Copilot just 9x'd Sonnet and 27x'd Opus and teams have no idea

GitHub quietly updated Copilot's model multiplier table last week, revealing the first visible crack in a subsidy model that was never sustainable. Opus 4.6 jumped from a 3x multiplier to 27x, while Sonnet 4.6 went from 1x to 9x. These multipliers determine how fast premium request pools drain, and the hikes reflect the true cost of serving compute-intensive models. Anthropic is compute-constrained, with Claude Code and agentic workflows consuming 10-100x more tokens per user than simple chat completions. Infrastructure to meet this demand takes 18-24 months to build, while week-over-week compute costs for GitHub Copilot nearly doubled since January. Microsoft and Anthropic have been absorbing the gap, but they're done absorbing it.

On June 1, GitHub moves to full usage-based billing. The multiplier hike is just the warning shot—actual dollar charges will hit corporate cards, traced back to individual usage patterns that nobody thought to govern. Millions of employees have Copilot provisioned by IT departments with zero visibility into model-level consumption, running Opus on everything from code review to boilerplate. Some engineering manager will have a very bad Tuesday in early June explaining why the AI budget is 15x over forecast. Every major provider—OpenAI, Anthropic, Cursor—is unwinding the flat-rate era. The pricing structures being put in place now are designed to make heavy agentic usage reflect its true cost. If your team's workflow depends on treating frontier model access as essentially unlimited, that assumption has an expiration date, and it's soon.