Costs and pricing

How the money moves

Underneath, every product now prices the same way: a seat with some usage included, then usage at or near the lab's API rates. List prices in October 2026, rounded:

Under the seats sit token prices. Anthropic lists Opus 5.5 at $4 per million input tokens, $0.20 for cached input and $20 for output, and Sonnet 5.5 at half that. Cursor prices its own Composer 2.5 at $0.50 in and $2.50 out, and Mistral's Devstral 2 costs about the same. OpenAI prices Codex in credits whose dollar value depends on the plan or agreement.

A seat buys little agent time. An agent working on a big task makes 60-150 model requests an hour, each carrying about 150,000 tokens of context, most of it cached. By my arithmetic that's about $5.50-14 an hour on Opus 5.5 and $3-7 on Sonnet 5.5, so a $19 Copilot Business seat covers roughly one and a half to three and a half hours on Opus 5.5, if GitHub's credits track list prices as it says. Anthropic's own figure for Claude Code is an average of about $13 per developer per active day, or $150-250 a month, with nine in ten users under $30 a day. The tail is the problem: companies told one newsletter of engineers spending $500 a day and of a $10,000 week caused by a caching error.

Flat plans kept breaking on heavy agent users in 2025 and 2026: a tail costing 10-100 times the plan price, then caps, then usage pricing, then protests. Cursor moved its Pro plan to usage in June 2025, then refunded surprise charges and its CEO apologized. Anthropic added weekly caps in August 2025 after one account ran tens of thousands of dollars of usage on a $200 plan. GitHub paused self-serve Business signups on April 22, 2026 for lack of capacity, metered everything from June 1 and reopened signups on September 3, and some Copilot users reported projected bills rising from $29 to about $750 a month.

The app layer's margin is set by what it pays the labs. In 2025 Windsurf's margins were reportedly "very negative". Cursor ran negative gross margins until early 2026 and reached a slightly positive one by routing to its own Composer model and to cheaper ones such as Kimi, according to four unnamed sources.

How companies of different sizes use it

Nobody publishes AI coding spend by company size, so the spend figures below are my rough estimates from list prices. Almost every price here is per developer plus that developer's usage, so I give each size a typical number of engineers. Even that is a loose guide: the share of heavy agent users moves the bill more than headcount, and two companies with the same number of engineers can spend five to ten times apart.

Which tools a company picks changes with size too. In The Pragmatic Engineer's survey of engineers early in 2026, Claude Code was the most used tool at the smallest companies (75% of respondents there) and Copilot at companies of 10,000 people or more (56%), while Cursor's use was roughly flat until it dipped at the largest. JetBrains' January 2026 data also shows Copilot's use rising at companies above 5,000 people. The authors of both surveys put the split down to enterprise buying, with Copilot coming through existing GitHub and Microsoft contracts, rather than to what developers prefer.

A startup (under 50 people, usually 3 to 25 engineers) runs agents in the editor or terminal on its engineers' own machines, mostly with their own credentials, the riskiest setup in What mistakes cost. Each engineer picks their tools, often two at once on individual plans paid by card, with free tiers and open-source command-line agents filling gaps. Anthropic, OpenAI and GitHub all run startup programs, and AWS's startup credits reportedly pay for other companies' models on Bedrock, but the useful part is usually API credits for the product rather than seats for engineers. The founders or the CTO own it, and my rough estimate is $5,000-100,000 a year. What pushes it on is a big customer's security questionnaire asking whether the tools train on the company's code. Several individual plans do unless the user opts out (see the plan table in How the rules work), so the answer is business plans with no training and an indemnity.

