Cursor Hits $2B ARR as GitHub Copilot Moves to Metered Billing

Two stories landed close together in 2026, and read side by side they tell you where AI coding tools are heading. Cursor hit $2B ARR, and on June 1 GitHub Copilot switched to usage-based billing. One is a momentum story. The other is a pricing story. Both come down to the same question developers now face on every task: how much is this agent run actually costing me?

Cursor, built by Anysphere, is the upstart that turned an AI-first fork of VS Code into the tool many engineers grab for big multi-file work. Copilot, owned by GitHub and Microsoft, is the incumbent that put AI completions in front of millions of developers first. Watching the leader rewrite its billing in the same season the challenger crosses $2B is not a coincidence.

What changed

The first shift is Cursor's revenue. In February 2026, Bloomberg reported that Cursor reached $2B in annual recurring revenue. That makes it the fastest B2B SaaS company to go from zero to $2B, getting there in roughly three years. In April 2026, Anysphere was reportedly raising around $2B at a valuation near $50B, with a16z and Thrive leading and Nvidia involved as a strategic investor. The company is forecasting more than $6B ARR by the end of 2026.

The second shift is Copilot's billing. On June 1, 2026, GitHub moved its paid plans to usage-based pricing built on GitHub AI Credits, where one credit equals $0.01. Each paid plan now ships with a monthly credit allotment. Pro stays at $10 a month and includes $10 of credits. Pro+ is $39 a month and includes $39 of credits plus access to premium models like Claude Opus. Premium-model calls and heavy agent runs draw down those credits. Plain code completions and Next Edit Suggestions stay unmetered, so the everyday autocomplete most people use all day is not affected.

Why it matters

For years the pitch for an AI coding assistant was simple: pay a flat monthly fee, get unlimited help. That model is breaking under the weight of agents. When a tool can spend ten minutes planning a refactor, reading dozens of files and calling a frontier model repeatedly, a single task can cost real money. Flat pricing only works until enough users run agents constantly.

Copilot's move makes that cost visible. Instead of one bill, you get a meter, and the meter tells you which work is cheap and which is expensive. Cursor already lives in this world; its plans run on usage credits, where a heavy agent user on frontier models can burn through the included $20 of Pro usage faster than expected and end up needing Pro+ or Ultra. The two tools are converging on the same truth from opposite directions, and the Cursor vs GitHub Copilot gap now turns as much on cost behavior as on raw output quality.

What it means for developers

Cost predictability is the thing to plan around. Under a metered model you can no longer assume a fixed monthly number; your bill scales with how much agent work you trigger. That is not automatically worse. For light users it can be cheaper, and Copilot's unmetered completions mean basic autocomplete still costs nothing extra. But anyone running agents all day should track usage the same way they would watch a cloud bill.

Picking the right model per task is the lever you actually control. A frontier model is worth it for a gnarly multi-file change; a smaller, cheaper model is plenty for a quick rename or a boilerplate test. Spending a premium credit on trivial work is how budgets disappear. Both tools now reward developers who match the model to the job rather than reaching for the most expensive option every time.

The third habit is watching agent spend before committing to a tier. If you mostly write code with autocomplete and occasional chat, the entry plans are fine. If your daily loop is "describe a feature, let the agent build it," budget for the higher tiers. The Cursor review and the GitHub Copilot review both walk through where the included quota stops being enough.

What to watch next

Expect more of the market to follow. Once the leader meters agent usage and the fastest-growing challenger is already credit-based, flat unlimited pricing for heavy agent work looks like a temporary artifact. The interesting question is whether vendors give developers better in-product cost controls, so you can cap a session or get warned before an agent run gets expensive.

Watch Cursor's numbers too. A jump from $2B to a forecast above $6B in a single year is the kind of curve that invites scrutiny, and it puts pressure on Copilot's ecosystem advantage. For now, the practical takeaway is the same whichever tool you pick: treat agent runs as a metered resource. If you are still choosing between options, the best AI coding tools guide ranks both in context.