Cursor did something on July 1 that reads like an accounting change and is actually a statement about where coding-agent pricing is headed. Every Teams seat now carries two separate pools of included usage — one for Cursor's own models (Auto and Composer 2.5), one for third-party API models routed to Anthropic, OpenAI, Google and the rest — and seats now come in two grades, Standard and Premium. It hit new customers immediately and reaches renewing customers on billing cycles starting July 1, 2026.
If you run a team on Cursor, the practical questions are: does this cost me more, and which seat do I buy? The answers are probably not and it depends on your bottleneck — and getting there is worth five minutes, because the same shape is about to show up everywhere you buy agent time.
What actually changed#
Two things, and they're independent.
The seat now has two meters. Instead of one blended pool of "usage," each seat gets a first-party pool (Auto, Composer 2.5) and a third-party API pool (everything you route to an outside model). The point is legibility: an admin can finally see whether a team's spend is going to Cursor's models or to outside APIs, and provision against the one that actually burns.
Seats come in two sizes. A Standard seat is $32/mo billed annually ($40 monthly). A Premium seat is $96/mo billed annually ($120 monthly) and ships with 5× the included usage for 3× the price — so per unit of usage it's cheaper, but only if you use it. Crucially, a team can mix the two freely: put your two agent-heavy engineers on Premium and everyone else on Standard.
Cursor's own estimate is that 90% of teams pay the same or less under the new structure. That's plausible for a simple reason: most teams don't saturate a seat, and the two-pool split stops one model class from quietly eating the budget meant for another.
The headline is the price. The story is the shape: a predictable per-seat base, plus a metered, separated pool for model spend.
Why the shape matters more than the numbers#
Zoom out and the same silhouette is forming across every serious coding agent. Anthropic sells Claude Code as flat tiers — $20 Pro, $200 Max — and recently doubled its Claude Code limits. GitHub moved Copilot to usage-based flex billing with a Max plan. Now Cursor has made the base-plus-metered split the most explicit version yet: a fixed seat you can forecast, and a usage pool you can watch.
This is what a market does when it stops selling a demo and starts selling infrastructure. Per-seat alone can't survive agents, because an agent's cost scales with what it does, not how many humans are logged in — one engineer running three agents overnight burns more than ten engineers typing. Pure usage-based can't survive either, because founders won't sign up for an unbounded meter attached to software that can loop. The stable answer both ends converge on is the one Cursor just shipped: a legible seat, plus a bounded, itemized pool for the part that varies. It's the same logic behind choosing per-seat vs usage-based vs outcome-based when you price your own product.
The floor forming underneath#
There's a second force pushing on this. While Cursor adds a Premium seat at the top, a real price floor is setting at the bottom: fixed monthly plans in the $10–$50 range — Alibaba's Qwen Coding Plan and its peers — that plug a cheap-but-capable model into the very same agent tools, including Claude Code, through an Anthropic-compatible endpoint. They don't match frontier quality, but for a solo founder or a small team doing high-volume, non-frontier work, they reset the reference price for "an agent that codes."
So the Cursor change isn't just a Cursor story. It's a checkpoint in a market splitting into two legible ends — a premium seat for teams that run agents hard, and a cheap subscription floor for everyone whose bottleneck is budget, not model quality — with the fuzzy flat-$20 middle getting squeezed from both sides.
The founder read#
- You're a small team on Cursor: do nothing yet. Odds are you're in the 90% and your bill is flat or lower. Check which of the two pools your team actually burns before your next renewal — provision against the real one.
- You have one or two agent-heavy engineers: mix seats. A Premium seat for them ($96/mo annually) at 5× usage is cheaper per unit than topping up Standard, and Standard for everyone else keeps the base down.
- Your bottleneck is cost, not quality: the interesting move is off the premium ladder entirely — a cheap Coding-Plan backend behind the same agent tools you already know.
The tool you standardize on and the model you pay for are separating into two decisions. Cursor just drew the line on its invoice.
