AI agent development cost, worked out line by line

How much does AI agent development cost in 2026?

There are two bills, and vague guides blur them. The build is a one-off piece of engineering: discovery, tools, an eval set, guardrails and a supervised pilot. The run is monthly: model tokens, hosting, monitoring, channel fees and the minutes your staff spend approving what the agent proposes. The build is mostly people. The run is mostly usage, and it is the part that surprises teams in month three.

Build cost: engineer-weeks by agent tier

These are the ranges we plan with for one agent doing one task, from the first workshop to the end of a supervised pilot. They are planning figures, not a quote, and they assume your systems have working APIs and someone on your side answers questions within a day.

Turn effort into money with your own numbers:

build cost ≈ engineer-weeks × blended weekly rate + your experts’ hours + one-off fees + contingency

• Blended rate: the weighted average of everyone on the work, from senior engineer to part-time reviewer. Use your in-house cost, or back it out of the proposals you hold.
• Your experts’ hours: someone who knows the task must label eval cases, approve tool designs and judge pilot output. Plan a few hours a week for the whole build.
• One-off fees: phone numbers, a security or legal review, a data protection impact assessment where the law or your client asks for one.
• Contingency: we add about a fifth for the integration nobody has looked at closely yet.

Weighing an in-house build instead? The same engineer-weeks apply; our guide to hiring AI developers in India covers that route.

What pushes an agent into a higher tier

• Writing instead of reading. Every write needs validation, an undo story and usually an approval screen.
• Each extra system of record, especially one with no sandbox or a partner approval process.
• Regulated data. India’s DPDP Rules, RBI outsourcing expectations or US HIPAA add reviews, retention rules and logging.
• Languages. Every language needs its own eval cases, and Indian-language voice needs testing with real speakers.
• Voice. Latency targets and call transfer are engineering work that chat agents never need.

Running cost: tokens × steps × price

An agent calls the model several times per task: once to pick the first tool, again after each tool result, and once to write the reply. Each call re-sends the instructions, the tool definitions and everything that has happened so far. So estimate per step and add the steps up:

model cost per task = Σ steps (uncached input × input price + cache writes × write price + cache reads × read price + output × output price)

Reasoning tokens count as output on both Anthropic and OpenAI models, so a model that thinks at length on every step costs more than its visible replies suggest.

List prices we used (read 1 November 2026)

US dollars per million tokens, standard tier, global processing. Cache writes cost more than plain input: 1.25 times for Anthropic’s 5-minute cache and OpenAI’s 30-minute cache, 2 times for Anthropic’s 1-hour cache and 2.5 times for OpenAI’s 12-hour cache. Gemini caching on these two models is implicit: you get the lower rate when a request hits the cache, from a 4,096-token minimum. Both Gemini models have price rises already published, so a budget written in November is wrong by January.

Worked example: an order-support agent (illustrative)

Not client work. The assumptions are the point, so change them to match your task:

• 3,000 tasks a month, six model calls per task.
• A fixed 3,000-token prefix: instructions plus tool definitions.
• Each call adds about 1,000 tokens of new context: the last output plus a tool result.
• 400 output tokens per call, reasoning included.

That is 39,000 input and 2,400 output tokens per task, or 117 million input and 7.2 million output tokens a month. With caching switched on, 9,000 of each task’s input tokens are cache writes and 30,000 are cache reads.

Gemini rows show the uncached price. On Gemini 4 Argon, an agent loop reuses the previous request from cache only when the new request adds at least 2,050 tokens, and our example adds about 1,000 per step, so its saving here would be small. Agents with bulkier tool results would see more.

Two lessons. Caching roughly halves the bill for this shape of agent, for free. Model choice moves it nearly fortyfold, and whether a small model is good enough is a question for your eval set, not the price list. Many agents route the easy steps to a small model and keep a larger one for the hard ones.

What does an AI voice agent cost per minute?

Voice adds meters that run while the caller talks: telephony, speech and the model behind the conversation, plus a platform fee if you use a hosted voice platform.

Illustrative: 2,000 inbound calls a month in the US, three minutes each, so 6,000 minutes. Rates from Twilio’s US voice pricing and OpenAI’s pricing page, read on 1 November 2026.

Conversation Relay turns the caller’s speech into text for your model and speaks the reply. GPT-Live 1 holds the spoken conversation itself and hands tasks to a backend model, which OpenAI bills separately. Either way the model is the smallest line: speech and telephony dominate, so for voice a per-minute rate matters more than a token price. Hosted voice platforms fold these lines into one per-minute price; compare it with your own sum.

India is different. Twilio’s number price list has no Indian numbers, and calls to Indian mobiles from its international numbers cost $0.0496 a minute. Indian cloud-telephony providers such as Exotel publish plan prices rather than per-minute rates on their site, so get a written quote. Outbound calling also falls under TRAI’s commercial-communication rules.

