The Free Lunch Is Over: GitHub Copilot Freezes Signups, Leaked Docs Show Token Metering by June

GitHub pulled the plug on new individual signups for Copilot Pro, Pro+, and Student plans on April 20. No warning. No timeline for reopening. The reason is simple math: weekly operating costs doubled since January as users leaned harder into agentic features that burn Opus-class tokens by the fistful.

But the signup freeze is just the appetizer. The Register obtained leaked internal documents showing a June 1 migration to token-based billing. Business tier moves to $19/month plus $30 in pooled credits. Enterprise jumps to $39/month plus $70. Opus models get pulled from Pro entirely. A single 50K-token agentic session on Opus 4.7 costs roughly $1.50, which means that $30 credit budget covers about 20 sessions per month. For anyone using Copilot as their primary coding agent, that's maybe one good afternoon of work.

I've been expecting this. Unlimited flat-rate pricing for AI coding tools was always a subsidy play to capture market share, and subsidies end. What I didn't expect was how fast the economics turned. GitHub went from "everyone gets unlimited AI" to "we can't even onboard new users" in under four months. The agentic workflow pattern, where models loop through tool calls and reasoning for minutes at a time, broke the unit economics that were already tight with simple autocomplete.

Here's what this means for builders right now. If you're locked into Copilot workflows, you've got about five weeks to audit your actual token consumption. Most developers have no idea how many tokens their agentic sessions burn because they've never had to care. Start caring. Run a week of normal work and track it.

More importantly, this is the moment to evaluate alternatives seriously. DeepSeek V4-Flash at $0.14 per million input tokens (story #2 below) changes the calculus entirely. Local models like Qwen3.6-27B running at 100 tokens/sec on consumer hardware change it even more. The open-weight escape hatch isn't theoretical anymore. It's the rational economic choice for a growing set of use cases.

The broader pattern: every AI tool vendor offering unlimited usage is running the same math GitHub just failed. Expect Cursor, Windsurf, and others to follow with usage-based pricing within 6-12 months. Build your workflows to be model-portable now, not after you hit a billing wall.