A scale-up (about 50 to 500 people, 20 to 200 engineers) standardizes on one team plan, such as Copilot Business, Cursor Teams, Claude Team or Codex Business, often with another tool on top for its heaviest users. It adds AI code review and background agents for work the tests can check. A VP of engineering or an early platform lead owns it. Finance notices when expensed plans turn into an invoice. The invoicing company sits at the top of this band, where the heaviest users set the bill: model choice, cache misses and parallel agents move it more than the choice of seat. A seats-only plan with overage off works as a budget cap rather than a product choice, since the heavy users burn the pooled credits early in the month and then agents pause for everyone. My rough estimate is $50,000-750,000 a year. Bill shock from the heavy tail, an incident from an agent with real credentials, or an auditor asking who approved an agent's change pushes it on, and procurement steps in once the total passes what an engineering leader can sign alone.

A mid-market company (about 500 to 2,000 people) looks different depending on whether it sells software. A software company this size has a few hundred engineers, standardizes on one or two tools under enterprise terms (single sign-on, audit logs, no training, an indemnity, budgets on pooled usage) and adds code security. A manufacturer or insurer may have a few dozen developers and, as far as I can tell, usually takes Copilot through its Microsoft agreement. The CTO owns it with a platform team, while the CISO and legal approve the data terms and procurement sets budgets. This is where buying moves onto commitments already signed. Each cloud sells its own tool on its own bill (Copilot through the Microsoft agreement, Kiro on AWS, Gemini Code Assist on Google Cloud) and resells lab models through Foundry, Bedrock and Vertex. AWS's marketplace also sells Claude for Enterprise seats with Claude Code, and since March 2026 Anthropic lets a company spend part of its Anthropic commitment on listed partners such as GitLab, Lovable and Replit. Whether a marketplace purchase counts toward a cloud commitment is something I'd check in the contract. My rough estimate is $100,000-2 million a year for a software company and far less elsewhere. Tool consolidation, an acquisition or an auditor's question about AI-written changes pushes it on.

An enterprise (about 2,000 to 10,000 people, 500 to 5,000 engineers) usually gives everyone a default seat tool, most often Copilot through its Microsoft agreement, and puts heavy users on a lab's agent such as Claude Code or Codex, or on Cursor, with AI review and migration agents on top. Data terms start to decide the choice: GitHub and Anthropic both charge a premium to keep processing in one region, and Cursor and Anthropic offer zero data retention to enterprise customers. A central developer-platform team runs the tools, the CTO or CIO owns the budget, and the CISO, legal and an AI governance group approve data terms and agent permissions. My rough estimate is $0.5-10 million a year. Even at the top that's a few percent of engineering payroll, so it pays for itself if engineers ship a few percent more, and that's what the board starts asking about as business units and regulators multiply.

A Fortune 500 company (thousands of engineers or more) runs several tools by business unit. Goldman Sachs said in 2025 it would deploy hundreds of Cognition's Devin agents on legacy code under human supervision, and the biggest tech companies build their own harnesses and route work to their own or cheaper models, as Microsoft does inside Copilot. Cursor says 64% of the Fortune 500 use it, which is its own figure, and GitHub says Copilot has tens of thousands of business customers; I found no comparable figure from the labs. Government work needs FedRAMP, and vendors are getting there by different routes: Claude Code is in public beta in a FedRAMP High environment, Copilot's model hosts are authorized at FedRAMP Moderate with Copilot itself on the way, and Kiro runs in AWS GovCloud. The CTO or CIO owns it with the CFO, an AI governance council, and the finance team that holds the cloud and Microsoft commitments. My rough estimate is $3-50 million or more a year. There's no next size up, so the pressure moves to leverage at renewal: the company's own keys with the labs, several agents side by side in GitHub's Agent HQ, or a credible in-house option.

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

What one merged change costs

My rough arithmetic for one background-agent task, from a small fix to a multi-hour job:

Human review is the biggest line in nearly every case, which is the bottleneck written in money. The merge rate multiplies everything: at 40%, the tokens behind each merged change cost two and a half times the per-task figure. At the invoicing company, if 40 senior engineers each absorb three to eight extra hours of review a week, that's roughly $0.5-1.4 million a year, more than its AI tool bill in most setups. Security tooling adds less: GitHub's Secret Protection and Code Security at $19 and $30 per active committer a month come to about $118,000 a year for 200 committers.