GitHub Copilot went token-based. Your $29 bill just became unpredictable.

On June 1, 2026, GitHub flipped Copilot from a flat $29-a-month subscription to usage-based, per-token billing. Within days, developers running heavy agentic workloads were posting screenshots of bills past $750 — and the reaction was immediate and loud.

The sticker shock is the headline. It isn't the story. The story is quieter and bigger: your AI coding cost just stopped being a number you can plan around.

Flat rate was never going to survive agents

A $29 seat made sense when AI coding meant autocomplete. It cannot survive agents. Agentic coding is structurally token-hungry — every turn re-sends a growing context window plus tool output: files read, diffs, test logs. A developer who lets agents run all day moves a startling number of tokens, and someone was always going to bill for them honestly.

So token pricing isn't a betrayal. It's the real cost arriving. The problem is the form it arrives in: not a plan, but a surprise — a line item that used to be $29 × seats and is now a function of how hard your agents happened to run this month.

$750 feels outrageous because you can't see what earned it

Here's the trap. A per-token bill tells you the total and nothing else. Was your $750 the migration an agent shipped that saved three weeks — or an agent that looped on a dead end overnight? The invoice can't tell them apart. Per-seat hid the spend; per-token exposes it but still won't attribute it.

The instinct now will be to cap it, throttle it, or flee back to whatever flat-rate tool is left. All three are the same move: managing a number you can't see by making it smaller.

We wrote about the cap version of this when Uber put a $1,500 ceiling on AI coding per developer. Copilot's pricing change is the same structural problem from the other direction — and this time it landed on every team at once.

The fix is the session, not the plan

You don't govern variable spend by wishing the flat plan back. You govern it with attribution — and the right unit for AI work is the session: one task, one transcript, with its model, its compute, and its cost bound together. That's the bet behind cerver:

• Cost per task, not per month. Each session carries its own dollar cost, so the $700 feature and the $40 overnight loop stop looking the same on the invoice.
• Route cheap work to cheap models. Because the session is model-agnostic, high-volume routine work goes to a cheap-good-enough model and the frontier model is reserved for what needs it — instead of paying top rates for everything by reflex.
• Predictable by design. Cerver's own price is a flat $2 per 1M tokens, visible per session — token-based, but a number you can actually forecast.

The takeaway

Flat-rate AI coding isn't coming back, and Copilot won't be the last to make the switch — usage-based is simply what agents cost. The teams that win the next year won't be the ones who spent the least. They'll be the ones who could see and steer the spend, session by session, instead of bracing for the invoice.

And there's an uncomfortable gap underneath all of this: nobody actually knows what "normal" AI-coding spend looks like yet. We're collecting it — and the early numbers are wider than anyone expects. Add yours and see where you land.

A note on the numbers: the $29 → $750 figures reflect GitHub's June 1, 2026 move to usage-based Copilot billing and developer reports of cost increases for heavy agentic use; read them as reported, not as a measurement of any specific account. The token economics echo the illustrative model in our Uber piece.