--- title: "How Much Does It Cost to Build an AI Agent in 2026?" url: "https://dupple.com/learn/how-much-does-it-cost-to-build-an-ai-agent" description: "What an AI agent costs to build and run in 2026: per-task model costs from OpenAI and Google list prices, no-code builders from $29.99/mo, and the hidden costs." published: "2026-11-08" updated: "2026-11-08" llms_txt: "https://dupple.com/llms.txt" --- # How Much Does It Cost to Build an AI Agent in 2026? Token Math, Not Agency Quotes Trusted by 790,000+ Techpresso subscribers · Editorial policy · How we make money Written by Louis Corneloup Founder at Dupple, covering AI tools and strategies for 790K+ readers. How our guides are researched . November 8, 2026 · Updated November 2026 10 min read **Key facts4 routes compared: Custom build on an API, Lindy, Microsoft Copilot Studio, Microsoft 365…** - **Updated:** November 8, 2026 - **Prices as of:** November 2026 - 4 routes compared: Custom build on an API, Lindy, Microsoft Copilot Studio, Microsoft 365 Copilot add-on - Custom build on an API (best for: Teams with an engineer and a clear, high-volume task): $0 platform fee; tokens from $3 per 1,000 triage tasks (gpt-6-luna) - Lindy (best for: Small teams automating inbox, meetings and follow-ups without code): $29.99 per user a month for 3,000 credits; free start with $50 of credits for 7 days - Microsoft Copilot Studio (best for: Organizations on Microsoft 365 that need agents inside it or on a website): $200 a month per pack of 25,000 Copilot Credits, or a pay-as-you-go meter - Microsoft 365 Copilot add-on (best for: Companies already licensing Microsoft 365): $30 per user a month, paid yearly Short answer: running an AI agent costs mostly tokens. **An inbox-triage task of about 20,000 input and 2,000 output tokens costs $0.003 on OpenAI's gpt-6-luna and $0.06 on gpt-6.1-sol**, so 1,000 tasks a month cost $3 to $60 in model spend (our arithmetic on list prices verified October 2026). No-code builders start at $29.99 a month on [Lindy](https://try.lindy.ai/xgu55xh248g4). The big quotes you see elsewhere are developer time. ## What 1,000 agent tasks a month cost *What 1,000 agent tasks a month cost: How Much Does It Cost to Build an AI Agent in 2026? Token Math, Not Agency Quotes. Last updated November 2026.* | Agent type | Tokens per task (our assumption) | Cheap model: gpt-6-luna | Mid model: gpt-6.1-sol | Flagship: gpt-6-astra | | --- | --- | --- | --- | --- | | (a) Inbox triage | 20,000 in, 2,000 out | $3 | $60 | $300 | | (b) Research, 5 web lookups | 60,000 in, 6,000 out | $9 | $180 | $900 | | (c) Coding or ops agent, cached context | 150,000 in (120,000 cached), 15,000 out | $11.70 | $222 | $1,170 | Costs per 1,000 tasks at OpenAI's standard API prices, read on [OpenAI's pricing page](https://developers.openai.com/api/docs/pricing) on 28 October 2026. Token counts are our assumptions for a typical task of each type. Row (b) counts tokens only, not search fees; row (c) prices the 120,000 repeated tokens at the cached-input rate and excludes cache-write charges. ## Why the agency quotes look so different Search this question and you will find agency articles quoting large build budgets. Those numbers price people: discovery workshops, integration work, testing and a maintenance retainer. They are real costs if you hire an agency, but they say almost nothing about what the agent itself costs to run. The running cost is something you can compute from a vendor's pricing page. A model provider charges per million tokens, separately for input (what you send: instructions, documents, conversation history, tool results) and output (what the model writes back, including any reasoning tokens). **For most small business agents, the monthly model bill is tens of dollars, not thousands.** The expensive part is the work of making the agent reliable. That split matters for budgeting. Model spend scales with volume and model choice, which you control. Labour is mostly a one-time build plus upkeep, and it depends on how many systems the agent touches. If you need a refresher on what counts as a token, read our explainer on [what a token is in AI](https://dupple.com/learn/what-is-a-token-in-ai). **Dupple data:** Toolradar, our software directory, tracks 1,022 AI agent tools as of 28 October 2026. Among the 319 that publish a price, the median cheapest paid plan is $29 a month ([Toolradar AI agents ranking](https://toolradar.com/best/ai-agents)). That is the going rate for a packaged agent before you write any code. ## How much does an AI agent cost per task? Per task, the formula is simple: input tokens times the input price, plus output tokens times the output price. Here is the arithmetic once, for