GitHub's transition from request-based billing to usage-based pricing for Copilot went live today, and developers are not handling the news well. Across social media and coding forums, users are sharing screenshots of their credit meter plummeting after just hours of what they'd consider normal AI-assisted workβ€”a stark contrast to the predictable monthly bills they knew before.

The New Pricing Structure

Under GitHub's updated model, Copilot subscribers receive monthly credits worth $0.01 each rather than a fixed number of requests. The Pro tier at $10/month grants 1,500 credits ($15 value), Pro+ at $39/month provides 7,000 credits ($70 value), and the top-tier Max plan for $100/month offers 20,000 credits ($200 value). But here's where it gets expensive: credit consumption scales directly with token usage and model selection. One million output tokens from OpenAI's GPT-5.4 nano runs just $1.25, while that same volume from the flagship GPT-5.5 model costs $30β€”24 times more.

The Token Tax Adds Up Fast

Ars Technica's testing revealed that even straightforward prompts carry surprising weight. A simple "build a Minesweeper game" request through Claude Haiku 4.5 consumed approximately 94 credits. But users report far steeper consumption: single complex prompts burning 171 credits, "a few prompts" totaling 700 credits, and alarmingly, just two Copilot-assisted commits using 5,000 credits. One developer shared that testing Claude Sonnet 4.6 for a single day with "super cautious" usage still consumed 840 credits without any particularly demanding work.

Why Long Chats Now Carry Hidden Costs

The shift punishes developers who maintain continuous chat sessionsβ€”a common workflow for complex debugging or refactoring. As coder Neil Hewitt noted on Bluesky, continuing a three-day-old conversation means sending the entire chat history as context each time: "input tokens use credits... it's not rocket science." Developers accustomed to keeping Copilot conversations alive for context now face an entirely different economic calculus around maintaining those sessions.

Users Threaten Exodus While Others Adapt

Many developers are publicly declaring intentions to cancel, with some already exploring alternatives. Reddit users discuss integrating Deepseek into their VSCode environments at roughly 7 cents per 15 million tokensβ€”a fraction of Copilot's rates. However, not everyone is hemorrhaging credits. Developer Henri Kinnunen managed a "productive day" on just 161 credits by staying "very focused and deliberate with AI," suggesting that disciplined usage patterns can stretch credit allotments significantly further than the average developer might expect.

Key Takeaways

  • Credit consumption depends heavily on model selection: GPT-5.4 nano costs 24x less per token than GPT-5.5
  • Input context (chat history, file contents) directly impacts monthly spending
  • Users report burning through thousands of credits daily under normal usage patterns
  • Strategic session management and deliberate prompting can dramatically reduce consumption

The Bottom Line

GitHub's move exposes just how heavily the company was subsidizing power users under flat-rate pricingβ€”and that subsidy is over now. Whether this usage-based model becomes industry standard or drives developers toward more economical alternatives will depend on whether Copilot's integration advantages justify what amounts to metered billing for a tool many have come to rely on constantly.