How Much Does It Cost to Build an AI Agent?
An AI agent has two costs: a one-time build and a monthly running bill. Published data puts most custom AI builds between $10,000 and $250,000, with Clutch reporting an average AI project of $120,594.55 as of September 21, 2026. The running cost is mostly model usage, and you can estimate it from the providers' per-token prices before any code is written.
No public dataset isolates AI agents from other AI projects. The figures below are for AI development in general, and each one is labeled with what it measures. Prices and quotes were read on October 3, 2026. Where a page carries its own date, that date is given.
What does an AI agent cost to build, according to published data?
Published figures range from under $10,000 to more than $250,000. The sources disagree because one counts what clients paid, one counts what vendors say they charge, and none separates agents from simpler AI work.
| Source | What it measures | Figure | Date |
|---|---|---|---|
| Clutch AI Pricing Guide | Client reviews of AI development companies | Average project $120,594.55; most common band $10,000 to $49,999; typical timeline 10 months | Updated September 21, 2026 |
| GoodFirms cost survey | What more than 100 software companies say they charge for AI builds | MVP $50,000 to $125,000 (71.4% of companies); medium $125,000 to $250,000 (23.4%); large above $250,000 (4.7%) | Updated September 28, 2026 |
| Lemon.io salary report | Average hourly rates paid to AI engineers on its platform, from a dataset of 2,500+ contracts | Mid-level $43, senior $60.3, strong senior $79.2 per hour | April 25, 2026 |
| US Bureau of Labor Statistics | Median US employee pay for software developers, QA analysts and testers | $64.44 per hour, $134,040 per year (2025) | Last modified August 27, 2026 |
Three notes on reading the table.
Clutch's average is pulled up by large projects. Its most common band is $10,000 to $49,999, and the same page gives an average monthly cost of $11,553.45.
GoodFirms collected its answers "between September and October 2025" from vendors, so these are asking prices. Its text gives 71.9% for the MVP band where its table gives 71.4%.
The hourly figures are pay to engineers. A company's billing rate adds management, overhead and margin on top.
What drives the build cost of an AI agent?
Five things drive it: scope, integrations, evaluation, guardrails and compliance. The choice of model has little effect on build cost and a large effect on running cost.
Scope. The first decision is whether the task needs an agent at all. Anthropic's guidance of December 19, 2024 recommends "finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all." A fixed workflow with one or two model calls costs less to build and test than an agent that plans its own steps.
Integrations. Each system the agent reads from or writes to is a separate piece of engineering: authentication, permissions, error handling and tests. OWASP's 2025 guidance on Excessive Agency says to limit the tools an agent can call "to only the minimum necessary", which also keeps the build smaller.
Evaluation. An agent needs a test set of real tasks with expected outcomes, run on every change to prompts, tools or models. Anthropic's guidance says "The autonomous nature of agents means higher costs, and the potential for compounding errors. We recommend extensive testing in sandboxed environments, along with the appropriate guardrails."
Guardrails. These are step limits, permission checks, input and output filters, and human approval for risky actions. OWASP recommends "human-in-the-loop control to require a human to approve high-impact actions before they are taken." Each approval step needs an interface and a queue.
Compliance. Regulated data adds audit logging, access reviews and data-location rules, and some of it shows up in prices. Anthropic's pricing page applies "a 1.1x multiplier on all token pricing categories" for US-only inference. OpenAI's pricing page says regional processing endpoints "are charged a 10% uplift for models released on or after March 5, 2026."
Adding AI to a software project raises its price in vendor surveys too. In the GoodFirms survey, most participants said "AI Integration alone can result in a 0-30% increase in cost."
What does an AI agent cost to run each month?
The monthly bill has four lines: model usage, infrastructure, monitoring and maintenance. Model usage is the line that grows with traffic.
Model usage. Agents use more tokens than chatbots because every step resends the context. Anthropic reported on June 13, 2025 that "agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats."
Some tools carry their own fees. Anthropic's pricing page charges "$10 per 1,000 searches" for its web search tool. Google's Gemini pricing page, last updated October 1, 2026, lists search grounding on Gemini 3 models at "$14 per 1,000 requests" after 5,000 free requests a month.