the inbox-triage agent on gpt-6.1-sol at $2 per million input tokens and $10 per million output tokens: - Input: 20,000 tokens x $2 / 1,000,000 = $0.04 - Output: 2,000 tokens x $10 / 1,000,000 = $0.02 - Total: **$0.06 per task**, the figure behind the mid-model column in the first table The same task on gpt-6-luna ($0.10 input, $0.50 output) costs $0.003, twenty times less. On gpt-6-astra ($10 input, $50 output) it costs $0.30. The model you choose moves the bill by two orders of magnitude, which is why picking the cheapest model that does the job well is the single biggest cost lever. Google's Gemini API gives a similar spread. Two of its prices are introductory and rise soon, so we show both: *How much does an AI agent cost per task?: How Much Does It Cost to Build an AI Agent in 2026? Token Math, Not Agency Quotes. Last updated November 2026.* | Agent type | Gemini 3.5 Flash-Lite | Gemini 3.8 Flash, through Dec 31, 2026 | Gemini 3.8 Flash, from Jan 1, 2027 | Gemini 4 Argon, through Nov 30, 2026 | Gemini 4 Argon, from Dec 1, 2026 | | --- | --- | --- | --- | --- | --- | | (a) Inbox triage | $11 | $22.50 | $45 | $60 | $120 | | (b) Research | $33 | $67.50 | $135 | $180 | $360 | | (c) Coding or ops, cached | $50.10 | $87.75 | $175.50 | $222 | $444 | Per 1,000 tasks at paid-tier standard prices on [Google's Gemini API pricing page](https://ai.google.dev/gemini-api/docs/pricing), read 28 October 2026, same token assumptions and caching rule as above. Argon costs $2 per million input tokens and $10 per million output through November 30, then $4 and $20 from December 1. Gemini 3.8 Flash costs $0.75 and $3.75 through December 31, then $1.50 and $7.50 from January 1, 2027. **If you build on Argon this month, budget for double the price in December.** Anthropic's Claude models are a third option; check the current per-token prices on Anthropic's site, as the same formula applies. ## How much does an AI agent cost per month to run? Multiply the per-task cost by your volume, then add the infrastructure around the model. For our three example agents at 1,000 tasks a month on a mid-tier model, the model bill is $60 for triage, $180 for research and $222 for the coding agent with caching. Volume changes the picture quickly. A support agent that handles 10,000 conversations a month at the triage profile costs $600 on gpt-6.1-sol and $30 on gpt-6-luna, by the same arithmetic. That gap is why many teams route easy requests to a small model and send only the hard ones to a larger one; an [LLM gateway](https://dupple.com/learn/best-llm-gateways) can do that routing for you. Beyond tokens, a custom agent usually needs a place to run (a server or serverless functions), a database or vector store for its memory, logging, and any paid tools it calls, such as search or a CRM API. Those costs depend on your stack and are often small next to the model bill at low volume, but they do not disappear. ## Is it cheaper to build or buy an AI agent? At low volume, buying is usually cheaper once you count your own time. At high volume, or when the agent is core to your product, building wins on unit cost. *Is it cheaper to build or buy an AI agent?: How Much Does It Cost to Build an AI Agent in 2026? Token Math, Not Agency Quotes. Last updated November 2026.* | Route | Entry price | What you pay for | Best for | | --- | --- | --- | --- | | Custom build on an API | $0 platform fee; tokens from $3 per 1,000 triage tasks (gpt-6-luna) | Tokens, hosting, developer time | Teams with an engineer and a clear, high-volume task | | Lindy | $29.99 per user a month for 3,000 credits; free start with $50 of credits for 7 days | Pooled credits; top-ups at $10 per 1,000 credits | Small teams automating inbox, meetings and follow-ups without code | | Microsoft Copilot Studio | $200 a month per pack of 25,000 Copilot Credits, or a pay-as-you-go meter | Credits per agent action, tenant-wide | Organizations on Microsoft 365 that need agents inside it or on a website | | Microsoft 365 Copilot add-on | $30 per user a month, paid yearly | Copilot Chat plus building agents inside Microsoft 365 | Companies already licensing Microsoft 365 | Sources: [Lindy pricing](https://www.lindy.ai/pricing) and [Copilot Studio pricing](https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio), read 28 October 2026. Some visitors to Lindy's page see a different lineup, starting at 5,000 credits; the table shows the lineup in its structured data. A no-code builder bundles the model, hosting, integrations and an interface, which is what you would otherwise pay an engineer to assemble. The trade-off is control: you pay per credit at the builder's rate, and a credit is not a token, so you cannot easily