Is the AI buffet gone? The cost of GitHub Copilot is too high to bear, starting from June 1st, it will be billed based on usage.

Unable to bear the soaring costs of computing power, GitHub Copilot announced that starting June 1 it will cancel its “unlimited” mode and switch to token-based usage billing. The move has triggered strong backlash from developers and a wave of subscription cancellations, highlighting the infrastructure challenges the AI industry is facing.

The era of “AI unlimited” is over—GitHub Copilot changes its billing model

Previously praised as “AI all-you-can-eat” in the industry, GitHub Copilot is now shifting from an all-you-can-eat buffet to an ordinary restaurant.

According to a report by The Register, Microsoft-owned GitHub acknowledged in its latest announcement that because it can no longer keep absorbing losses, it has decided that starting June 1, 2026, it will change its billing method from request-based to usage-based.

Under the original model, subscribers could submit a fixed number of advanced requests, but it did not account for the complexity of tasks. As a result, the cost of prompts that require large amounts of computation far exceeds the subscription revenue received.

As GitHub product lead Mario Rodriguez revealed, simple chats and AI coding tasks that last for several hours may cost exactly the same. The company has absorbed continuously rising inference costs, making the current model difficult to operate sustainably.

Image source: GitHub Copilot announcement—The end of the AI all-you-can-eat era, GitHub Copilot changes its billing model

Starting June 1, GitHub Copilot introduces virtual billing units

In the past, GitHub Copilot offered unlimited AI assistance for a fixed monthly fee, earning it the nickname “AI all-you-can-eat.” Compared with mainstream alternatives such as Codex, Cursor, and Claude Code, it has long been a low-profile but high value option for engineers.

However, once it shifts to usage-based billing, GitHub Copilot’s charges will be directly tied to tokens.

Because different models have different pricing rates, GitHub has created a virtual unit called GitHub AI Credits, equivalent to $0.01 USD. Microsoft will convert users’ consumed input, output, and cached tokens into credit costs based on the listed API fee rates.

Rodriguez said that future GitHub Copilot subscription plans will include a fixed monthly amount of AI credits, and paid plans can also be supplemented with additional add-ons.

Since usage-based billing involves uncertainty, users cannot confirm in advance how many tokens a specific input will consume, and calculating this across other tools is even more complicated. Therefore, GitHub plans to launch a billing preview feature in early May, so users can assess spending before the June transition period.

Reddit community pushes back—users vow to cancel

Unsurprisingly, the change in GitHub Copilot’s pricing has sparked widespread backlash on Reddit forums.

One user replied that if billing is based on usage, services like OpenRouter—which don’t require subscriptions—already exist in the market. The new system effectively makes users pay the full API list price, completely removing the value of subscribing.

Many annual subscribers feel their rights are being harmed. They estimate that the cost for certain models could jump by dozens of times, and many have vowed to demand refunds and cancel their subscriptions.

The community has also seen increased calls to switch platforms. Many developers said they will use tools like Claude Code or Cursor instead, and some even plan to upgrade hardware to run open-source models such as Alibaba’s Qwen 3.6 27B locally.

Image source: Reddit—The changes in GitHub Copilot’s pricing sparked a strong backlash on Reddit

OpenClaw sparks a new wave—AI infrastructure overload

GitHub Copilot’s changes reflect the broader infrastructure challenges facing the AI industry.

This year’s February saw a surge of attention for an open-source AI assistant nicknamed “Lobster”—OpenClaw. It led many developers to experiment with AI agents running 24/7 all day long to handle various tasks, and improvements in model capability have also encouraged more people to explore AI coding.

As a result, AI companies that previously offered subscription subsidies are now facing a huge demand that far exceeds the capacity their inference infrastructure can handle. Even giants such as Anthropic and OpenAI have encountered capacity problems. Claude Code only recently fixed a major bug that improved output quality and reduced issues with poor output and latency (commonly known as “getting dumber”/“de-intelligence”), and it also reset users’ usage quotas.

Until the industry finds a way to balance costs and user experience, the resource consumption driven by massive AI compute demand—and its “price correction effect”—will continue to spread throughout the AI industry.

Further reading:
Claude Code really got dumber! The official admits there are 3 major bugs, and all users’ subscription quotas have been reset

Lawmakers propose banning data centers—environmental groups criticize an ecological disaster! And the First Lady is walking into the White House with an AI robot