The End of the All You Can Eat AI Buffet
GitHub is finally pulling the plug on the fantasy that everyone could gorge on AI computing power for a flat fee. Starting June 1, Copilot is ditching its subscription model for a pay per token scheme, a move that reveals just how unsustainable the AI arms race has become. The company admitted it can no longer absorb the escalating inference costs, a polite way of saying that the math on offering unlimited premium requests was always a house of cards.
The new system introduces AI Credits, a meter that ticks based on token consumption for every input, output, and cached request. Cheap models like GPT-5.4 Mini will cost $4.50 per million output tokens, while the heavy hitters like GPT-5.5 will run you $30 per million. This is a direct admission that a simple chat and a multi-hour autonomous coding session are not the same product, and they should not carry the same price tag.
The Real Reason: Agentic Coding and Empty Pockets
This isn’t just about fairness. It is about survival. Leaked internal documents cited by critic Ed Zitron show that Copilot’s week-over-week costs nearly doubled since January, a spike directly caused by agentic AI assistants like Openclaw. These always on workflows burn through tokens like there is no tomorrow, and GitHub was footing the bill. The company has paused new signups, pulled Claude Opus from lower tier plans, and is now flashing the bill to its heaviest users.
This move is predatory in its timing. GitHub and Anthropic are both signaling that the era of subsidized AI is over. Anthropic is already charging enterprise clients for full compute costs and restricting Claude Code during peak hours. The message is clear: if you want AI to do real work, you will pay real money. The days of a $20 monthly subscription giving you a backdoor to infinite compute are over. The meter is running, and it is ticking fast.
Source: Arstechnica
