AI agent development cost in 2026: the build bill and the run bill

An AI agent has two bills. The build bill covers design, integrations, evaluations, guardrails and human handoff, and depends on autonomy and how many systems the agent touches. The run bill covers model tokens, retrieval, monitoring and voice minutes, and it is maths you can do before you start. Here is that maths on official October 2026 prices.

Most "AI agent development cost" pages that rank today are written by agencies, give price bands "from our experience" with no data behind them, and still quote models that have been retired for over a year. We do the opposite. We don't publish build prices, because two agents with the same name can differ enormously in scope. Instead we explain what drives the build, and we show the run bill line by line, using only list prices we read on the vendors' own pages on 23 October 2026.

What are the two bills: build vs run?

The two are linked. A build that adds prompt caching, a cheaper model for simple steps and a hard cap on tool loops can cut the run bill by more than half. A build that skips evals usually pays for it later in both bills.

What drives build effort?

We quote every agent from a written scope, so here are the drivers we size, without price bands:

• Autonomy. An answer bot that cites your help centre is far simpler than an agent that decides its own steps, calls tools and acts in other systems. Read AI agent vs chatbot for where your use case sits.
• Tools and integrations. Every system the agent reads or writes (CRM, helpdesk, ERP, booking, payments) needs an integration, permissions, error handling and tests. Model Context Protocol helps; see MCP vs API.
• Retrieval. Grounding answers in your documents means ingestion, chunking, a vector index, access control and freshness rules. If you're weighing retrieval against fine-tuning, see RAG vs fine-tuning.
• Evals. A test set of real conversations, scored automatically on every prompt or model change. This is the work most cheap quotes leave out.
• Guardrails. Input and output checks, tool permissions, spending caps, loop limits and refusal rules.
• Human handoff. Routing to a person with context, and a way for staff to correct the agent.
• Risk level. An agent that touches money, health data or insurance claims needs more review, logging and approval steps than one that books meetings. Regulated builds such as insurance software carry more of this work.
• Disclosure and compliance. EU AI Act transparency duties (below), privacy notices, data retention and audit logs.
• Channels. Web chat, WhatsApp, email, phone and in-app each add integration and testing work.

How do you calculate LLM token costs?

Model providers charge per million tokens (MTok), with separate prices for input, cached input and output. The formula for one month:

1. Calls = conversations × model calls per conversation.
2. Input tokens = calls × input tokens per call, split into the cacheable part (system prompt, tool definitions, policies) and the fresh part (the user's messages and retrieved content).
3. Output tokens = calls × output tokens per call.
4. Cost = fresh input × input price + cached input × cache-hit price + output × output price, all per MTok.

Agents make several calls per conversation (plan, call a tool, read the result, answer), and every call resends the context. That is why agents cost more to run than single-turn chatbots, and why caching matters so much.

Worked example: one support agent on Claude, OpenAI and Gemini

Illustrative example, list prices read 23 October 2026, USD per MTok. Assumptions: 10,000 conversations a month × 6 model calls × 8,000 input tokens per call, of which 6,000 are a cacheable prefix, plus 400 output tokens per call. That gives 480M input tokens (360M cache hits + 120M fresh) and 24M output tokens.

Sources: Anthropic pricing, OpenAI API pricing and Gemini API pricing. What the table leaves out: cache writes (Anthropic charges 1.25× the base input price for a 5-minute cache write and 2× for a 1-hour write), Gemini's cache storage charge, embeddings, retries and taxes. Treat these figures as a floor and add headroom for retries and growth.

Two lessons stand out. First, model choice moves the bill by an order of magnitude, so route simple steps (classification, extraction, short replies) to a small model and keep the large one for reasoning. Second, without caching, the same Sonnet 5.5 workload would cost about $1,200 a month instead of $516.

Caching and batch: what are the biggest levers?

• Prompt caching. Put the stable parts of every call (system prompt, tool definitions, policies, few-shot examples) first, and the changing parts last. In our example, a cache hit on Claude Sonnet 5.5 costs $0.10 per MTok against $2 for fresh input, a 95% cut on the cached share.
• Batch processing. Work that doesn't need an instant answer (overnight evals, ticket tagging, document summaries) can go through batch APIs. Anthropic lists "a 50 percent discount on both input and output tokens", Google lists "50% cost reduction" for its Batch API, and OpenAI's batch rates for gpt-6-sol are half its standard rates ($1.00 input, $5.00 output).
• Model routing. A small model triages, a large model handles only the turns that need it.
• Shorter context. Retrieve fewer, better chunks; summarise long threads; cap tool output length.
• Loop limits. A hard ceiling on tool calls per conversation stops a confused agent from spending all night.

Which price cliffs should you plan for?

List prices change, and three changes are already published:

If your business case only works on a promotional or pre-increase price, it doesn't work. Build your estimate on the post-change price and make switching models a configuration change, not a rebuild.

How much does a voice agent cost per minute?

Voice agents are priced per minute rather than only per token. Illustrative example: 5,000 calls a month × 3 minutes = 15,000 minutes.

OpenAI notes that for gpt-live-1, "Backend model and tool usage is charged separately", so add the token maths above. Add telephony (phone numbers and carrier minutes) too, which neither price includes. See ElevenLabs API pricing and our AI voice agents service for how we design voice flows.

What else is in the run stack?

The people line is the one budgets forget. An agent that nobody reviews drifts: policies change, products change, and the knowledge base goes stale.

What does compliance cost? EU AI Act Article 50 is live

The European Commission's AI Act page says "The transparency rules of the AI Act will come into effect in August 2026." For most business agents, that means telling people they are talking to an AI system unless it's obvious, and marking AI-generated content in some cases. Law firms report that the Digital Omnibus proposal gives a grace period to 2 December 2026 for the Article 50(2) content-marking duty; treat that as reported, not settled. Budget for disclosure copy, logging and a review of what your agent generates. This is general guidance, not legal advice; your counsel confirms what applies to you.

Why do over 40% of agentic AI projects get canceled?

Gartner predicted on 25 June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027. Press coverage of the release attributes this to escalating costs, unclear business value or inadequate risk controls. Each of those maps to a decision you make before the build:

• Escalating costs: do the token maths above, cap loops and choose models per step.
• Unclear value: pick one workflow with a measurable outcome (tickets resolved, hours saved) and a baseline.
• Risk controls: evals, permissions, human approval for anything irreversible, and logs you can audit.

Is a chatbot cheaper than an AI agent?

Usually, yes, on both bills. A chatbot that answers from your help centre makes one or two model calls per turn and integrates with few systems. An agent makes several calls per task, uses tools and writes to other systems, so it needs more integration, testing and guardrails, and more tokens per conversation. Start with the least autonomous design that meets the goal. Our AI chatbot development and LLM integration services cover the simpler end.

What does an agent product look like in practice?

Our flagship build, Kenerate AI, was designed and engineered by Elgrith. It is an all-in-one AI image and video studio with 200+ AI models in one app. Two parts are agentic: Kenerate Agent, a chat in which an agent picks the model and runs the workflow for you, and MCP and CLI access so Kenerate works inside Claude, ChatGPT, Cursor and 19+ AI apps. Read the Kenerate AI case study for what we built.

How can Elgrith help?

Elgrith has delivered 1,000+ projects for 250+ clients since 2015, from our New Delhi HQ and client-partnership centres in New York, London and Shanghai. Our AI agent development team scopes the build, writes the run-cost model with you before any code, and designs caching, routing and evals in from the start. You own the code, prompts and evals, and you get a fixed written quote after a free call. Book a call or start a project.