GitHub Copilot is ditching its flat-rate subscription for usage-based pricing starting June 1, 2026. Instead of a set number of requests, Microsoft will charge users based on actual model compute costs. The company framed the shift as Copilot evolving from an in-editor assistant into an "agentic platform" capable of multi-hour autonomous coding sessions.

The real story runs deeper than product evolution. Per the Wall Street Journal, Microsoft was losing over $20 per user monthly on GitHub Copilot, with some power users costing the company $80 a month against a $10 subscription. For three years, Microsoft absorbed the escalating inference costs while users burned more tokens than their subscriptions covered.

The subscription model fundamentally doesn't work for generative AI. AI startups are annihilating cash—Anthropic reportedly allowed users to burn $8 in compute for every dollar of subscription. Companies assumed they'd raise prices later or that inference costs would drop. Neither happened. Newer "reasoning" models burn way more tokens, making inference more expensive over time.

Every AI service you've used has subsidized compute, and every service has lost money as a result. The "subprime AI crisis" has arrived: companies must now pass real costs to users or face unsustainable losses.