I Burned Through 81% of My AI Credits in One Day After GitHub Copilot's June 1 Overhaul

What Changed

On June 1, 2026, GitHub fully migrated Copilot's billing unit from "Premium Requests (PRU)" to "GitHub AI Credits." The monthly fee stayed the same, but the consumption model changed from the ground up. This article is based on my experience with GitHub Copilot Pro ($10/month). The Pro plan comes with 1,500 AI Credits per month.

Under the old system, GitHub absorbed the heavy token consumption of agent sessions as "a single request." That was no longer sustainable, and the essence of this change is that real costs have now been passed on to users.

How It Feels Before and After the Overhaul

When I look at my own usage patterns, the difference is stark.

That means 81% of my monthly credit allocation was gone on day one. The gap is extreme. It is said that the impact is small if code completion is your main use case, but when agent mode is your primary workflow, costs spike this dramatically.

GPT-5 Mini Is No Longer Free Either

One easily overlooked aspect of this change is that GPT-5 mini, which was previously free to use, has also become a paid feature under the AI Credits system.

Under the old PRU system, lightweight models like GPT-5 mini did not consume Premium Requests. Under the new system, all usage is billed on a token-consumption basis, and GPT-5 mini is no exception. The fact that even the lightweight models I used casually now eat into my credits is part of why they are depleting faster than expected.

Why I Did Not Choose Claude Code

There is a comparison with Claude Code behind my continued use of GitHub Copilot. When I tried Claude Code previously, I ran into the constraint of a forced stop after 5 hours. If you are working in long sessions, that interruption is quite inconvenient. Based on that experience, I had been choosing GitHub Copilot for its greater continuity.

However, this overhaul has eliminated that reason for my choice.

Current Status of Each Tool

Summary

GitHub Copilot's June 1 change was an asymmetric revision — a small impact for light users, but a significant cost increase for heavy users of agent mode. For users like me where agent usage is a given, the gap from the old system was simply too large.

Going forward, I plan to try ChatGPT Codex and Gemini Code Assist and compare their costs and usability. I intend to write another article on the results.