want to build an AI agent, an MVP, or an app?
I design and build AI products from the first version to production: MVPs, React Native mobile apps, web apps, and AI agents. Tell me what you want to build and I will tell you where to start.