Why Cursor May Be Heading From Flat Seats to Metered Usage

The push toward metered usage at Cursor isn't a billing detail: it's the sign that inference cost is straining the flat-subscription model of AI SaaS.

Cursor's changelog (cursor.com/changelog) shows, month by month, a product that stopped being an editor with smart autocomplete and became a fleet of agents running in the cloud. In September 2026, the company launched Projects: a coordinating agent that plans the work, delegates to "thousands of subagents," and maintains context over months, with each Project running "on its own computer in the cloud." Before that came cloud agents that wake up on their own with Slack events and PRs, subagents with their own virtual machines, and even code hosting (Origin). The product reading is obvious: more autonomy. The business reading is the one that matters to whoever builds AI SaaS, and it's more uncomfortable.

What the Flat Seat Hides

The classic SaaS model assumes a marginal cost per user close to zero. You charge a monthly seat, and the ten-thousandth user costs nearly the same to serve as the first. That's what sustains the historically high gross margins that justify software companies' revenue multiple.

A product that embeds an LLM breaks this premise from the start. Every request to a model carries a real inference cost, measured in tokens, paid to the provider (OpenAI, Anthropic) or to the GPU itself. When the product was autocomplete, the cost per seat was predictable: a dev generates a more or less stable volume of suggestions per day. When the product becomes an autonomous agent that runs in a loop, on a cloud machine, without waiting for a prompt, the cost per seat stops having a ceiling. A single user can trigger a Project that spins up thousands of subagents for weeks. The changelog describes exactly this scenario: agents that "hold on to a goal until it's done" and run long sessions without intervention.

This is where the flat seat stops being a business model and becomes a bet against the user itself: the company only profits if the customer uses it little. The entire product Cursor is building pushes the customer to use it a lot.

Metered Usage Pressure as a Consequence, Not a Choice

Migrating to usage-based billing, in this context, would not be a greedy move. It would be the way to align revenue with the variable cost that the agent architecture itself creates. When the cost of serving is dominated by inference and scales with usage intensity, charging by consumption is what any honest spreadsheet demands. It's the same reasoning cloud operators have applied for two decades: charge for what runs.

The changelog reinforces this logic on the infrastructure side. Self-hosted machines and dynamic pool scheduling (September 2026) allow capacity to "grow when requests arrive and shrink when workers disconnect," with idle machines hibernating so as not to "keep expensive capacity warm just for the next prompt." The phrasing is revealing: Cursor is treating GPU as a cost to be squeezed in real time. A company with comfortable margins doesn't write documentation obsessed with not leaving capacity idle.

What This Would Do to the CAC/LTV of Those Who Copied the Model

During the 2023-2025 boom, many AI products sold subscriptions with unlimited use at a fixed monthly price, based on the low CAC and the inflated LTV that the seat model suggests. The problem is that LTV assumes stable margin over the customer's lifetime. With variable inference cost, the most engaged customer, the one who should be the most valuable, may be the one generating losses. The best user becomes the worst unit economics.

A potential migration to metered usage fixes the margin, but it charges the funnel dearly:

• CAC doesn't change, but conversion does. Variable pricing introduces bill anxiety. The user who accepted a fixed value hesitates in front of a running meter. Part of the top of the funnel evaporates.
• LTV becomes honest, and smaller on paper. You stop counting usage revenue that was actually generating losses. The number drops, but it becomes real.
• Retention becomes tied to delivered value. Whoever pays by consumption only stays if the agent delivers useful work. This is healthy, but it exposes any product that depended on forgotten subscriptions.

The Counterpoint: What If Inference Cost Plummets?

The thesis that metered usage is inevitable has a serious enemy: inference prices are falling fast. If the cost per token drops an order of magnitude in the coming years, as it already has, the premise of "high marginal cost" weakens, and the flat seat once again makes the math work. In this scenario, whoever migrates early to metered usage may scare away users over a problem the market would solve on its own.

There's a counter-argument to the counter-argument. The drop in price per token has been more than offset by the explosion in the number of tokens consumed per task. Agents that run in loops, generate and test code in isolated sandboxes, and maintain context for months burn orders of magnitude more tokens than a chat. As long as the product frontier remains "more autonomy, more subagents, more machine time," the total cost per customer rises even with cheaper tokens. Cursor is clearly betting on this side: the entire changelog is about making the agent run more, for longer, on more machines.

What This Means for Those Building in Brazil

For the Brazilian founder building a product with an embedded LLM, the Cursor case is a warning about financial architecture, not just pricing. Three practical implications:

1. Instrument consumption from day one. Even if you charge a flat seat now, you need to know the inference cost per customer, per feature, and per session. Without that meter, you don't know which customer is profitable, and you'll find out late.
2. Design pricing to migrate without trauma. Cursor has the scale and brand to reprice and absorb the friction, if and when it does so. An early-stage startup doesn't. Hybrid models (a fixed base with included usage credits and marginal billing above the cap) preserve the predictability the customer wants and the margin you need.
3. Be careful with the "unlimited" promise. In Brazil, where average SaaS ticket size is already squeezed by the dollar exchange rate on inference costs, selling unlimited agent use is selling losses at a discount. The cost is in dollars; the revenue, often, in reais.

Cursor's product trajectory doesn't say that AI SaaS has bad margins. It suggests that AI SaaS margin is not structural, it's operational: it depends on how well you match billing with variable cost. Whoever treats inference as if it were a fixed server cost will find out, at the first cloud bill that doubles, that they copied the product without copying the spreadsheet.