compare a Lindy or Copilot Studio bill with the token tables above. Lindy's own guide puts everyday asks at 2 to 250 credits and deep work at 250 to 1,000. For plan-by-plan detail, see our [Lindy AI pricing](https://dupple.com/learn/lindy-ai-pricing) and [Microsoft Copilot Studio pricing](https://dupple.com/learn/microsoft-copilot-studio-pricing) breakdowns. To compare builders on features, start with our ranking of [AI agent platforms](https://dupple.com/learn/best-ai-agent-platforms); if you plan to write code, our list of [AI agent frameworks](https://dupple.com/learn/best-ai-agent-frameworks) and our step-by-step guide to [building an AI agent](https://dupple.com/learn/how-to-build-an-ai-agent) are the next reads. ## Which model is cheapest for agents? On list price, OpenAI's gpt-6-luna is the cheapest of the models in our tables, at $0.10 per million input tokens and $0.50 per million output. Gemini 3.5 Flash-Lite is next at $0.30 and $2.50. Cheapest is not the same as best value, though: a model that fails a task and retries, or needs a human to fix its output, costs more than its token price suggests. A practical approach is to start on a mid-tier model to prove the agent works, then test whether a cheaper model passes the same checks. If you want to avoid API bills entirely for private or offline work, open models can also run on your own hardware; our guide to [running an LLM locally](https://dupple.com/learn/how-to-run-an-llm-locally) covers the trade-offs. For free options with a hosted interface, see our list of [free AI agents](https://dupple.com/learn/best-free-ai-agents). ## What are the hidden costs of AI agents? These are the charges that do not show in a headline per-token price. Each figure comes from the vendor pricing pages cited above, read 28 October 2026. - **Introductory prices that end.** Gemini 4 Argon doubles on December 1, 2026, to $4 input and $20 output per million tokens. Gemini 3.8 Flash doubles on January 1, 2027. An agent costed on today's Argon price needs twice the budget a month later. - **Search grounding.** Google includes 5,000 grounded search requests a month, shared across Gemini 3 and newer models, then charges $14 per 1,000. Our research agent makes 5 lookups per task, so 1,000 tasks use the free allowance exactly; at 2,000 tasks, the extra 5,000 requests cost $70 a month (our arithmetic). - **Cache writes.** Caching cuts the cost of repeated context sharply, but writing to the cache costs more than normal input on OpenAI: gpt-6.1-sol charges $2.50 per million tokens for a 30-minute cache write against $2 for standard input. Caching pays off only when the same context is reused many times. - **Context caching storage on Gemini.** Gemini 3.8 Flash adds a storage charge of $0.50 per million tokens per hour through December 31, 2026, and $1.00 after that. - **Long prompts and data residency.** OpenAI prices its flagship models separately for prompts above 272K input tokens, and charges a 10% uplift on regional processing endpoints for models released on or after March 5, 2026. - **Thinking tokens.** Google bills reasoning ("thinking") tokens as output, and Argon also counts internal tool instructions when grounding is used. Reasoning-heavy agents can produce far more output tokens than their visible answers suggest. - **Free tiers that train on your data.** Google's free tier marks content as used to improve its products; the paid tier does not. For business data, budget for the paid tier. - **Per-user minimums on builders.** Lindy bills every active user in the workspace, and Copilot Studio sells capacity in packs of 25,000 credits, so a light user still pays the full entry price. Caching is also where the biggest saving hides. Without it, our coding agent's 150,000 input tokens cost $450 per 1,000 tasks on gpt-6.1-sol; with 120,000 of them cached, the cost falls to $222 (our arithmetic). For more ways to cut the bill, Batch processing on OpenAI halves the standard price for jobs that can wait. ## How much does a no-code AI agent builder cost? Entry prices in October 2026 run from $29.99 per user a month on Lindy to $200 a month for a 25,000-credit pack on Copilot Studio, with the Microsoft 365 Copilot add-on at $30 per user a month for organizations already on Microsoft 365. Toolradar's median of $29 a month for AI agent tools that publish a price suggests most packaged agents sit at the low end. The number to watch is not the entry price but the credit burn. A builder that charges per action can cost more than a custom build once an agent runs thousands of times a month. Before committing, run one real workflow through the free allowance and note how many credits it used, then multiply by your