Copilot ends the illusion of unlimited AI: coding agents need a budget

On April 27, 2026 GitHub announced that Copilot is moving to usage-based billing for premium AI requests. The useful framing is simple: this is the end of the illusion of unlimited AI.

Not because Copilot stops making sense. Quite the opposite. AI coding is becoming real work, and real work consumes compute.

What changes

Premium requests beyond included allowances will be charged based on usage. That matters most for agentic work. Autocomplete is a short interaction. An agent that reads a repository, plans a change, runs tests and iterates is a different cost category.

Why this is healthy

In API automation we already know this: sending every step to the strongest model is expensive. Coding assistants hid some of that behind monthly seat pricing. But once the assistant starts acting like an agent, the economics change.

More context means more tokens. Longer tasks mean more iterations. Stronger models should be reserved for harder tasks. Badly specified work can burn budget without producing value.

What I would set in a company

I would not ban Copilot, Claude Code, Codex, or similar tools. I would manage them operationally: team or project budgets, rules for premium model use, reporting by repository, review for expensive agent runs, and clear policies for production code, secrets, and data.

Conclusion

Usage-based billing is not the end of AI coding. It is the end of pretending AI coding has no operational cost. Companies that want to benefit from agents need the same discipline they use for cloud services: limits, metrics, reporting, and responsibility.

---

Sources: GitHub Blog, GitHub Docs, AI ta Krajta, Jirka Herník.