Copilot's surprise bill shows what AI agents really cost

GitHub Copilot is an AI tool that helps people write code. On June 1, 2026, it changed how it charges. Before, you paid one flat price each month. Now you pay for how much AI you use. This matters because AI "agents" — helpers that do whole jobs on their own — can eat up a lot of AI in minutes.

Here are the numbers. Copilot Pro now gives you $10 of AI credits each month. Copilot Pro+ gives you $39. Plain autocomplete, where it finishes your line of code, still costs nothing extra. But the bigger agent jobs burn credits fast. Some developers say their bills jumped from $29 to $750, and from $50 to $3,000 a month. GitHub said its agent tools simply cost far more computing power to run.

This is not just about Copilot. Every AI tool that runs agents has the same problem. An agent does not ask the AI once and stop. It loops: it tries, checks, fixes, and sometimes calls in smaller helper agents. Every step uses more tokens — the little chunks of text an AI charges you for. Those steps add up fast.

So the first move is to check how your team really uses Copilot. If you mostly use autocomplete, little will change. If you let agents roam across many files, the bill climbs quickly. A clear, narrow task costs much less than a vague one that sends the agent wandering.

The advice is simple. Set a spending limit. Say clearly what each task should do. Watch your credits like you watch any other bill. Cheap, unlimited AI coding is ending. What comes next is just honest about the cost.