GitHub Copilot's pricing gamble - Human Who Codes

On June 1, GitHub switched the pricing model for Copilot from a request-based model to a usage-based model. This followed years of losing money1 on every Copilot plan in an effort to gain AI market share. Tech companies use this approach regularly, offering free or low-cost introductory rates to attract their user base before scaling back those offerings in favor of higher-priced services (often moving upmarket in the process). This is a well-worn path trodden by many companies who went on to make a lot of money on their products. It’s no wonder that GitHub, along with Anthropic and OpenAI, priced their products to attract the most users in the shortest period of time while losing money2. Most users didn’t realize that these prices were heavily subsidized and that those subsidies would eventually come to an end.

It appears that, after three years, GitHub decided that it was tired of losing money. (This change may or may not be related to the Microsoft-OpenAI divorce3, which means less revenue for Microsoft’s Azure service and potentially higher costs for OpenAI model access.) Instead of subsidizing Copilot users’ access to LLM APIs, Copilot now passes those costs directly to the user, functioning more like OpenRouter4. The result has been staggering.

Disclosure: As a GitHub Star5, I have received the Copilot Pro+ plan for free for the last two years and was privy to NDA-covered conversations around the pricing change.

What is usage-based pricing?

Usage-based pricing means that you’re charged based on your usage of the APIs rather than paying a fixed flat fee. For GitHub, this usage is allocated in dollars as AI credits, and each AI credit buys you a certain number of tokens for each model6. Users receive both a fixed allotment of credits based on their subscription and a “flex” allotment that may change at any time, giving GitHub the ability to throttle requests as needed.

The previous pricing model was based on the number of premium requests per month. This also had a fixed allotment per month, but each request could use as many tokens as it wanted without users incurring further charges.

I’ve used Copilot Pro+ for over two years and rarely used more than 40% of my allotted monthly requests in the previous pricing plan. That included use on open source projects, contract work, and private projects. During the first week of June on a very light work schedule, I used 33% of my allotted credits. It would easily have been 50% or more on a typical work week.

To understand the context of this problem, consider what would happen if Netflix followed a similar pricing scheme. Instead of paying a flat monthly fee and watching as many videos as you want, you’d pay based on the amount of transmitted data. That data is the product of the video resolution (determined by Netflix), the quality of the video (also determined by Netflix), and the length of the video. The five minutes of end credits? That counts toward your data even if you don’t watch it. The three-minute recap of the previous episode? That also counts. The ads Netflix inserts? You’re paying for that data too. Now imagine how many straight hours of Netflix you’d watch if you were presented with a bill at the end of each show. That’s where GitHub Copilot is right now.

The problem

GitHub always knew that subsidies were going to come to an end. However, two factors contributed to this decision:

Copilot’s agentic evolution. The way people use Copilot (and AI in general) has changed dramatically in the past three years.

Pushing more agent use. GitHub itself has continued to implement more agentic user experiences.

Both factors contribute to the growing frustration of GitHub users who feel like there’s been a bit of a bait-and-switch here.

Copilot’s agentic evolution

When Copilot first launched, it was essentially a better (and often worse) version of autocomplete. The “ghost text” would appear as you typed and try to guess what code you wanted to write. Autocomplete gave way to the sidebar conversation where you could directly interact with Copilot and ask it to complete different tasks. In hindsight, it’s easy to see that things would evolve beyond that point, but in 2023, no one really knew what to expect from AI.

Fast-forward to 2026 and Copilot can run on its own, read relevant documentation, fetch resources from the web to learn more, and implement complex features that take hours to complete. This move from a “copilot” editor to a fully-autonomous agent meant Copilot consumed and produced ever-larger amounts of tokens on an ongoing basis, 10-100x more tokens according to some estimates7. Further, Visual Studio Code continued adding features to allow multiple agents to operate simultaneously in a single project. Cost-per-token is decreasing, but not at a rate that can offset the increased token usage.

We started with Copilot as something you’d work with interactively on a few files and...