Our AI voice agents page covers latency, handoff and telephony set-up in the US and India.

Infrastructure, channels and people

• Hosting. The model runs at the vendor, so the agent itself needs little compute. An AWS EC2 t4g.medium in Mumbai costs $0.0224 an hour on demand, about $16 a month, in AWS’s price list of 29 October 2026. Production plus staging is two of them; disks, backups and data transfer are extra.
• Vector search. If the agent searches documents, pgvector adds it to the PostgreSQL you already run, under the open-source PostgreSQL licence. Managed vector databases have floors: Pinecone’s Standard plan bills at least $50 a month.
• Tracing. Langfuse Cloud’s Core plan is $29 a month with 100,000 units included, and the open-source edition can be self-hosted for free. LangSmith Plus is $39 per seat a month with 10,000 base traces.
• Channels. Web chat has no per-message fee. On WhatsApp, each business number gets 1,000 free service messages a month, then a reported ₹0.115 each plus 18% GST for Indian recipients since 1 October 2026. Our WhatsApp AI chatbot cost guide has the full rate card and its sources, and our WhatsApp Business API page covers the integration.
• People. The minutes staff spend approving what the agent proposes are a running cost. Measure them in the pilot; they often outweigh the token bill.

12-month total cost of ownership: one illustrative agent

The same order-support agent with a year around it: tier B, web chat and WhatsApp, Claude Sonnet 5.5 or GPT-6.1 Sol with caching, 3,000 tasks a month. Illustrative, not client work. Every line is an assumption you can swap for your own, and we assume ₹90 to the dollar.

Read the bottom row twice. At this volume the year’s cash bill is a few thousand dollars, and the real cost is people: building, upkeep and approvals. At 30,000 tasks a month the token and allowance lines grow about tenfold, to roughly $1,860 a month, and approval time passes 110 hours a month, so both sides matter at scale.

Hidden costs most quotes leave out

• Model retirements. Anthropic promises at least 60 days’ notice before retiring a publicly released model; Claude Sonnet 4.5 was deprecated on 30 September 2026 and retires on 30 November 2026 (Anthropic model deprecations). OpenAI commits to at least six months’ notice for generally available models (OpenAI deprecations). Each forced move means re-running evals and re-tuning prompts. We budget one to two engineer-weeks a year for it.
• Tokenizer changes. Anthropic says the tokenizer used by Claude 4.7 and later produces about 30% more tokens for the same text. Same price per token, bigger bill, unless you re-measure after an upgrade.
• Data residency premiums. OpenAI adds 10% for regional-processing endpoints on models released on or after 5 March 2026. Anthropic charges 1.1 times for US-only inference on Claude 4.6 and later.
• Long prompts. OpenAI doubles rates above 272K input tokens on most models. Claude Haiku 5.5 costs five times as much per token once a prompt passes 100,000 tokens. An agent that stuffs whole documents into context can cross both lines.
• Eval upkeep. Every new failure becomes a test case, and every model, prompt or tool change means a full run. The tokens are cheap; the labelling time is not.
• Incidents. Someone has to read traces when a customer complains, fix the cause and sometimes put the customer’s record right by hand.
• Message category drift. On WhatsApp, Meta can move a template from utility to marketing, which multiplies its price about 7.5 times.

Cost levers, and how big their effect is

Build vs buy: agent platform or custom agent?

Agent platforms fold the build into a usage price, and for many teams that is the right call. The test is whether your task fits the platform and what its price comes to at your volume. Here is each vendor’s published price applied to our 3,000-task example. The arithmetic and the actions per task are our assumptions.

Below a few hundred tasks a month, a platform usually wins, because 14 engineer-weeks buys a lot of outcomes. Above a few thousand, or when the agent must act inside systems the platform cannot reach, a custom build tends to win. Seat-priced assistants are a different purchase: they help staff do their own work rather than run a task end to end. If that is what you are after, start with ChatGPT integration instead.

What to ask for in an AI agent development quote

1. Engineer-weeks by phase, and the seniority of the people doing each phase.
2. The tool list, with read or write permission and an approval rule for each tool.
3. Eval set size, who labels it, and the score the agent must reach before it goes live.
4. A cost-per-task estimate showing steps, tokens per step, model, caching and the 95th-percentile task, not only the average.
5. Monthly running cost split into model, hosting, monitoring, channel and staff time.
6. Step limits and a spend cap per task and per day.
7. Who pays for re-testing when a model is retired.
8. Who owns the code, prompts, eval set and traces at the end.
9. Data processing terms for India’s DPDP Act, GDPR or a US BAA, as your data requires.
10. How you would switch model or vendor later, and what that would cost.

Red flags: a price with no tool list, “unlimited conversations” on a fixed fee (someone is paying for those tokens), no eval set, and token charges hidden inside a platform fee you cannot audit.