In early May, GitHub shipped a billing preview tool. It runs in your browser, reads your Copilot usage history, and projects what you would have paid under the token-based billing model that takes effect June 1. The numbers it produces have been described by users as the kind of revelation that "could sober up an alcoholic from mere shock."1
The Sequence
On April 10, GitHub paused new Copilot Pro trials.2 On April 20, they paused new signups for Copilot Pro, Pro+, and Student plans entirely. No estimated timeline for reopening.3 The same day, they removed all Opus models from the Pro tier and announced that Opus 4.5 and 4.6 would be removed from Pro+, leaving only Opus 4.7 in medium thinking mode.3
On April 27, GitHub announced that all Copilot plans would transition to usage-based billing on June 1. The flat-rate premium request model was "no longer sustainable."4
Then came the preview tool. Log into GitHub, visit your Billing Overview, and the tool shows you what your current usage patterns would cost under the new system. GitHub described it as giving users "visibility into projected costs."5
They also set a refund deadline: May 20. Cancel before then and get your money back for the remaining subscription time.6
Read the sequence again. Pause signups. Remove the expensive models from cheaper tiers. Announce the pricing change. Ship a tool that shows users the damage. Set a refund deadline. GitHub built an off-ramp because they knew users would need one.
The Multipliers
Under the current system, a request to Claude Sonnet 4.6 costs 1x your base rate. On June 1, the same request costs 9x. Claude Opus 4.6 goes from 3x to 27x. GPT-5.4 mini, which was 0.33x, becomes 6x. That is an 18x increase for the cheapest non-free model in the lineup.7
| Model | Current | June 1 | Change |
|---|---|---|---|
| Claude Opus 4.6 | 3x | 27x | +800% |
| Claude Opus 4.7 | 15x | 27x | +80% |
| Claude Sonnet 4.6 | 1x | 9x | +800% |
| GPT-5.4 | 1x | 6x | +500% |
| GPT-5.4 mini | 0.33x | 6x | +1,718% |
| Gemini 3.1 Pro | 1x | 6x | +500% |
| GPT-4.1 | 0x (free) | 1x | ∞ |
| Haiku 4.5 | 0.33x | 0.33x | unchanged |
The only models that don't change are the ones nobody is choosing for serious work. Haiku stays at 0.33x. Gemini 2.5 Pro stays at 1x. Everything a developer would actually select for an agentic coding session got multiplied by 5 to 27.7
The Math
Under token-based pricing, one AI Credit equals one cent. Copilot Pro includes $10 in monthly credits. Copilot Pro+ includes $39.5
Claude Opus output tokens cost $25 per million. GPT-5.5 output costs $30 per million. A single agentic session with Opus, running a plan-and-implement on a substantial feature, can consume the entire Pro+ monthly allotment in one sitting. One user documented the cost: $35 for a single chat session.1
At Pro+ pricing, that leaves $4 for the rest of the month.
Another user calculated that a Pro+ subscriber using Opus would get roughly 140 requests per month under the new system.1 That is about seven requests per working day. For a $39 monthly subscription to the most advanced AI coding assistant on the market.
One Hacker News commenter ran the numbers on a long-running task under the old system. At API prices, it would have cost approximately $260. Under the old Copilot model, they paid $0.12.8
The preview tool did not invent these numbers. It revealed them.
— Hacker News commenter
The Subsidy Was the Product
The reason the preview numbers are shocking is that the old numbers were fictional. GitHub was absorbing most of the inference cost. A $10 monthly subscription was buying hundreds of dollars in compute. The gap between what users paid and what the models cost was not a rounding error. It was the business model.
GitHub's CPO, Mario Rodriguez, said it directly: "Today, a quick chat question and a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable."4
The weekly operating cost of running Copilot approximately doubled between January and May 2026, driven by agentic workloads.9 Users who adopted agent mode, who used Copilot the way GitHub marketed it, who let it run multi-step coding sessions across files for fifteen minutes at a stretch, are the users whose bills will spike the most. The power users are the expensive users. The feature that sold the subscriptions is the feature that broke the pricing.
Copilot grew from 15 million to 20 million paid enterprise seats in a single quarter.10 Each of those users consumes more tokens per session than they did three months ago, because each update made the tools more capable and more agentic. The growth that justified the subsidy is the growth that ended it.
