Elena Voss
Nearly doubled.
That is the phrase Microsoft used on April 20, per the internal document disclosed to Ed Zitron, to describe what had happened to the week-over-week cost of running GitHub Copilot since mid-January. Not the annual run rate. Not the cumulative expense. The week-over-week cost. Copilot in the third week of April cost Microsoft nearly twice as much to operate as it did in the second week of January. Fifteen weeks, and the unit economics flipped.
Microsoft responded the same day with four simultaneous moves:
- Paused new signups for the Copilot Pro, Pro+, and Student tiers.
- Removed Anthropic's Opus models from the Pro tier entirely.
- Restricted Opus 4.7 to Pro+.
- Began migrating Copilot from request-based to token-based billing.
All four moves run in the same direction: stop the variable-cost bleeding.
The same week, Cursor — the AI coding assistant Copilot was supposed to make redundant — was reported to be closing $2 billion in funding at a $50 billion pre-money valuation, roughly double its previous round. The round was co-led by a16z, with Nvidia participating.
Copilot paused the signup button. Cursor doubled. Same week.
The Baseline
GitHub Copilot launched in June 2022 at $10 per month for individuals. That price held for almost four years. In May 2025, Microsoft reported Copilot had reached 15 million users, up 4x year-over-year. In October 2023, The Wall Street Journal had already called the product "the $10/month AI assistant" and noted Microsoft was subsidizing the cost.
The $10 price was not a market-clearing price. It was a strategic one. Microsoft had decided that winning developer mindshare was worth running the product at or below cost. Azure would recapture the economics elsewhere. Copilot was loss-leader infrastructure for the Microsoft AI bundle.
The math of that subsidy has two variables: cost per token, and tokens per user. Both were supposed to improve. Token prices were supposed to fall faster than per-user consumption would rise. That was the bet in 2022. It has been the bet every quarter since.
In fifteen weeks between January and April 2026, both variables moved in the wrong direction at the same time.
The Model Microsoft Doesn't Own
The specific driver is the model mix. Anthropic released Claude Opus 4.7 in early 2026. It is the model Copilot users have been defaulting to for the hardest problems — long-context refactors, multi-file reasoning, agentic work. It is also the most expensive frontier model in production by per-token cost. Users don't choose models by price. They use the most capable one their plan allows.
Opus is an Anthropic model. Microsoft doesn't own it. Anthropic's inference runs primarily on AWS and Google Cloud. The Azure capacity Anthropic committed to buy in November 2025 is, at the margin, a training purchase inside a multi-cloud footprint. When a Copilot Pro subscriber runs an agentic Opus 4.7 session, Microsoft pays a competitor's wholesale token bill on a competitor's infrastructure, and charges the user $10.
That is not a pricing error. That is a structural admission. The architect of the AI-coding market no longer owns the model that makes its flagship product work. Once the best model isn't yours, the flat-rate subsidy is no longer a distribution strategy — it's a transfer payment to your model supplier, routed through every user who runs the feature the subscription exists to deliver.
Microsoft didn't raise the price. It paused the product.
Cursor
The Cursor raise is the counter-number. Cursor is a fraction of Copilot's user base. It has never had a flat-rate unlimited tier — its pricing has been usage-aware since launch. And the investors who co-led the $2 billion round did so after watching the dominant incumbent pull its signup form.
Two days earlier, xAI had agreed to let Cursor train Composer 2.5 on Grok's infrastructure. Three days later, SpaceX announced it was working with Cursor to build "the world's most capable coding agent." A company that charges by consumption is being handed the largest engineering teams in tech. A company that charges flat-rate is closing its signup form to new users.
This is not the narrative mean-reversion of "AI is getting expensive." It is a regime change in how AI coding products are priced. The subsidized flat-rate tier was an assumption about what the incumbent could sustain at scale. The assumption was wrong by a factor of two, in fifteen weeks.
Fifteen Weeks
The $10 price point held for 200 weeks. Its retirement took 15. By any reasonable benchmark, that is not a pricing adjustment. That is the bet the market just made about what AI coding costs and who captures the economics.
Cost nearly doubled. The product paused. The challenger doubled. Whatever comes next will be metered.