Cost to build an AI agent in 2026: prices and estimates

Hephaestus Technology8 min read

The cost to build an AI agent splits into a one-off build cost, which is mostly people's time, and a running cost, which is mostly model usage billed per million tokens. On the price lists published today (8 November 2026), model usage ranges from $0.10 to $10 per million input tokens and from $0.50 to $50 per million output tokens, depending on the model tier. Our hypothetical worked example below shows the same agent costing roughly $55 or roughly $1,000 a month in model fees, depending on which tier you choose.

Key takeaways

  • Budget for two things separately: the build (people's time for design, integration, evaluation and hardening) and the running cost (tokens, tools, hosting).
  • Published per-million-token prices on 8 November 2026 span about 100x between the smallest and largest models at Anthropic, OpenAI and Google.
  • Prompt caching and batch processing are the biggest published discounts: batch is 50% off at all three providers, and cached input is billed at a small fraction of the normal input price.
  • Only about one in five Irish and EU enterprises used AI in 2025, and a lack of expertise was the most common reason for not adopting it.
  • Treat any cost estimate as a model of your own traffic: measure tokens per task in a prototype before you commit to a budget.

What do AI models cost per token in November 2026?

Model usage is billed per million tokens (MTok), with output tokens costing about five times as much as input tokens on most current models. The table shows a cross-section of the price lists published by Anthropic, OpenAI and Google on 8 November 2026, in US dollars, at standard (non-batch) rates.

Provider and model Input per MTok Cached input per MTok Output per MTok
Anthropic Claude Haiku 5.5 (prompts up to 100,000 tokens) $0.10 $0.01 $0.50
Anthropic Claude Sonnet 5.5 $2.00 $0.10 $10.00
Anthropic Claude Opus 5.5 $4.00 $0.20 $20.00
OpenAI gpt-6-luna $0.10 $0.01 $0.50
OpenAI gpt-6.1-terra $1.00 $0.05 $5.00
OpenAI gpt-6.1-sol $2.00 $0.10 $10.00
OpenAI gpt-6-astra $10.00 $1.00 $50.00
Google gemini-3.1-flash-lite $0.25 $0.025 $1.50
Google gemini-3.8-flash $0.75 $0.075 $3.75
Google gemini-4-argon $2.00 $0.10 $10.00

Three caveats matter when you read this table. First, tokens are not identical across providers: Anthropic notes that its newer models use a tokenizer that produces approximately 30% more tokens for the same text, so compare cost per task, not just cost per token. Second, prices move: Google's page lists scheduled increases for gemini-4-argon on 1 December 2026 and for gemini-3.8-flash on 1 January 2027, and Google adds an hourly storage charge for some cached content (Google pricing). Third, Google's cached figure is its context caching rate, which works differently from the other two.

What discounts are published?

Batch processing and prompt caching are the two discounts every major provider now lists. Anthropic's Batch API gives a 50 percent discount on both input and output tokens for asynchronous work. Google describes its Batch API as a 50% cost reduction, and OpenAI's batch prices are half its standard prices on the same pricing page.

For caching, Anthropic charges cache reads at 0.1x the base input price on most models (0.05x on Opus 5.5 and Sonnet 5.5), with cache writes at 1.25x for the five-minute cache. OpenAI says caching is enabled by default and cached input is discounted up to 95%. Because an agent resends its instructions and tool definitions on every step, caching usually has more effect on an agent's bill than on a simple chatbot's. Our guide to keeping LLM costs under control covers how to structure prompts so the cache actually hits.

What else appears on the bill besides tokens?

Tool calls, storage and data residency are billed separately from tokens. Both Anthropic and OpenAI list web search at $10 per 1,000 searches, plus the tokens the search results add. OpenAI lists hosted vector storage for file search at $0.10 per GB per day, with 1 GB free, and $2.50 per 1,000 file search calls.

For EU customers, data residency can add a premium. OpenAI states that its regional processing endpoints are charged a 10% uplift for models released on or after March 5, 2026, and its Europe region covers the EEA and Switzerland. Add your own hosting, logging, monitoring and any database or vector store you run yourself.

How many businesses are actually building with AI?

About one in five enterprises in Ireland and the EU used AI in 2025. The CSO reports that 20.2% of enterprises in Ireland used AI technologies in 2025, up from 8.1% in 2023, with large enterprises at 57.7% and small ones (10 to 49 people) at 17.2%. Eurostat puts the EU figure at 19.95% in 2025.

