A production AI agent costs from $8,000 to about $150,000 to build, at the flat fees on our pricing page. Running it costs tokens plus operations. Tokens can be under $10 a month for 10,000 simple tasks on a small model, or thousands of dollars for long agent loops on a larger one. Managed operations are $3,000 to $20,000 a month with us, or the time of your own engineers. Model prices below were checked on the vendors' pages on 30 October 2026.
AI agent costs at a glance
| Cost | What it covers | Typical amount |
|---|---|---|
| Build | One agent, in one workflow, with evals, monitoring and guardrails | From $8,000 flat, up to about $150,000 (Gigabit) |
| Diagnostic, optional | Two-week AI Transformation Sprint ending in a deployed pilot | $25,000 (Gigabit), credits toward a build |
| Model tokens | Each model call the agent makes | See the examples below, from under $10 to thousands a month |
| Operations | Monitoring, eval regressions, prompt versions, vendor changes | $3,000–$20,000/mo (Gigabit), or your engineers' time |
| Other services | Hosting, vector databases, the tools it connects to | Depends on your stack, so get these from your vendors |
What it costs to build an AI agent
Our price is a flat fee agreed before work starts. A standard agent starts at $8,000, and the agent catalog lists what each one does. Four things push a build toward the top of the range:
- Integrations. An agent that reads one system and writes to another is the simplest case. One that orchestrates five tools with approval steps is far more work.
- Data readiness. Clean, queryable data saves weeks. Knowledge stuck in inboxes and PDFs needs a pipeline first. In Informatica's 2025 survey, data and technology tied as the top obstacles to production, each named by 43% of data leaders.
- Risk. A customer-facing agent needs guardrails, evals and escalation paths that an internal tool may not.
- Operating bar. An agent that runs unattended overnight needs more than one that works in a demo.
For real figures instead of ranges, see what ten client projects cost: the build price, the monthly cost after launch, the timeline and the payback, taken from our published case studies.
What it costs to run: model prices
Prices are in US dollars per million tokens at standard rates. Vendors change these often, so check the source before you budget.
| Model | Input | Output | Source |
|---|---|---|---|
| Claude Haiku 5.5 (up to 100K tokens) | $0.10 | $0.50 | Anthropic |
| GPT-6 Luna | $0.10 | $0.50 | OpenAI |
| Gemini 2.5 Flash | $0.30 | $2.50 | |
| Gemini 3.8 Flash, through 31 Dec 2026 | $0.75 | $3.75 | |
| GPT-6.1 Terra | $1 | $5 | OpenAI |
| Claude Sonnet 5.5 | $2 | $10 | Anthropic |
| GPT-6.1 Sol | $2 | $10 | OpenAI |
| Claude Opus 5.5 | $4 | $20 | Anthropic |
Google has announced that Gemini 3.8 Flash goes to $1.50 and $7.50 on 1 January 2027, so prices can rise as well as fall.
Worked examples: tokens for 10,000 tasks a month
The token counts are example inputs we chose, and your numbers will differ. Measure them on real traffic. The totals are our arithmetic on the Anthropic prices above, with no caching or batch discount.
| Example task | Tokens per task | Haiku 5.5 | Sonnet 5.5 | Opus 5.5 |
|---|---|---|---|---|
| One model call, such as a simple support answer | 3,000 in, 700 out | $6.50 | $130 | $260 |
| An agent loop of eight calls, each with 20,000 in and 1,000 out | 160,000 in, 8,000 out | $200 | $4,000 | $8,000 |
The second row is why loops matter. The same 10,000 tasks cost about 30 times more when each one runs eight calls with a large context. An agent that keeps retrying or carries a long history gets expensive quickly.
How to cut the token bill
- Prompt caching. On Sonnet 5.5 a cache read is $0.10 per million tokens, against $2 for fresh input. In the loop above, if 75% of input tokens were cache reads, the Sonnet cost would fall from $4,000 to about $1,720 by our arithmetic, before the higher price of cache writes.
- Batch processing. Anthropic lists 50% off for batch work, and OpenAI's batch prices are half its standard ones. This suits work that can wait.
- Match the model to the step. Use a small model for classification and extraction, and a larger one only where it changes the result.
- Cap the loop. Set a limit on calls per task and alert when it is hit.
Prices at a given capability level have fallen fast. Epoch AI measured a 40x fall per year for GPT-4-level performance on one benchmark, and 9x to 900x across the six it studied. It also says it is unclear whether the fastest declines will continue, so budget on today's prices.
What it costs to run: operations
After launch, models change, vendors ship breaking changes and edge cases build up. Someone has to watch for that. Our Managed AI Operations cover monitoring, eval regressions, prompt versions, drift and vendor changes, in three tiers:
| Tier | Price | Covers |
|---|---|---|
| Light | $3,000/mo | 1 system |
| Standard | $8,000/mo | 3 systems |
| Heavy | $20,000/mo | 5+ systems or production agent platforms |
You can do this in-house if an engineer has the time and the tooling. Either way, budget for it. An agent nobody watches can give bad outputs for weeks before anyone notices.
Where the money gets wasted
Failed builds cost more than tokens. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. Forbes, reporting on an MIT study, says 95% of GenAI pilots fail. In RSM's 2025 survey, 91% of mid-market respondents used generative AI, but only one in four of those users had it fully integrated into core operations.
A demo that collapses under real load and has to be rebuilt means paying twice. We wrote up the engineering side in why your AI pilot never reached production.
Is it worth it? A worked example
Take an example workflow that uses 20 hours a week at $75 an hour. That is $78,000 a year. If an agent handles 50% of it, the saving is $39,000 a year. A standard agent at $8,000 pays back in about 2.5 months on those figures, before running costs. Add our Light operations tier at $3,000 a month, or $36,000 a year, and most of the saving is gone.
So at this size, either run it with your own team or pick a bigger workflow. At 70% automation of a 40-hour workflow, the saving is $109,200 a year and the same tier fits easily. Those are our sums on example inputs, and the AI ROI calculator lets you run your own.
How to keep the cost down
- Scope to one workflow. Cost and risk both grow with surface area.
- Measure tokens on real traffic. Do it before you pick a model or sign an operations plan.
- Gate releases on evals. They catch regressions before customers do. Here is how we build them.
- Start with a diagnostic if you are unsure. A two-week Sprint gives you the real number before you commit, and credits toward the build.
- Consider building in-house. If AI is the core of your product and you can hire senior engineers, it can win long term. Our build vs buy vs embed guide goes through it.
