AI Subscriptions Are a Ticking Time Bomb. GitHub and Anthropic Both Shift to Usage-Based Billing

Your AI coding budget just got a lot harder to predict.

A viral analysis on Hacker News (413 points, 396 comments) makes the case that every major AI lab has been running a loss-leader program, and the correction is starting. Two concrete dates matter. GitHub transitions all Copilot plans to token-based AI Credits on June 1: Pro ($10/mo) gets $15 in combined credits, Pro+ ($39/mo) gets $70, and a new Max tier ($100/mo) gets $200. Anthropic splits Claude subscriptions into usage pools effective June 15. Code completions and next-edit suggestions stay unlimited and free, but everything agentic now has a meter running.

The catalyzing number: Uber reportedly burned through its entire 2026 AI budget by April. Let that sink in. A company with massive engineering resources couldn't forecast agentic compute costs accurately enough to make their annual budget last four months.

This isn't surprising if you've been running agentic workloads. A chat completion is a few thousand tokens. An agent loop that reads files, writes code, runs tests, reads errors, and iterates can burn 100K+ tokens per task. I've watched single Claude Code sessions in my personal projects consume what used to be a day's worth of API credits in 20 minutes. Flat-rate pricing was never going to survive that math.

Windsurf is making the same move. They bumped Pro from $15 to $20/month, added a $200/month Max tier, and shifted Bugbot to usage-based billing effective June 8. The bundling of Devin Cloud softens the price increase, but the direction is unmistakable. Every tool is migrating from "all you can eat" to "pay for what you consume."

The builder move: audit your agentic workload costs this week, before the switchover dates. Measure actual token consumption per task type. Budget for 3-5x what chat-based usage costs. And look seriously at the custom model story below, because the cost pressure is exactly why platforms are training their own models.