Infrastructure. The agent's own code, database and job queue run on ordinary cloud hosting. Hosted agent runtimes charge by time: Anthropic lists its Managed Agents session runtime at "$0.08 per session-hour", with tokens billed separately.
Monitoring. Tracing each step is how you find failures and waste. Langfuse's pricing page lists a free Hobby plan with 50k units a month, a Core plan at $29 a month and a Pro plan at $199 a month, with additional usage at $8 per 100k units. LangSmith's pricing page lists its Plus plan at $39 per seat per month with 10k base traces included.
Maintenance. The GoodFirms survey reports that companies "spend an average of 10–20% annually on post-launch maintenance, bug fixes, and updates." Agents add one item to that list: models are retired. Anthropic's pricing table already marks Claude Opus 4.1 and Claude Sonnet 4 as retired on its own API, and each migration means running the evaluation set again.
How do you estimate monthly model spend from per-token prices?
Multiply the number of model calls by the tokens in each call and by the price per million tokens, with input and output counted separately.
- Count model calls per task. In an agent, each tool call is followed by another model call.
- Measure input tokens per call: system prompt, tool definitions, conversation history and tool results.
- Measure output tokens per call.
- Multiply by tasks per month and apply the published prices.
Worked example. A support agent handles 10,000 conversations a month. Each conversation takes 6 model calls, and each call averages 6,000 input tokens and 400 output tokens. That is 60,000 calls, 360 million input tokens and 24 million output tokens a month. The volumes are assumptions for illustration. The prices are those listed on October 3, 2026.
| Model | Input / output price per 1M tokens | Input cost | Output cost | Monthly total |
|---|---|---|---|---|
OpenAI gpt-6-luna |
$0.10 / $0.50 | $36 | $12 | $48 |
| Google Gemini 3.8 Flash | $0.75 / $3.75 | $270 | $90 | $360 |
| Anthropic Claude Haiku 4.5 | $1 / $5 | $360 | $120 | $480 |
| Anthropic Claude Sonnet 5.5 | $2 / $10 | $720 | $240 | $960 |
OpenAI gpt-6.1-sol |
$2 / $10 | $720 | $240 | $960 |
| Anthropic Claude Opus 5.5 | $4 / $20 | $1,440 | $480 | $1,920 |
OpenAI gpt-6-astra |
$10 / $50 | $3,600 | $1,200 | $4,800 |
The same workload costs $48 or $4,800 a month depending on the model. Google lists the Gemini 3.8 Flash price "through December 31, 2026", after which it doubles.
Prompt caching lowers the input line. If 3,000 of the 6,000 input tokens per call are a stable prefix, such as the system prompt and tool definitions, Claude Sonnet 5.5 bills those 180 million tokens at its cache-hit rate of $0.20 per million. That is $36 where the full rate would be $360, and the monthly total falls from $960 to $636 before cache-write charges.
Anthropic's own pricing page shows how much the agent loop matters. Its example of single-call support tickets at about 3,700 tokens each comes to "~$37.00 per 10,000 tickets" on Claude Haiku 4.5. The six-call agent above costs $480 on the same model.
Replace the assumed volumes with numbers measured from a prototype. Token counts also differ by model: Anthropic notes that the tokenizer in Claude 4.7 and later "produces approximately 30% more tokens for the same text."
Is it cheaper to buy an agent than to build one?
For standard customer-service work at modest volume, buying is usually cheaper at the start, because there is no build cost. Published per-outcome prices also show the level at which a custom build begins to pay.
Intercom's Fin is priced at "$0.99 per outcome", and the page adds that "You're not charged when a conversation is simply passed to your team without an outcome." Salesforce Agentforce lists $2 per conversation, or $500 per 100,000 Flex Credits when paying per action.
At 10,000 billable outcomes a month, that is $9,900 with Fin and $20,000 with Agentforce conversations. A custom agent has a build cost and the running lines above in place of that monthly fee. Building tends to make sense when the workflow is specific to your business, when volume is high, or when data cannot pass through a vendor's product.
Why do AI agent projects go over budget?
They go over budget when the scope grows, when token use is not measured, and when the agent is allowed to loop. Gartner predicted on June 25, 2025 that "Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls".
The same release quotes Gartner analyst Anushree Verma: "Many use cases positioned as agentic today don't require agentic implementations."
