Beginning June 1, 2026, GitHub Copilot will transition from its current flat-rate subscription model to a new token-based billing system. This significant change aligns Copilot's pricing structure with the common API charge models used by large language models (LLMs), which are prevalent in business plans.

The previous subscription model was straightforward: users received a predetermined number of 'Premium Requests' based on their tier. Whether a complex, multi-hour coding task or a simple query, each interaction consumed one premium request. The upcoming token-based scheme, however, will measure most requests according to the tokens utilized by, input into, and output from the underlying LLM powering Copilot.

A 'token' is generally estimated to represent roughly three-quarters of a word. Consequently, providing an LLM with a 10,000-word text for analysis would amount to 12,000-13,000 tokens of content. In the context of development, if Copilot were to analyze a body of code comprising 10,000 'words' (encompassing expressions, statements, and variable names) in a single query, this would deplete 12,000-13,000 tokens from a user's monthly allotment. Both prompt text (inputs) and Copilot's generated outputs will contribute to the token count.

While the pricing tiers themselves will remain at their existing levels, users will now be allocated 'AI Credits' equivalent to their subscription value, replacing the previous system of a fixed number of queries per month. For instance, a base-tier Copilot Pro subscriber paying $10 per month will receive 1,000 credits. GitHub has indicated that currently, one AI Credit is equivalent to one US cent. The actual number of tokens each credit can purchase will vary, influenced by factors such as the specific model used, the input/output mix, the size of the LLM's memory cache for context, and the particular feature requested.

This new model means that developers primarily engaging in simple queries may find they do not need to acquire additional credits each month. Conversely, complex, multi-agent queries involving extensive codebases will consume AI Credits at a much faster rate. Furthermore, queries directed to the most advanced 'frontier models' will incur higher costs compared to those sent to less powerful alternatives.

GitHub's pricing adjustments do include some benefits to users, notably that code completions (similar to smartphone auto-complete functions) and Next Edit suggestions will continue to be provided free of charge. These changes are reflective of broader industry shifts, with companies like Anthropic and OpenAI having already transitioned their enterprise customers to token-based billing systems.

Unlike Anthropic and OpenAI, Microsoft, the owner of GitHub, is a broadly profitable entity capable of subsidizing GitHub Copilot through revenues from its other divisions, such as software and cloud services. Prior to this change, users could effectively utilize three to eight times the token volume covered by their monthly subscription costs without incurring penalties. Microsoft's strategic shift will compel new and existing Copilot users to become acutely aware of their token expenditure per query – a metric that was previously abstracted away by fixed monthly subscriptions. This billing model, while potentially more economically sound for Microsoft, might inadvertently discourage the initial exploration and testing crucial for new users.

For businesses that integrate AI coding agents into their development teams, the cost implications of this industry-wide shift in pricing policies are substantial. Uber, as reported by The Information, saw its CTO state that the company had already expended its entire 2026 AI budget within the current year, noting that 11% of Uber's code updates are now generated by AI, primarily using Anthropic's Claude agents. Beyond the IT department, companies deploying AI automation should recognize that complex tasks, especially those involving unsupervised agentic LLMs operating over extended periods, could soon be subject to similar per-token charges. Consequently, the anticipated efficiency gains from AI in the workforce will necessitate careful measurement against the potential rise in AI vendor bills.