GitHub Copilot's Metered Billing Delivers Its First Bill Shock

GitHub moved every Copilot plan to usage-based billing on June 1. This week the first full month of invoices came due, and the reaction from developers has been loud.

Under the old model, a flat monthly fee covered a set number of requests. The new system meters actual token consumption, counting input, output and cached tokens against each model's published API rate. Every plan now includes a monthly allotment of GitHub AI Credits, priced at one cent each, with the option to buy more once the allowance runs out. GitHub laid out the change on its company blog in May.

For anyone typing the occasional autocomplete, the shift barely registers. The pain lands on developers who spent June running Copilot's autonomous coding agents. When an agent takes on a task, the model itself decides how much context to load and how many internal steps to run, so a single instruction can quietly trigger dozens of billable calls. GitHub's own research, published in May, found that agentic coding can burn through roughly a thousand times more tokens than a simple one-shot query.

The result was sticker shock. TechCrunch reported developers projecting jumps from $29 a month to $750, and from $50 to $3,000. The Register documented users vowing to abandon the product. Some had left agents running around the clock in June, assuming the flat rate still applied, and only learned the true cost when the meter caught up.

The end of the flat rate

The episode fits a pattern we have tracked. The industry's brief era of all-you-can-eat AI pricing, what some called tokenmaxxing, is closing as providers push consumption-based models that pass compute costs straight to the user. It also rhymes with the startup employee who accidentally ran up an $81,000 bill letting an agent loose. Autonomous tools are powerful precisely because they do a great deal of work without supervision, which turns out to be the same reason the invoice can surprise you.

GitHub argues the new pricing reflects real costs more honestly than a subsidized flat fee ever could, and that the heaviest users were always a small minority. That is fair as far as it goes. The harder problem is visibility. Developers want a meter they can watch while they work, not a total that arrives after the tokens are already spent.