The End of the Free Lunch
GitHub is finally admitting what everyone in AI infrastructure already knew: subsidizing unlimited coding queries on expensive models is a money pit. Starting June 1, Copilot subscribers will pay per AI credit, a shift from the flat fee model that let heavy users feast on inference compute like it was an all you can eat buffet. GitHub claims this aligns pricing with actual usage, but let’s call this what it is. A desperate attempt to staunch the bleeding after internal documents leaked to critic Ed Zitron showed weekly Copilot costs nearly doubling since January. The rise of always on agentic assistants like Openclaw has turned GitHub’s generous token allocations into a fiscal nightmare.
The Fine Print Will Bite You
Under the new scheme, your monthly subscription buys a pool of AI credits, with additional usage charged at per token API rates that vary wildly by model. Premium thinking with GPT 5.5 will cost $30 per million output tokens, while cheaper models like GPT 5.4 Mini run at $4.50. GitHub will still let you use basic code completion without bleeding credits, but code reviews now burn Actions minutes. This follows Anthropic’s move to charge enterprise Claude users for actual compute costs and its brief test of removing Claude Code from the Pro plan. The entire industry is realizing that the current subscription model is a Ponzi scheme for compute time, and the free ride for power users is over.
CVE 2026 12345 and the AI Profit Paradox
These pricing moves expose a fundamental tension AI companies face: translating massive demand into actual profitability. Despite record revenue, giants like Microsoft, Anthropic, and OpenAI are still chasing profits while grappling with compute shortages. The days of flat fee AI access are dying. As GitHub tightens usage limits and removes Opus models from lower tiers, users should expect more surprise bills. For anyone relying on agentic workflows that burn tokens like a drag racer burns fuel, start budgeting now because the AI sugar daddy is cutting you off.
Source: Arstechnica
