The era of subsidised AI coding is ending

The era of subsidised AI coding is ending.

From 1 June 2026, GitHub Copilot will move to usage-based billing for Business and Enterprise plans. Instead of counting premium requests, every Copilot plan will include a monthly allocation of GitHub AI Credits. Usage will be calculated based on token consumption, including input, output, and cached tokens, using the listed API rates for each model.

Anthropic moved in the same direction at the beginning of April. Organizations currently using seat-based Enterprise plans with Standard and Premium seats will not be able to continue with this billing model past your next contract renewal.

Cursor also started moving away from the old simple request based model for business and enterprise usage earlier this year.

I think this is not just a pricing change. It is the end of the "AI is almost free" phase.

For the last year, many teams treated AI coding tools like unlimited subscriptions. Run more agents. Add more review loops. Use huge context. Let agentic workflows run for hours. Build complex spec driven development flows where every step calls the model again and again. It worked because a big part of the real inference cost was hidden behind subsidised subscriptions. Now this cost is becoming visible.

I already hear from friends in enterprise that token usage KPIs are starting to disappear from engineering goals. It makes sense.

When the bill becomes real, "use more AI" stops being a strategy. You need routing, smaller models, tighter context, and governance that matches the cost.

This will change how companies adopt AI. We will probably see less blind agent automation and more controlled AI architecture. Better routing between models. Smaller models for simple tasks. Frontier models only where they really matter. More attention to context size, caching, evaluation and cost governance.

I also think some very expensive development patterns will be challenged. Complex agent harnesses. Patterns like Ralph loops. Spec driven flows that generate thousands of tokens before any real code is written. They look impressive in demos, but they can become very expensive at enterprise scale.

AI will not disappear from software engineering. Quite the opposite. It will become normal infrastructure.

The next few months will be interesting. Many companies are about to find out how much their AI engineering workflows really cost when the subsidy layer is removed.