What GitHub’s New Token-Based Billing Actually Changes
GitHub Copilot’s token-based pricing is a consumption model that bills developers for the volume of AI-generated tokens their coding assistant processes, instead of charging a flat subscription tied to user seats or request counts. Under the new system, GitHub Copilot pricing tracks how much work the AI performs: long code completions, multi-file edits, and detailed explanations all translate into more tokens and higher AI coding assistant costs. Reports collected by Ars Technica and PCMag show that many users who were comfortable under a fixed monthly plan are now seeing credit balances melt away after only a day or two of normal coding. This marks a decisive move away from the earlier, heavily subsidized subscriptions that hid the real cost of large language models behind a single predictable fee.
From Flat Subscription to Per-Token Charges and 10x Bills
The switch from fixed subscriptions to per-token charges is having an immediate and painful effect on some developers’ wallets. According to PCMag, “one user previously used $39 (approx. RM180) a month, but could now be expected to receive a bill for almost $1,800 (approx. RM8,280) a month,” a more than 40x jump under the new GitHub Copilot pricing. Other Copilot users on X and GitHub’s community forum report burning through half or more of their monthly token allowance in a single workday. In GitHub’s community discussion, one user said their 12% credit block disappeared on “very minor” edits amounting to a few lines changed across six files. These anecdotes show how quickly AI coding assistant costs can spike once every token of output is tallied instead of being absorbed by a flat monthly subscription.
Why AI Credits Are Disappearing Faster Than Developers Expect
The new token-based billing surprises many users because they were conditioned by years of flat-fee access to powerful models. Under the old system, Copilot’s subscription plans effectively allowed developers to burn far more tokens than the subscription value covered. Source reports argue these plans acted as loss leaders, with Microsoft subsidizing usage to grow adoption. Now, the per-token model exposes the true compute cost of running large language models at scale, including development, training, infrastructure, and maintenance. That is why small-looking tasks can carry a hidden price: multi-file context, long responses, and frequent refinements each consume substantial tokens. Users writing feedback on GitHub’s forum say their “AI credits burned like anything” on tasks they assumed were cheap, revealing a gap between perceived light usage and the token-heavy reality of modern AI coding assistants.
How Per-Token Billing Will Reshape Developer Workflows
This shift is already changing how developers think about using AI. Some Copilot customers tell PCMag and Ars Technica they plan to move to alternative tools, while others intend to use AI in a “very focused” way, reserving tokens for complex or high-value coding. Instead of streaming suggestions continuously, teams may restrict AI coding assistant costs by shortening prompts, disabling background completions, or only enabling Copilot during specific tasks such as refactoring or documentation. The psychological effect is also significant: seeing a visible token budget encourages budgeting behavior similar to cloud compute credits. Over time, this could slow the casual, always-on use of Copilot and favor workflows where developers batch questions or code changes to minimize per-token charges.
A Broader AI Industry Shift Toward Consumption Pricing
GitHub Copilot’s pricing overhaul is part of a wider move across the AI industry toward consumption-based token billing. PCMag notes that Anthropic shifted Claude Enterprise to token-based billing in April, and Microsoft has now done the same for Copilot, signaling the end of what many users experienced as the “wild west” of cheap frontier models. For AI vendors, this model aligns revenue with actual resource usage and reflects the high ongoing cost of running large models. For customers, it replaces predictable subscriptions with metered utility-style charges, pushing organizations to measure and justify AI usage more carefully. As AI coding assistant costs become more transparent and variable, procurement teams and engineering leaders will likely weigh benefits against per-token charges, influencing which tools survive and how deeply AI is integrated into everyday development.
