Costs and pricing

How the money moves

Nine meters run on one agent: input, cached and output tokens (reasoning bills as output), tool fees, runtime or sandbox hours, memory, retrieval, tracing and evals, and the people who take escalations. The platform's fee sits on top.

Rough list prices from the three labs' pricing pages in October 2026; they change often, so treat them as orders of magnitude. Google's Pro Preview costs more above 200,000 input tokens, and Gemini 3.8 Flash doubles on January 1, 2027. All three take 50% off for batch. Anthropic and OpenAI add about 10% for US-only or regional processing on newer models. Cache writes cost extra at Anthropic and OpenAI; Google's implicit cache has no write fee. Every tool definition bills on every call, used or not: in Anthropic's own measurement, 58 tools from five MCP servers took about 55,000 tokens before the conversation began. Web search costs roughly a cent a call at all three, and Google gives a few thousand free calls a month.

One refund run, eight model calls

My assumptions: an 8,000-token stable prefix (system prompt, 12 tool definitions); the eight calls of the refund run, with a policy search returning 4,000 tokens; 2,800 output tokens in all, reasoning included; a prompt growing from 8,150 to 17,650 tokens, 111,200 input tokens in total; "warm" means other runs already cached the prefix; Google's implicit cache assumed to hit; list prices.

Caching cuts this run's cost by about 71%. At 100,000 tickets a month, that's about $25,000 uncached against about $7,000 warm on a $2/$10 model, about $400 on the cheapest small model and about $36,000 on a top model. So model choice spreads cost about 100x, caching about 3.5x. Once caching is on, output and cache writes dominate: OpenAI's cheaper reads win at this size, and Google's missing write fee makes up for its higher output price. A runaway loop of 41 extra steps ends at a 69,650-token context and costs almost $4 uncached on any $2/$10 model, about 15 times a normal run. The agent tax: Anthropic prices a single-call support chat at about 3,700 tokens; this run uses about 30 times that.

How companies of different sizes use it

Nobody publishes agent spend by company size, so the spend figures below are my rough estimates from list prices, the refund run above, Ramp's card data and two surveys of large companies. Token spend, not headcount, decides when a company moves up a step. I've kept the usual five headcount steps because they say who owns the agents and how they're bought, but a 30-person company whose product is an agent can outspend a bank's first agent. Anthropic and OpenAI both cap self-serve buying at about $200,000 a month, after which it's a contract; I didn't check where Google, Meta, Mistral or xAI draw that line. The paragraphs describe companies running agents inside the business; where the agent is the product, tokens are a cost of sales and the bill grows with customers.

A startup (under 50 people) runs one or two agents on one lab or one cloud, with an open-source framework and a free tracing tier. The clouds' startup credits, worth up to six figures, now buy frontier models too: AWS Activate covers Anthropic's, Meta's and Mistral's models in Bedrock, Google's AI program covers Gemini, and Microsoft's reportedly cover the models in Foundry. The labs' own programs are smaller (Anthropic gives a one-time API credit and team seats; OpenAI's credits come mostly through investors and partners), so the cheapest tokens come through a cloud. The founders or the CTO own it. My rough estimate runs from nothing to about $150,000 a year, mostly in credits, with the top end a company whose product is the agent. It moves up when the credits expire, when it hits a lab's monthly spend cap and requests fail until the next month, or when a big customer's security review asks about zero data retention and where data is processed.

A scale-up (roughly 50 to 500 people) sells agents to paying customers, or buys a support or IT agent from a system of record it already uses. Those building their own sit on a lab's top self-serve tier or a cloud's catalogue, add a fallback model, move long runs onto a workflow engine such as Temporal, and start paying for tracing and evals (Braintrust, Arize, LangSmith). The head of engineering or an early AI lead owns it, and finance starts asking about the token line. My rough estimate is $70,000 to $1 million a year. It usually moves on after a jolt: hitting the self-serve ceiling or rate limits at peak, a forced model retirement that breaks production and forces a real regression suite, an outage or tighter limits at a lab (OpenAI and Anthropic both rationed capacity in spring 2026), or a bank, insurer or EU customer asking about logs, residency and human oversight.

A mid-market company (roughly 500 to 2,000 people) runs several agents across functions, some bought and some built. It builds on the cloud it already uses (Bedrock and AgentCore, Microsoft Foundry and Copilot Studio, or Vertex AI and Gemini Enterprise), buys agents from its systems of record, and traces runs in the observability tool it already has, such as Datadog. Committed-spend deals with a cloud usually begin here, and a lab's model bought through the cloud counts toward the commitment and skips a new vendor approval, so buying tilts toward the cloud's catalogue. A platform or AI team under the CTO or CIO builds, business units buy, and the CISO and legal review. My rough estimate is $300,000 to $2 million a year in technology. The people cost more: in my rough model of one busy support agent, the staff who take escalations cost far more than the technology, so the share of conversations the agent finishes alone matters more than the token price. It moves on when nobody can say how many agents exist or what they can touch, when auditors or the board ask, or after an incident such as a wrong refund or a leaked document.

An enterprise (roughly 2,000 to 10,000 people) runs dozens of agents across several clouds and labs, with a central platform team running a gateway, an agent registry, evals and policy. It buys the clouds' agent services under its commitment, agents from its systems of record (Agentforce, ServiceNow, Workday, SAP Joule), Microsoft's agent tools in its Microsoft 365 deal and security modules from Palo Alto Networks or CrowdStrike, plus an independent platform where it wants one control layer across labs. The CIO or CTO owns it with a central AI team, the CISO and legal review, and procurement negotiates with labs and clouds. My rough estimate is $1 million to $8 million a year on models and agent tooling, plus systems-of-record licences. Above the self-serve ceiling everything is a contract (OpenAI's Scale and Reserved tiers, Anthropic's Custom tier, a cloud's reserved throughput), carrying the terms in Liability allocation. The hard question becomes how many agents there are, what they can touch and who approved them. Board attention, multi-year capacity deals and regulators in several countries push it further.

A Fortune 500 company (revenue above about $7.5 billion) runs everything at once: hundreds to thousands of agents, every big cloud, several systems of record with their own agents, internal platforms and direct lab contracts. OpenAI named State Farm, Uber and Intuit among early customers of Frontier, its enterprise agent platform, and Google says most of the Fortune 100 use Gemini Enterprise; I found no comparable list for Anthropic, Meta, Mistral or xAI. The CEO and board own the outcome, often with a chief AI officer, while procurement and FinOps own the cloud commitments and multi-year capacity reservations. My rough estimate is $10 million to $50 million or more a year on models and agent platforms, before systems-of-record licences. There's no next size up; the pressure comes from regulators, the board and the market for computing capacity.

What changes as a company grows (spend is my rough estimate):

Buying a finished agent moves the risk. By my rough estimate, vendors charging per conversation, action or resolution price at 2 to 10 times their technology cost, because they sell the result, so I'd read the outcome definition first. Intercom counts a resolution when the customer confirms, doesn't ask for more help, or a hand-off completes; if the customer writes back three days later because the fix failed, that wording decides whether you paid. A model change that inflates tokens hits the vendor's margin under outcome pricing and the customer's bill under consumption pricing.