How much do AI agents actually cost in 2026

What “AI agent” actually means

An AI agent, as vendors sell it in 2026, is software built on a large language model that carries out a multi-step task with limited supervision, not just a single answer. A chatbot answers “what is your refund policy”; an agent reads an inbox, drafts three replies, flags one for a human, and updates a spreadsheet, without being told each step. That autonomy also makes an agent harder to price than a chatbot: a chatbot subscription is usually a flat monthly fee, while an agent’s bill depends on how many tasks it runs, which is why three pricing models now compete for the same small-business budget.

Per-seat subscriptions: the familiar bill

The simplest way to buy agent capability is a monthly seat. OpenAI’s ChatGPT Business charges $20 per user monthly on annual billing, or $25 month-to-month, for a Standard seat, and $100 annually (or $125 monthly) for a Premium seat with five times the usage and no five-hour cap, per 2026 pricing-tracker reporting sourced from OpenAI’s own business page (justinmckelvey.com, seen September 12, 2026; OpenAI’s own pricing pages returned access errors on direct verification). Anthropic’s Claude Team lists the same $20/$25 split for a Standard seat and $100/$125 for a Premium seat that adds Anthropic’s coding-agent tools, while Claude Enterprise is billed at $20 per seat plus usage at standard API rates (claude.com/pricing, seen September 13, 2026). Microsoft’s 365 Copilot is priced separately: $21 per user monthly on an annual Business plan, with a promotional $18 through December 31, 2026, or $30 for Enterprise (justinmckelvey.com, seen September 13, 2026; Microsoft’s own pricing page timed out on verification). Copilot never stands alone — it is billed on top of a Microsoft 365 base subscription a business already pays for.

• ChatGPT Business Standard seat: $20/month annual, $25/month monthly
• Claude Team Standard seat: $20/month annual, $25/month monthly
• Microsoft 365 Copilot Business: $21/month annual, plus a required Microsoft 365 license

Usage-based pricing: the meter nobody puts on a landing page

Underneath every seat, and for anyone building a custom agent instead of buying one, compute is billed per token. In plain terms, a token is a small chunk of text, roughly three-quarters of a word; providers charge one price per million tokens sent in and a higher price per million tokens the model writes back. OpenAI’s published API rates run from $0.05 input / $0.40 output per million tokens for GPT-5-nano up to $10/$50 for GPT-6 Astra, with mid-tier GPT-5.6 “Terra” at $2/$12 (developers.openai.com/api/docs/pricing, seen September 13, 2026). Anthropic’s Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens, and Claude Haiku 4.5 costs $1 and $5. Anthropic’s own documentation gives a worked example: processing 10,000 support tickets, averaging roughly 3,700 tokens each, on Haiku 4.5, totals about $37 (platform.claude.com/docs/en/about-claude/pricing, seen September 13, 2026) — about $0.0037 per ticket. As an illustrative example only, not a vendor-published figure: applying that per-ticket rate to a hypothetical agent triaging 50 emails a day works out to under $0.20 a day, or roughly $4 to $6 a month in raw model calls, before the extra tokens an agent’s own tool calls add and before the cost of whatever platform runs the workflow.

Agent platforms and marketplaces: paying by the outcome

A third model skips both the seat and the raw token meter: a platform charges for the automation itself. Zapier’s core plans start at $19.99 a month on annual billing (or $29.99 monthly) for a Professional plan with 750 tasks (zapier.com/pricing, seen September 13, 2026). Since June 15, 2026, an AI step inside a Zap consumes 1x, 3x or 5x a normal task depending on its model tier, with tool calls billed the same way. Salesforce’s Agentforce, aimed at larger service teams, sells three meters at once: $2 per resolved conversation, Flex Credits at $0.10 per action, or a flat $125-per-user add-on for unmetered access (eesel.ai, last edited July 14, 2026). The common thread is that the bill tracks how much work the agent does, not how many people can log in.

Hidden costs: what doesn’t show up on a price list

None of the figures above include the time a human still spends checking an agent’s output. That review does not disappear because a task is automated: someone has to notice a bookkeeping agent miscategorizing a refund, or an email agent sending a badly worded reply, and fixing it usually costs more than the API call did. Integration work — connecting an agent to an existing inbox, accounting package or customer database — is often a separate line item, whether that means a Zapier plan upgrade or a dependent license like Copilot’s. From August 2, 2026, the EU AI Act’s transparency rules add a compliance duty rather than a cost line: any AI system that interacts directly with a person, including a chatbot or an agent that autonomously contacts someone, must make that clear inside the interaction itself, not buried in terms and conditions. A company that gets this wrong can be fined up to €15 million or 3 percent of global turnover (European Commission statement dated August 2, 2026, seen September 13, 2026).

Reality check: one person, several small bills

For a one-person business in Europe running one lightweight agent for email and light bookkeeping: a Standard seat on ChatGPT Business or Claude Team runs about $20 a month; a Zapier Professional plan to connect that agent to an inbox and a spreadsheet adds roughly $20 more; and model usage for a few hundred short tasks a day, at nano- or Haiku-tier rates, adds $5 to $15 depending on volume. Adding those figures together — this article’s own illustrative arithmetic, not a vendor’s quote — gives a realistic all-in range of roughly $45 to $55 a month for one lightly used agent, before the unpaid hours spent reviewing its output, a cost no vendor pricing page states.

We use AI in our daily operations; posts are reviewed by people before publishing; automated/generated posts are marked as such where needed.