The barriers are relevant to budgeting. Among EU enterprises that considered AI but did not use it, 70.89% cited a lack of relevant expertise, 52.52% a lack of clarity about legal consequences and 48.83% concerns about data protection and privacy. In practice, expertise and compliance work are often a larger part of the budget than the model fees.

A worked estimate: what might a support agent cost to run?

This is a hypothetical example built only from the published prices above, not a client project or a market average. Assume a customer support agent that handles 20,000 conversations a month. Each conversation takes six model calls (read the question, look up the order, check the policy, draft, check, reply). Each call sends 8,000 input tokens, of which 6,000 are the same instructions and tool definitions every time and can be cached, and returns 400 output tokens.

  1. Mid-tier model ($2 input, $0.10 cached, $10 output per MTok): per call, 2,000 fresh tokens cost $0.004, 6,000 cached tokens cost $0.0006 and 400 output tokens cost $0.004. That is about $0.0086 per call, $0.052 per conversation and about $1,030 a month.
  2. Same model without caching: 8,000 input tokens cost $0.016 per call, so the total is $0.02 per call and about $2,400 a month.
  3. Small model ($0.10 input, $0.01 cached, $0.50 output per MTok): about $0.00046 per call and about $55 a month, if the small model is accurate enough for the task.

The example ignores cache write charges, retries, failed runs and tool fees, so treat it as a lower bound. The gap between lines 1 and 3 is why model choice and evaluation belong in the budget conversation from day one.

What does the build itself cost?

The build cost is mainly people's time, and it varies far more than token prices. A useful way to estimate is by phase, then multiply by the day rate you are actually quoted. The phases below are a hypothetical planning template with illustrative effort ranges, not industry data:

Phase What it covers Illustrative effort
Discovery and design Use case, success criteria, data access, risks 5 to 10 days
Prototype Prompts, one or two tools, measuring tokens per task 10 to 15 days
Integration Connecting real systems, permissions, error handling 20 to 30 days
Evaluation and guardrails Test sets, human review points, abuse limits 10 to 15 days
Launch hardening Monitoring, cost alerts, documentation 5 to 10 days

Multiply the total by quoted rates, which vary widely by country; our overview of software development rates in 2026 is a starting point. Anthropic's engineering guidance is worth reading before you commit: it notes that agentic systems often trade latency and cost for better task performance and recommends the simplest solution that works. If a fixed workflow or a chatbot will do, the build is smaller, as our comparison of an AI agent vs a chatbot vs workflow automation explains.

How should a founder budget for an AI agent?

Start with a short prototype that measures real tokens per task, then budget from those measurements. Set a monthly spend cap with your provider, pick the smallest model that passes your evaluation, and recheck prices quarterly because the price lists above change several times a year. Keep 15 to 25% of the build budget for evaluation and hardening, because that is the work a demo skips and production traffic exposes.

Frequently asked questions

Is it cheaper to build an AI agent on a small model?

Usually much cheaper to run: on today's price lists the smallest models cost around one twentieth of a mid-tier model per token. Whether it is cheaper overall depends on accuracy. If a small model needs more retries or more human corrections, the saving shrinks, so test both tiers on the same evaluation set.

Are AI model prices quoted in euro?

The Anthropic, OpenAI and Google API price lists cited here are in US dollars. Your invoice currency and any VAT treatment depend on your account and billing country, so check your billing settings and convert at the rate your bank or card applies.

How often do LLM prices change?

Often. On 8 November 2026 Google's pricing page already listed scheduled increases for two models in December 2026 and January 2027, and Anthropic's page records a planned increase for one model that was cancelled. Recheck the official pages before signing off any budget.

What is the biggest hidden cost of an AI agent?

Usually the work around the model rather than the model itself: integration with your systems, evaluation, monitoring and handling the cases the agent gets wrong. Lack of expertise was the most common reason EU enterprises gave for not adopting AI in 2025, which reflects how much of the effort is skilled people's time.

Does EU data residency make an AI agent more expensive?

It can. OpenAI, for example, lists a 10% uplift on regional processing endpoints for models released on or after 5 March 2026. Other providers and cloud platforms price residency differently, so compare the total for your region, not just the headline token price.

If you want a second opinion on an AI agent budget before you commit, book a free 30-minute call.

Cover photo by https://kaboompics.com/ on Pexels.

Hephaestus Technology

We help founders and businesses validate, build and launch products across software, AI and hardware, from offices in Ho Chi Minh City, Dublin and Delaware. See how we work.

Tell us what
you're building.

A free 30-minute call. You'll get an honest view of the idea, the risks we see and what we'd do first, whether or not we end up working together.