The Flat Rate Is Dead. GitHub Just Made It Official.

GitHub Copilot switched to token-based billing today, ending the flat-rate model that has subsidised agentic coding sessions since Copilot expanded beyond autocomplete. The sticker prices are unchanged. The effective prices, for heavy users, are not. Nine hundred downvotes and four hundred comments on GitHub's community thread confirm that developers noticed.

What actually changed. Copilot Pro stays at $10 per month. Copilot Business stays at $19 per seat per month. What changed is the unit of consumption. The old system sold "premium request units" — an opaque abstraction that bundled model calls into a quota developers could mostly ignore. The new system sells GitHub AI Credits, denominated directly in tokens, priced per model per tier. One AI Credit equals $0.01. The monthly plan buys a fixed credit allotment; overages bill at published per-token rates.

The problem is that agentic coding sessions are expensive. A single session using Claude Sonnet 4 or GPT-4.1 through Copilot's agent mode reportedly costs $30 to $40 in credits, per the GitHub community thread and MLQ.ai's coverage. A Copilot Pro subscriber's entire monthly allotment is $10. One heavy agentic session, if that arithmetic is right, costs three to four times the monthly subscription. The community's summary — "you will get less but pay the same price" — is arithmetically correct for anyone running agent mode regularly.¹

That figure is the reported credit cost of a single agentic coding session under the new billing, against a $10/month Pro plan.

The inference-economics story. GitHub launched Copilot in 2021 on autocomplete use-cases. A code completion is cheap: a short prompt, a short completion, milliseconds of compute. Frontier model API costs at that scale are tiny fractions of a cent per interaction. At $10 or $19 per seat, the unit economics worked fine.

Agentic coding is a different product running on the same pricing sheet. An agent loop — where the model reads context, plans steps, writes code, runs tests, reads error output, iterates — produces tokens at volume. The research file cites Claude Sonnet 4 at 0.06 credits per thousand input tokens and 0.24 credits per thousand output tokens on GitHub's published rates. A multi-step agent session with a large codebase context is not a code completion; it is something closer to an API call that runs for minutes, consumes hundreds of thousands of tokens, and costs accordingly. The $10/month price point was never a price for that product. It was the price for the old product, applied to a new one because the old pricing sheet existed.

The inference-economics frame (the structural shift from training cost to inference cost as the binding constraint) makes a clean prediction here: as inference costs dominate, flat-rate products built around cheap inference get repriced or killed. Copilot repriced. Sora was killed. The prediction held.

SaaS apocalypse, developer-tools edition. The more interesting structural story is what this does to the category. Copilot, Cursor, Claude Code, and OpenAI's Codex are now all consumption-priced for agentic sessions. The unit of sale has shifted from "access" (a seat, a subscription) to "task" (a session, a token bundle). This is the SaaS apocalypse (the thesis that per-seat pricing collapses as AI agents replace human-hours) playing out in one specific vertical.

For vendors, the revenue dynamics change substantially. ARR (annual recurring revenue, the run-rate of subscription income) becomes harder to forecast when consumption variance is high. A seat-based model has predictable churn mechanics: the customer cancels or doesn't. A consumption model has a different failure mode: the customer caps spend, installs a hard budget limit, and effectively becomes a sporadic user. GitHub's billing system includes budget controls and projected-bill previews. Those features are not optional nice-to-haves; they are the mechanism by which enterprise buyers avoid political exposure to a surprise $40,000 quarterly overage charge.

The upside for vendors is a higher revenue ceiling. A developer who runs ten heavy agentic sessions per day generates far more revenue under consumption pricing than under any seat model. Net dollar retention (revenue kept from existing customers after expansion and churn) should expand for vendors whose users increase agentic workflows, assuming those users don't cap out or switch.

The question is whether developers doing heavy agentic work are price-sensitive enough to switch, or valuable enough to keep paying.

The Microsoft internal-external asymmetry. One detail in the research file is worth naming directly. Microsoft reportedly cancelled internal employee Claude Code token bills around May 29, two days before the external billing change went live. Microsoft staff running Claude Code: free. External Copilot customers running equivalent agentic sessions: metered.

This is not unusual corporate behaviour. Large software companies routinely provide staff with product access at cost or for free; it's a combination of productivity investment and dogfooding. The interesting part is the timing proximity, which makes the value-capture logic unusually legible. Microsoft is telling us two things simultaneously: the token cost is real enough to be worth absorbing as a staff benefit, and it is also real enough to bill customers for. Both of those things can be true. The first one undermines the framing that the new billing is primarily about "transparency."

The contrarian case. The community backlash is real, but it skews heavily toward heavy users. GitHub's position — that code completions remain unlimited, and that light users (autocomplete, occasional chat) will not see cost increases — is probably accurate for the median seat. The 900 downvotes represent developers who are running agent mode regularly, which is a minority of the installed base.

More importantly, the critics are implicitly arguing that the old pricing was correct. It was not. It was a subsidy designed to drive adoption. Developers who run $40 of agentic sessions per session and generate meaningful productivity value from them are not obviously being exploited at consumption prices; the relevant question is whether the productivity value exceeds the cost. If it does, the complaints are about price psychology. If it doesn't, those developers will switch, and the market will tell us something useful about where the actual value lies.

What to watch. The competitive benchmark that matters is whether Cursor or Claude Code can undercut Copilot's effective per-session rate for the heavy-agentic user. If Copilot at $30–40 per session is more expensive than a comparable Cursor session, the developers most likely to comparison-shop (the sophisticated, high-volume users who drove the community backlash) have a credible migration path. If the rates are roughly comparable across the category, the backlash becomes noise: there is nowhere cheaper to go.

Watch also the enterprise churn signal in Microsoft's next quarterly disclosure. Copilot Business at $19/seat is a meaningful revenue line; if enterprise buyers hit budget limits and reduce effective seat utilisation, that shows up in net dollar retention figures before it shows up in seat count.

The flat rate is dead. Every developer-tools vendor in this category now has to explain what an agentic session costs. That is probably good for the market in the long run, and it is definitely uncomfortable for customers in the short one.

Footnotes

¶