A developer is deep in a complex refactoring session, commanding an AI agent to modify ten different files simultaneously to ensure architectural consistency. The workflow is seamless until the AI suddenly stops responding mid-sentence. When the developer attempts to switch to a more powerful premium model to resolve the hang, the selection menu is greyed out. A notification appears at the bottom of the editor, warning of a usage limit. This is not a transient server glitch or a temporary API outage, but a systemic throttle triggered because the user's token consumption has crossed a critical threshold.

The Mechanics of the New Copilot Constraints

GitHub Copilot is fundamentally altering its access and consumption model to protect service stability for its existing user base. The company has officially suspended new sign-ups for the individual plan and is implementing a more aggressive set of usage caps. These restrictions are split into two distinct categories: session limits and weekly limits. Session limits act as a short-term governor to prevent server overload during peak traffic hours, while weekly limits impose a hard cap on the total number of tokens a user can consume over a seven-day rolling window.

To accommodate power users, GitHub is introducing a tiered differentiation between the Pro and Pro+ plans. Users who upgrade to the Pro+ tier receive a weekly token limit that is more than five times higher than that of the standard Pro plan. The system handles limit exhaustion through a tiered fallback mechanism. If a user hits their weekly cap but still possesses premium request credits, the service continues to function via an automatic model selection feature. However, the ability to manually select specific high-end models is disabled until the weekly cycle resets.

Transparency regarding these limits is integrated directly into the developer's environment. Users can monitor their remaining token allowance in real-time through VS Code and the Copilot CLI. For those who find these new restrictions unacceptable, GitHub has provided a window for exit. Users may cancel their subscriptions through their billing settings and request a prorated refund for the remaining period, provided the request is made by May 20.

The Economic Collision of Agentic Workflows

This policy shift is not a reaction to a simple increase in user count, but rather a response to a fundamental change in how developers interact with AI. The industry is moving from a request-response paradigm to agentic workflows. In the original Copilot model, a user asked a question and the AI provided a snippet of code, creating a short, linear token trajectory. Modern agentic features, however, allow the AI to plan, execute, and verify tasks across multiple files and iterations autonomously. These agents maintain parallel sessions and run for extended periods, consuming vast amounts of compute resources compared to a single chat prompt.

This evolution has created a structural cost imbalance. GitHub has observed that a small fraction of users performing agentic tasks can consume enough tokens to make their subscription cost-negative for the provider. The cost of processing these long-trajectory requests frequently exceeds the flat monthly fee paid by the user. By separating premium request permissions from overall usage limits, GitHub is creating a dual-layer guardrail. While the premium permission determines which model the user can access, the usage limit controls the raw volume of tokens to ensure the underlying infrastructure does not collapse under the weight of autonomous loops.

This transition reveals a broader truth about the current state of generative AI. The era of the unlimited flat-rate subscription is becoming unsustainable as AI tools evolve from passive assistants into active agents. By segregating heavy users into the Pro+ tier, GitHub is effectively moving away from a simple subscription model toward a resource-consumption model. The cost of intelligence is no longer a fixed overhead but a variable expense tied directly to the complexity of the autonomous work being performed.

AI subscription fees are evolving from fixed monthly costs into variable expenses that scale with the computing power required by the user's agents.