GitHub's AI Bill Shock and Claude's Hijacking Problem

Some GitHub Copilot users discovered this week that their entire month of AI credits was gone before lunch on day one. That's not an exaggeration — it's what happens when a platform quietly absorbs ballooning AI costs for years and then suddenly hands the bill to the people who were happily clicking away.

Here's the backstory. GitHub spent months running a system where users got a fixed number of "requests" per month, regardless of how computationally heavy those requests actually were. A quick autocomplete and a three-hour autonomous coding marathon counted roughly the same. That was generous, maybe unsustainably so, and GitHub has now pulled the curtain back with a usage-based pricing model that went live this week.

The new structure ties costs to actual token consumption — the raw input and output flowing through whichever AI model you're using. Paid plans now come with monthly "credits," where one credit equals one cent of usage. The entry-level $10 Pro plan includes 1,500 credits. The $39 Pro+ plan bumps that to 7,000. And the $100 Copilot Max plan gives you 20,000. Sounds reasonable until you realize that users sharing their old usage stats are finding that their previous habits would have cost them thousands of dollars under the new math.

The model choice matters enormously here, and that's where things get genuinely tricky. Running a million output tokens through a lightweight model like GPT-5.4 nano costs around $1.25. Doing the same thing through GPT-5.5, a frontier model, runs you $30. The gap is staggering, and the problem is that Copilot's "Auto" mode — which many users rely on to just pick the right tool for the job — can quietly escalate to expensive models even for straightforward tasks. Users who aren't watching their model selection closely are essentially handing the keys to an algorithm that has no particular interest in keeping their bill small.

Early reports from users paint a picture of costs that spiral fast. One complex prompt reportedly consumed 171 credits on its own. Another user burned 700 credits across just a few prompts. Ars Technica ran a test — building a basic Minesweeper game via Claude Haiku 4.5 — and clocked around 94 credits. That's manageable for a toy project. Scale that up to real-world work involving large codebases, multi-step reviews, or iterative debugging sessions, and you can see how a developer could genuinely exhaust their monthly allotment in a single focused afternoon.

What makes this sting a little more is the transparency gap. GitHub provided a tool to estimate what past usage would cost under the new model, and some users are clearly only now realizing what they were getting away with before. The old system functionally subsidized heavy users, and that subsidy is over.

The deeper issue is that this isn't really a GitHub problem — it's an industry-wide reckoning. AI inference costs are real, they're significant, and every platform that offered flat-rate AI access was eventually going to hit this wall. GitHub just hit it loudly, in public, with a very unhappy user base watching in real time. Developers who want to keep using Copilot without budget anxiety are going to need to get a lot more intentional about which models they reach for and when.