The Reckoning

Today is June 30, and for millions of developers who built their workflows around GitHub Copilot, the invoices now arriving are the first full accounting of a billing revolution nobody asked for. The flat-rate era is dead. The usage-based era is here, and the numbers are brutal.[1]

Since June 1, Copilot has charged per token for every input, output, and cached token an AI agent consumes. Developers who were paying $29 a month are now seeing bills projected at $750. Those on $50 plans are looking at $3,000. Not anomalies, not edge cases. The predictable result of a billing model where a single agentic coding session can consume more tokens in 20 minutes than a human developer generates in a week of manual queries.[1]

Why Agents Cost So Much

The core issue is a technical mechanism the old flat-rate model completely obscured: agentic AI billing scales non-linearly.

When a developer asks Copilot's agent mode to refactor a module, the agent does not make one model call. It reads files, each becoming input tokens. It plans an approach, generating output tokens. It produces proposed changes. If the results are unsatisfactory, it loops, and on each iteration the full conversation history re-enters the context window. A 20-turn agentic session where each turn carries 10,000 tokens of context does not consume 200,000 input tokens total. It processes approximately 200,000 input tokens on the 20th turn alone, because the transformer architecture processes the entire conversation at every step.[1]

GitHub's own research, published in May 2026, found that agentic coding tasks can consume roughly 1,000 times more tokens than standard single-turn queries. One thousand times. A billing model that treated a one-line chat query and a six-hour autonomous coding session identically was subsidizing the heaviest users at everyone else's expense. GitHub's position is that this was unsustainable. They are not wrong about that.

What Changed, Specifically

On April 27, 2026, GitHub announced all Copilot plans would move from Premium Request Units to GitHub AI Credits. Subscription prices stayed the same: Pro at $10, Pro+ at $39, Business at $19 per user, Enterprise at $39 per user. What changed is what those prices buy.[2]

Each plan now includes a monthly credit allowance. Pro gets 1,500 credits. Pro+ gets 7,000. One credit equals one cent. When the balance hits zero, usage stops unless an additional-usage budget is configured, and that budget defaults to unbounded. There is no longer an automatic fallback to a cheaper model when credits run out. Either you have credits, or the request is rejected.

Two safety nets disappeared. Annual plans were retired. The fallback model was removed. The only remaining protection is a spending cap that requires manual opt-in. Most users did not know it existed until their June invoices arrived.

Real Reports

One Pro+ subscriber burned through 8 percent of their 7,000-credit monthly allotment in two hours, projecting full exhaustion in under two days. Another spent more than $6 on a single change request. A session using Claude Opus 4.8 to fix website issues consumed 1,180 credits, 16 percent of a Pro+ monthly allowance, for results the developer described as mediocre. One user watched a file review that produced no code changes consume 20 percent of their monthly balance.[1]

The Model Context Protocol compounds the problem. When an agent has 30 tools registered, every request includes schema definitions for all 30 tools in the system prompt. That is potentially thousands of tokens per request that add zero value if the agent uses only two of them.

The Broader Pattern

This is not just a GitHub problem. It is the industry discovering what agentic AI actually costs to run. A flat-rate subscription for an AI agent that can autonomously decide how many model calls to make, how much context to load, and how many iterations to run was always going to collide with the reality that each of those calls costs real money on a GPU somewhere.

GitHub is offering promotional credits through August 2026 to ease the transition: an additional $30 per user per month for Business plans, $70 for Enterprise. Those disappear September 1. A new Copilot Max tier at $100 per month targets heavy agentic users who need a larger credit pool. Code completions and Next Edit Suggestions remain free and unlimited under all paid plans, which matters because that is the feature most developers actually use day to day.[2]

What To Do Today

If you use Copilot, open Settings, go to Billing, find the GitHub Copilot section, and enable a hard spending cap. Right now. Before the next billing cycle starts. By default, GitHub does not stop usage when credits run out. It bills whatever the agent consumes.

The flat-rate era was nice while it lasted. It was never going to last.