Three causes are within the buyer's control:
- Scope creep. In the GoodFirms survey, "65.6% of the survey respondents reported that scope creep typically increased project cost by 10-25%."
- Unbounded usage. OWASP's 2025 list names "Denial of Wallet", in which a high volume of requests exploits "the cost-per-use model of cloud-based AI services". Rate limits, step limits and spend caps prevent it.
- Multi-agent designs chosen too early. They use about 15 times the tokens of chat, by Anthropic's figures.
Key takeaways
- Published AI build costs run from under $10,000 to more than $250,000, and no public source isolates agents.
- Scope, integrations, evaluation, guardrails and compliance drive the build cost.
- Model usage drives the running cost, and the same workload can cost 100 times as much on the largest model as on the smallest.
- Estimate model spend as calls × tokens × price, then measure it on a prototype.
- Per-outcome products at $0.99 to $2 set a reference price for standard support work.
Frequently asked questions
How much does an AI agent cost per month to run?
It depends on volume and model. In the worked example above, 10,000 conversations a month cost between $48 and $4,800 in model usage. Monitoring tools start free and list paid plans from $29 to $199 a month, and hosting is extra.
How long does it take to build an AI agent?
Clutch reports a typical timeline of 10 months for AI development projects in its reviews as of September 21, 2026. That covers all AI projects. A narrow agent with few integrations takes less, and no public source gives a timeline for agents alone.
Does a multi-agent system cost more than a single agent?
Yes. Anthropic reported in June 2025 that multi-agent systems use about 15 times the tokens of chat, against about 4 times for single agents. It advised using them where "the value of the task is high enough to pay for the increased performance."
What is the cheapest way to build an AI agent?
Start with a fixed workflow and add agent behavior only where the task needs it. Use the smallest model that passes your tests, cache the stable part of the prompt, and limit the agent to the tools it needs.
Do I pay for agent runs that fail?
Yes. Tokens are billed whether or not the task succeeds. Step limits and retry caps keep a failing run from repeating.
Easital Technologies Ltd. designs, builds and runs custom AI agents, from a scoped use case to a monitored production system. One example is Calldone, a voice agent platform built by Easital for a client in the United States. See AI agent development, LLM cost optimization, hire AI developers and our work.
Sources
All sources were opened and checked on October 3, 2026.
- Clutch, "AI Pricing Guide 2026", Anna Peck, updated September 21, 2026. https://clutch.co/developers/artificial-intelligence/pricing
- GoodFirms, "Custom Software Development Cost Survey 2026", updated September 28, 2026. https://www.goodfirms.co/resources/custom-software-development-cost-survey
- Lemon.io, "Software Developer Salary Report 2026", April 25, 2026. https://lemon.io/salary-report/
- US Bureau of Labor Statistics, Occupational Outlook Handbook, "Software Developers, Quality Assurance Analysts, and Testers", last modified August 27, 2026. https://www.bls.gov/ooh/computer-and-information-technology/software-developers.htm
- Anthropic, "Building effective agents", December 19, 2024. https://www.anthropic.com/engineering/building-effective-agents
- Anthropic, "How we built our multi-agent research system", June 13, 2025. https://www.anthropic.com/engineering/multi-agent-research-system
- OWASP Gen AI Security Project, "LLM06:2025 Excessive Agency", 2025. https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
- OWASP Gen AI Security Project, "LLM10:2025 Unbounded Consumption", 2025. https://genai.owasp.org/llmrisk/llm102025-unbounded-consumption/
- Anthropic, "Pricing", Claude API documentation, read October 3, 2026. https://platform.claude.com/docs/en/about-claude/pricing
- OpenAI, "Pricing", OpenAI API documentation, read October 3, 2026. https://developers.openai.com/api/docs/pricing
- Google, "Gemini Developer API pricing", last updated October 1, 2026. https://ai.google.dev/gemini-api/docs/pricing
- Langfuse, "Pricing", read October 3, 2026. https://langfuse.com/pricing
- LangChain, "LangSmith pricing", read October 3, 2026. https://www.langchain.com/pricing
- Intercom, "Fin pricing", read October 3, 2026. https://fin.ai/pricing
- Salesforce, "Agentforce pricing", read October 3, 2026. https://www.salesforce.com/agentforce/pricing/
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", press release, June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027