Costs and pricing

How the money moves

The builder pays per token, and the money flows down to whoever owns the chips, the buildings and the power. List prices per million tokens in October 2026, input / output:

And the modifiers that apply on top:

Under these sheets, prices move in different directions. The price of a fixed level of capability falls: a16z measured about 10 times a year at a constant benchmark score, and Epoch AI found GPT-4-level performance on a hard science benchmark getting 40 times cheaper a year. The top of the ladder stays put, because each generation adds a rung: OpenAI's flagship launch prices fell from $30 per million input tokens for GPT-4 to $1.25 for GPT-5 and climbed back to $10 for gpt-6-astra, while Anthropic cut Opus from $15 to $4 and added Fable at $10. And the cost per task rises: on Artificial Analysis's index Opus 4.6 used twice the output tokens of Opus 4.5 on the same tests. In OpenRouter's 100-trillion-token study a 10% price cut brought only 0.5-0.7% more usage, so spending grows from new uses such as agents and coding far more than from cheaper tokens.

Prices also went up. Bedrock keeps retired Claude 3.5 Sonnet under "extended access" at twice its launch price, Gemini 3.6-3.8 Flash list a 2027 price twice the 2026 one, and Zhipu, Moonshot and MiniMax raised some API prices. Compute got dearer too: one-year H100 rental contracts rose about 40% between October 2025 and March 2026, and B200 spot prices more than doubled in six weeks in spring 2026.

How companies of different sizes use it

Headcount is a weak ruler for model spend. An AI-native company, one whose product is mostly model calls so the token bill is its cost of goods, can spend more on tokens with 50 people than a large bank does. The ladder that matters is the labs' own spend tiers: self-serve tiers with monthly caps, then a custom tier through sales, then cloud commitments and reserved capacity. I've kept the usual five sizes and noted in each where an AI-native company differs. Nobody publishes model spend by company size, so every spend figure here is my rough estimate.

A startup (under about 50 people, up to Series A) that's adding AI runs one or two features on one lab's API, plus coding agents for its engineers. It starts at OpenAI, Anthropic or Google, where the newest models land soonest, tries others through OpenRouter and lives in the self-serve tiers. Credits pay for most of year one and shape the route: a lab's credits are small and work only on its own API, while the clouds' startup credits are far larger and pull a company toward that cloud's catalog. Whether they cover another lab's model there depends on the cloud's terms, so I'd check before counting on it. The CTO or a founding engineer picks the model, and the founder sees the bill on a company card. My rough estimate is under $50,000 a year on tokens. An AI-native startup carries the token bill as cost of goods from launch, at my rough estimate of $50,000 to $1 million a year. Either kind moves on when a tier cap stops the feature until the 1st of next month, when credits expire, or when the first enterprise customer's security review asks for zero data retention, residency and an IP indemnity.

A scale-up (about 50 to 500 people, Series B to D) has several AI features in production and coding agents for most engineers, and the model bill becomes a line the CFO asks about. It usually runs two or three labs plus an open-weight model for the easy traffic, behind a gateway such as LiteLLM, Portkey or OpenRouter with budgets and a fallback. The head of engineering or a newly hired AI platform lead owns it, and finance starts a monthly review because token bills swing. My rough estimate is $100,000-750,000 a year. An AI-native scale-up spends millions and reaches the top self-serve cap, $200,000 a month at both Anthropic and OpenAI, which forces a sales contract. What pushes either kind on is that ceiling, enterprise customers who want an SLA with credits and data terms in writing, and a first cloud commitment.

A mid-market company (about 500 to 2,000 people), like the help-desk company in this guide, runs a portfolio of AI features and internal agents with a central gateway and a shared eval set. The cloud it already runs on becomes the default way to buy models, because the spend draws down a commitment and procurement has already approved the vendor. Claude is sold on Bedrock, Foundry and Vertex AI, OpenAI's models on Foundry and Bedrock, and Gemini on Vertex AI, and Bedrock's SLA with credits, which a lab's standard tier doesn't offer, often settles it. The route moves the price less than people expect, though. For the help-desk button, the model tier and the reasoning setting move the monthly bill by orders of magnitude, from a few hundred dollars on a small model to tens of thousands at the top, while the route, caching and batch move it by tens of percent. A CTO or CIO owns it with an AI platform team, FinOps owns the commitment, and security and legal check retention, residency and indemnities. My rough estimate is $250,000-2 million a year. An AI-native company this size buys capacity more than features: a custom tier and provisioned throughput, priced as cost of goods. Commitment renewals, regulated customers and an acquisition that brings a second cloud push it further.

An enterprise (roughly 2,000 to 10,000 people) has many teams on several labs, and a central AI platform with budgets per team and chargeback. It buys lab enterprise agreements and the cloud platforms side by side, reserves capacity for features where speed matters and sends back-office work to batch. Seven-figure accounts with a single lab are common at this size. It self-hosts some open-weight models for residency or fine-tuned weights; on cost alone, self-hosting pays only at tens of billions of tokens a month with steady load, because providers run newer chips at higher utilization. A chief data and AI officer or the CIO owns it, with procurement on the agreements, legal on indemnity and retention, and security on the gateway and its keys. My rough estimate is $2-15 million a year. Guaranteed capacity is harder to buy than a discount: it comes through reserved products such as OpenAI's Reserved Tier and the clouds' provisioned throughput, and Anthropic no longer sells its Priority Tier to new buyers. An AI-native company spending this much is one of its labs' biggest customers, with more leverage than most banks and more exposure if a lab starts selling the same product. Regulation, acquisitions that add clouds and retirements that force re-tests across dozens of teams push it on.

A Fortune 500 company (the largest US companies by revenue) runs everything at once: three or more labs on two or three clouds, self-hosted open-weight models, internal agent platforms, and seat products such as ChatGPT Enterprise, Claude for Work, Microsoft 365 Copilot and Gemini Enterprise. Each of the big labs and clouds says most of the largest companies are its customers, so what matters is how the spend is split. A board-level AI or risk committee sets the rules, the CIO or chief data and AI officer runs the platform, procurement and FinOps own the commitments, and the CISO owns the gateway and data controls. My rough estimate is $10-100 million or more a year for the heaviest, and many spend far less across several vendors. Still, the biggest single buyers of tokens are AI-native apps; Cursor, built on model APIs, reached about $2 billion of annual revenue in just over two years. There's no next size up; the pressure comes from regulators, residency rules and boards asking how much depends on one lab or one cloud.

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

Where the dollar goes

• Lab direct: customer to lab, at list less any negotiated discount (I found few public ones), then to the lab's compute. Gross margins are unaudited estimates for every lab: The Information, as reported by Reuters, put OpenAI's adjusted gross margin at 33% in 2025, down from 40% as inference costs rose, and I found no reliable figure for Anthropic, Google's Gemini or xAI.
• Cloud marketplace: customer to AWS, Azure or Google, drawing down a commitment, then to the lab. Marketplaces were 47% of Anthropic's 2025 revenue in its leaked draft filing, as reported.
• Inference provider: customer to Together, Fireworks or Baseten, then to GPU clouds or their own clusters, then to Nvidia. Cerebras, the one with public numbers, reported a GAAP gross margin of 14% in the quarter to June.
• GPU cloud and chips: lab or hyperscaler to CoreWeave, Nebius or Oracle, then to lenders, utilities and Nvidia, which earned $59.7 billion of net income on $96.2 billion of revenue in one quarter, the richest seat in the chain.