Why GitHub Copilot's Billing Model is Fundamentally Broken

The Flaw of Message-Based Billing

Modern AI development has shifted from simple chat interfaces to agentic workflows where a single user prompt can trigger dozens of recursive tool calls. Theo argues that billing these interactions as 'messages' is a fundamental error. In a traditional chat, one message equals one response. In an agentic loop, a single message can trigger multiple search queries, file reads, and reasoning steps, each consuming significant GPU compute. Because the cost delta between a simple query and a complex, multi-step agentic task can range from pennies to hundreds of dollars, fixed-price subscriptions are economically unsustainable for providers and prone to abuse.

The Economics of Inference

There are four primary ways to bill for AI inference: subscriptions with rate limits, subscriptions with message limits, subscriptions with spend limits, and dedicated compute. Theo highlights that 'message limits' are the most dangerous for small businesses because they fail to account for token variance. He shares his experience with T3 Chat, where a small percentage of users—often using tools like Repix to compress entire codebases into a single prompt—were costing the business hundreds of dollars in inference fees while paying only a nominal subscription price. This forced a pivot to stricter limits to prevent bankruptcy.

The Copilot Exploitation

GitHub Copilot, backed by Microsoft's massive capital, has historically subsidized these costs. However, as Copilot evolves into a full agentic solution, the cost per user has skyrocketed. Theo demonstrates that by running complex cryptography challenges that force the model into long-running reasoning loops, a single 'message' can cost over $60 in raw inference. By leveraging these agentic capabilities, a user on a $40/month plan could theoretically consume nearly $100,000 worth of compute in a single month. The current billing model fails to account for this because it treats all 'messages' as equal units, ignoring the underlying token consumption and GPU time required for agentic reasoning.

The Trade-off of Transparency

Platforms like Cursor or OpenRouter offer more transparent spend-based models, but they lack the 'all-you-can-eat' psychological comfort that users demand. The tension lies between providing a predictable user experience and protecting the business from the 'long tail' of power users who can bankrupt a platform through strategic, high-compute usage. Theo concludes that as models become more capable, the industry must move toward token-based or usage-based billing, as the 'message' has ceased to be a meaningful unit of value.