GitHub Copilot shifted to metered billing on June 1, ending unlimited usage for paid subscribers and introducing an explicit cost structure that reveals what AI-assisted coding actually expenses.

The Pro plan remains $10 monthly but now includes a finite pool of AI credits priced at one cent each. Credit consumption varies by model and token count. Developers immediately complained about the change, forcing GitHub to clarify pricing across its different coding assistant tiers.

This move marks a critical moment for AI tooling economics. For years, vendors subsidized agent and tool usage to build adoption and lock-in. GitHub could absorb Copilot's compute costs because the service drove retention and network effects. That calculus shifts once usage scales or infrastructure costs rise.

The pricing reveal matters because it strips away abstraction. When developers see explicit per-token costs, purchasing behavior changes. A penny per credit sounds cheap until an eight-hour coding session consumes thousands. Enterprises suddenly face accountability conversations with engineering leaders about AI tool sprawl.

GitHub's move also signals confidence in the market. The company believes developers will pay even with transparency. It also suggests GitHub's own economics require charging for usage rather than absorbing it. The shift from flat-rate to consumption-based pricing typically indicates either rising demand that strains margins or investor pressure to demonstrate monetization paths.

Competitors face immediate pressure. Tools like Cursor, JetBrains AI Assistant, and others must decide whether to follow suit or exploit the opening by maintaining unlimited plans. Staying unlimited only works if margins exist or funding extends. Most won't survive on that strategy long.

The broader pattern extends beyond Copilot. Claude's API pricing, OpenAI's token costs, and Anthropic's enterprise deals all expose the real cost structure beneath AI products. When abstractions fall away, developers optimize. Some shift to smaller models. Others reduce usage. A few simply