'What a joke.' That phrase now clings to GitHub Copilot’s new pricing like static, as Microsoft swaps a simple subscription for a token-based meter that many developers say feels more like a throttle than an upgrade.
The change looks less like innovation than retrenchment. Under the new scheme, usage is counted in tokens, a unit tied to the underlying transformer model’s context window and prompt length, and once a user hits the cap, Copilot can slow, degrade, or stop. Developers argue that they are being asked to prepay for uncertainty, since few know how many function stubs, refactors, or code reviews translate into a month of tokens, and GitHub’s documentation so far offers only abstract ranges rather than hard operational guarantees.
For many, the anger is about trust, not just money. Copilot was sold as an always-on coding companion, a flat-fee assistant that turned large language model inference and vector search into an invisible utility; now the product suddenly behaves like metered cloud compute, with silent limits that can trigger mid-sprint. Critics warn that this structure misaligns incentives: heavy users, the ones who feed Copilot the most training signal through telemetry and feedback, are exactly those who will hit ceilings fastest, eroding both goodwill and the data exhaust that sustains Microsoft’s broader AI strategy.
