Why Your GitHub Copilot Bill Suddenly Exploded

On the first day of GitHub Copilot's new billing, a tech journalist checked his account and found Copilot projecting him a $180 bill — and he got off easy. Other developers watched a $10 plan balloon into thousands. Same subscription, same price. The meter did it.

Microsoft launched seven of its own AI models the same week, and the headlines all said Microsoft and OpenAI were breaking up. That read is mostly wrong. The real story is an accounting problem that breaks the whole business model of selling AI.

Normal software is almost pure profit — one more user costs almost nothing. AI breaks that: every answer burns real compute, so the more someone uses it, the more it costs the company running it. The Wall Street Journal reported in 2023 that Copilot was losing ~$20 a month per user — up to $80 on power users — on a $10 plan.

Then the smoking gun, 24 hours apart. June 1st: Copilot goes metered, the free fallback removed. June 2nd: Microsoft drops its own model into the picker — and its own rate card prices GPT-5.5 at $30 per million words of output versus $4.50 for Microsoft's model. About 6.7 times cheaper, a price Microsoft set for itself, so routing a heavy user to it collapses the cost on that user.

A startup already lived this arc. Cursor resold Anthropic's model at a reported negative 30% gross margin — a $200 plan against ~$5,000 of compute — then shipped its own model and turned positive. Same law that pushed Apple off Intel onto the M1.

But this isn't a divorce. Microsoft's first move against the trap wasn't a model — it was the meter, which moved the loss onto whoever pays the bill. And it kept OpenAI licensed through 2032. The honest counter holds too: $4.50 only proves what Microsoft charges itself, and the "AI margins are turning the corner" story is mostly compute margin — all-in it was still about 33%. So MAI is a bargaining chip, not an exit.

The next time a price jumps on a tool you depend on, ask whether it's really greed, or just the trap.

Deep Dive goes beneath the headline — AI, finance, security, and the systems running underneath.

RELATED EPISODES

How Anthropic Actually Makes Money — the same per-token margin math, from the lab''s side.

Will AI Replace Software Engineers? What the Data Actually Says — Cursor''s rise; here, its compute bill.

The Real Cost of AI: Who's Actually Paying for the AI Build-Out — there, your power bill; here, your Copilot bill.

The AI Chip War: Why Everyone's Watching the Wrong Fight — the same build-your-own logic, one layer down.

CHAPTERS

00:00 Quick financial disclaimer

00:08 The day-one bill shock

01:05 Not a breakup — an accounting problem

01:28 Why AI breaks the SaaS model

03:12 June 1st: Copilot becomes a meter

04:10 June 2nd: Microsoft ships its own model

05:14 Cursor already lived this trap

06:24 The law: a big buyer builds its own

08:08 Why it's still not a divorce

09:59 The honest ledger, and the calls

SOURCES

Copilot lost ~$20/user/mo on a $10 plan — AI has no SaaS economies of scale (WSJ 2023)

Inference price falling ~50x/yr while usage explodes; ~24x token growth by 2030 (Epoch AI; Goldman)

June 1 Copilot goes metered (free fallback removed); June 2 MAI-Code-1-Flash ships into the picker (GitHub)

Microsoft's rate card — GPT-5.5 $30 vs MAI $4.50 per 1M output tokens (~6.7x cheaper); 4.7M paid subscribers (GitHub)

Dev bills jumped $29→~$750 and $50→~$3,000 (TechCrunch)

Cursor ran ~-30% margin reselling Anthropic ($200 plan ≈ $5,000 compute), then shipped its own model (Newcomer)

OpenAI licensed through 2032, $38B revenue-share cap; GitHub VP says Copilot >$100M ARR; all-in margin ≈ 33% (Microsoft; The Information)

———

This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser. All investing involves risk. Consult a licensed financial professional before making investment decisions.