expected monthly volume. ## How we calculated these costs We read OpenAI's API pricing page, Google's Gemini API pricing page, Lindy's pricing page and Microsoft's Copilot Studio pricing page on 28 October 2026, and used standard (non-batch) list prices. Token counts per task are our assumptions for three common agent types: 20,000 input and 2,000 output tokens for triage, 60,000 and 6,000 for research, 150,000 (120,000 cached) and 15,000 for a coding or ops agent. All per-task and monthly totals are our arithmetic. Labour, hosting and agency fees are not priced here because no vendor page prices them. We plan to recheck every price every 60 days, sooner when Google's introductory prices end. ## FAQ **How much does it cost to build an AI agent in 2026?** The model cost of running one is often small: 1,000 inbox-triage tasks cost $3 to $60 a month on OpenAI's gpt-6-luna to gpt-6.1-sol at October 2026 list prices. Building it costs developer time, or a no-code builder subscription from $29.99 per user a month on Lindy. **How much does an AI agent cost per month?** It depends on volume and model. At 1,000 tasks a month on gpt-6.1-sol, our three example agents cost $60, $180 and $222 in tokens. Add hosting, storage and any paid tools the agent calls. **What is the cheapest model for an AI agent?** Among the models we priced, OpenAI's gpt-6-luna is cheapest at $0.10 per million input tokens and $0.50 per million output tokens, followed by Gemini 3.5 Flash-Lite. Test that a cheap model handles your task before you commit, since retries and fixes add cost. **Is it cheaper to build or buy an AI agent?** Buying a no-code builder is usually cheaper at low volume because it bundles hosting and integrations. Building on an API is cheaper per task at high volume, but you pay for engineering time. **Will AI agent costs go up?** Some will soon. Gemini 4 Argon's price doubles on December 1, 2026, and Gemini 3.8 Flash's doubles on January 1, 2027, according to Google's pricing page. OpenAI's page shows no scheduled change for the gpt-6 models we priced. **What does Copilot Studio cost?** Microsoft sells Copilot Studio as a tenant license with packs of 25,000 Copilot Credits at $200 a month, or through a pay-as-you-go meter. Each agent action consumes a varying number of credits. **Do AI agents need web search, and what does it cost?** Research agents usually do. Google includes 5,000 grounded searches a month across Gemini 3 and newer models, then charges $14 per 1,000 requests. ## Bottom line For most teams, the model is the cheap part of an AI agent: tens of dollars a month for a thousand tasks on a mid-tier model, and a few dollars on a small one. The real costs are building it well and keeping it reliable, plus a few traps worth budgeting for, chiefly Google's introductory prices ending and per-user minimums on no-code builders. Price your agent per task first; the build-or-buy answer usually follows from that number. **Cite this:** Dupple, "How Much Does It Cost to Build an AI Agent in 2026? Token Math, Not Agency Quotes", November 2026. **Want to put AI agents to work without guesswork?** [Start the Dupple X trial](https://checkout.dupple.com/checkout/dupple-x-yearly-trial/) for practical AI courses and tool guides. Some offers on this page may be paid placements or contain affiliate links. Related Articles Blog Post ### The Best AI Math Solver in 2026: 8 Tools Compared I compared 8 AI math solvers in 2026: Wolfram Alpha, Mathos AI, Photomath, Symbolab and more. Real pricing, accuracy, and which one to pick. Blog Post ### 8 Best AI for College Students in 2026 (Study, Not Cheat) The 8 best AI for college students in 2026, plus which student discounts exist. ChatGPT, Claude, Perplexity, NotebookLM with real free-tier limits. Blog Post ### 7 Best AI News Sources for Educators in 2026 (Teachers' Reading List) The 7 best AI news sources for educators in 2026: EdSurge, Education Week, Tech & Learning and more, re-checked in October 2026 for format and price. Article ### Is GitHub Copilot Free in 2026? The Free Tier, Student Access and Paid Plans GitHub Copilot Free costs $0 with capped inline completions. Verified students get Copilot free, and some teachers and maintainers may be eligible for Pro free. Article ### 10 Best AI Money Apps for Students in 2026 Enrolled students can get YNAB free for a year, then $109/year. Cleo Plus is $5.99/mo. Prices for these student money apps were checked in September 2026. Blog Post ### Best AI Education Tools in 2026: Tutors, Teacher Platforms, and Study Apps Compared The best AI education tools in 2026, compared: Khanmigo, MagicSchool, NotebookLM, ChatGPT Study Mode, Grammarly, Quizlet and more, with real pricing.