The Exits
GitHub set the refund deadline for May 20. Annual subscribers who don't renew will fall to Copilot Free when their term expires. The only individual plan still accepting new users is the free tier.3
The community response has been direct. One user on GitHub's discussion forum: "like you expect someone with brain-defect to accept this deal." Another described the Opus removal from Pro as a "rug pull" that "broke my agent coding workflows." Another: "I've stopped using Microsoft GitHub Copilot altogether."1
On Hacker News: "The recent pricing changes have just made my Copilot subscription go from great deal to awful value overnight." And: "I'm honestly thinking about cancelling and switching to alternatives while also looking at investing in a local LLM setup."8
The alternatives absorbing the switchers are Cursor, Claude Code, Windsurf, Continue, and Cline. Developers now average 2.3 AI coding tools simultaneously.9 The loyalty that flat-rate pricing was supposed to build never materialized. When the price of a tool reflects its actual cost, users comparison-shop.
A community-built tool called copilot-arewecooked appeared on Hacker News the same week. It analyzes local session logs, applies the new per-token pricing, and generates reports comparing usage against plan tiers. The creator's warning: "Your costs might be significantly higher than the current premium request model."11
The Pattern
This is the fifth Rationing story we have published. Anthropic moved to pay-per-overage. Windsurf replaced credits with quotas. Cursor tightened limits. OpenAI added advertising. GitHub paused signups, removed models, and shipped a tool that shows you the bill.12
The pattern is the same every time. Subsidize to acquire users. Make the tools good enough to create dependency. Then adjust the pricing to reflect actual costs. The adjustment always arrives as a surprise, even though the math was never in doubt. Inference costs money. Models that think longer cost more money. Agents that run for fifteen minutes cost much more money. The flat-rate era was a customer acquisition cost amortized across millions of users, funded by investors who expected the extraction phase to arrive eventually.
It arrived.
GitHub did one thing differently. They built a preview tool. They showed users the number before charging it. They set a refund deadline. In an industry that has spent two years obscuring the true cost of AI inference behind flat-rate subscriptions and "unlimited" plans, the billing preview is the most honest artifact any vendor has produced.
The honesty is the problem. The number was always this high. The only thing that changed is that now you can see it.
Disclosure
This article was written using Claude, an AI model made by Anthropic. Anthropic competes with GitHub Copilot through Claude Code. The billing data and user quotes are sourced from GitHub's official documentation, community discussions, and developer forums. The multiplier table reflects GitHub's published rates as of May 12, 2026.
Sources
- GitHub Community Discussion #192948, "GitHub Copilot is moving to usage-based billing." User reactions, projected cost calculations, and session cost documentation. Link.
- GitHub Changelog, "Pausing new GitHub Copilot Pro trials," April 10, 2026. Link.
- GitHub Blog, "Changes to GitHub Copilot Individual plans," April 20, 2026. Signup pause, model removals, refund policy. Link.
- GitHub Blog, "GitHub Copilot is moving to usage-based billing," April 27, 2026. Mario Rodriguez quotes, AI Credits system, June 1 transition. Link.
- GitHub Docs, "Preparing for your move to usage-based billing." Preview bill tool, plan allotments, AI Credits pricing. Link.
- GitHub Community Discussion #192963, "Announcement & FAQ: Changes to GitHub Copilot Individual Plans." May 20 refund deadline, cancellation process. Link.
- GitHub Docs, "Model multipliers for annual plans staying on request-based billing." Current vs. June 1 multiplier table. Link.
- Hacker News discussion, "GitHub Copilot is moving to usage-based billing." User cost calculations, bill shock reactions, alternative tool discussion. Link.
- Where's Your Ed At, "Exclusive: Microsoft To Shift GitHub Copilot Users To Token-Based Billing, Tighten Rate Limits." Weekly cost doubling, industry subsidy analysis. Link.
- MindStudio, "GitHub Copilot's CPO Says the Flat-Rate AI Pricing Model Is Dead." Copilot seat growth, Nadella earnings call statements. Link.
- Hacker News discussion, "copilot-arewecooked." Community tool for projecting usage-based costs from local session logs. Link.
- Sloppish Rationing series: The Rationing, Rationing: The Double Cross, Rationing: The Industry. Coverage of the industry-wide shift from subsidized to usage-based AI pricing. Link.