Anthropic and OpenAI both changed how they charge enterprise customers in April 2026, aligning seat pricing to raw API token costs and effectively ending the deep discounts that large companies had relied on. Simon Willison, the developer and longtime AI observer, published an analysis on May 27 arguing this is not opportunism. It is the signature of genuine product-market fit.
The arithmetic is blunt. Willison ran the ccusage tool against his own Claude Code and OpenAI Codex consumption over the prior 30 days and found roughly $2,180 in API-equivalent value against $200 in subscription fees. His conclusion: the $10-to-$20-per-seat consumer pricing model was never going to cover frontier inference costs. Coding agents broke the equation in the labs’ favor because each agent session burns far more tokens than a chat exchange, and the professionals who run those sessions are already spending heavily on the work those sessions replace.
That math connects to numbers we have covered previously. Claude Code is tracking toward $2.5 billion in annualized revenue. Cursor, which built its product on Anthropic’s API, is running at a $3 billion annual rate. Both figures emerged before April’s pricing change took full effect across enterprise contracts. The Colossus compute deal Willison cites, at $1.25 billion per month paid by Anthropic to SpaceX through May 2029, implies inference budgets that no consumer subscription base could sustain.
The structural point is worth stating plainly. Per-seat SaaS pricing at $10 to $20 per user was designed for software with near-zero marginal cost per seat. Frontier-model inference has steep marginal cost per query. Coding agents collapse that tension because each agent run is a discrete, high-value, billable unit: the developer using Claude Code to automate a full pull request has a willingness to pay that reflects the hours of effort being substituted, not the entertainment value of a chat response. That willingness to pay is what allows the labs to price at API rates and have enterprise buyers still sign.
Willison’s PMF framing reads cleanly for the coding vertical. It is less clean elsewhere. Legal AI, to take one example, is growing at roughly 7 percent annual revenue at Harvey (per Harvey’s own reported figures) and still requires senior attorney review on every substantive output. Qualitative analysis, strategic writing, and research synthesis require human judgment that clients are not yet willing to remove from the loop. PMF in coding does not transfer automatically to those verticals, and Willison, who is a builder and not a financial analyst, does not try to claim it does. The risk is that readers extrapolate the coding-agent thesis into a general claim that AI has found its business model across the board.
Two counter-pressures are worth noting. DeepSeek’s V3 and R1 releases drove commodity inference prices down sharply in early 2026. That price war is ongoing. If open-weight model capability continues closing the gap with frontier models, the API-rate pricing that looks sustainable today faces structural pressure from substitution. Anthropic’s own compute-cost-per-revenue-dollar has been improving, but the $1.25 billion monthly Colossus commitment shows how exposed the business remains to inference economics.
The Uber and Microsoft stories Willison examines also carry a different read than the PMF one. Uber maxing its AI budget and Microsoft pulling Claude Code licenses could signal that buyers are finding the value-to-cost ratio uncertain at enterprise API rates, not that they are grudgingly accepting a fair price. The two interpretations are not mutually exclusive, but enterprise procurement teams will stress-test that distinction hard over the next two quarters.
Enterprise buyers evaluating coding-agent contracts should now build token-consumption benchmarks into procurement, not just seat counts. The April pricing shift means that usage variance, not headcount, drives the budget exposure. Locking a year-long deal without per-user token caps is how finance teams end up with Uber’s budget problem by Q3.
Posted by Simon Willison on simonwillison.net on 2026-05-27.
