AI tools went metered in June. Read the new math.

In June 2026 the tools people build software and businesses with changed how they charge. GitHub Copilot, Cursor, Zapier, and Notion all dropped flat per-seat pricing for usage meters. The bill no longer tracks how many seats you buy. It tracks how much you run.

One cause sits under all four moves: agents. A developer who runs autonomous agents all day burns far more compute than one who taps autocomplete a few times an hour, and a flat seat was never priced for the first kind of user. So the vendors split them apart. Light users mostly pay the same or less. The heavy user, the operator who squeezed the most out of a $20 seat, now pays the real cost of what they run.

If you build with these tools, the read is short: your AI tooling moved from a fixed line item to a variable cost, and the variable is your own usage. Here is what actually changed, why, and what to do about it.

The receipts: four tools, one month

GitHub Copilot. On June 1, 2026, GitHub moved every Copilot plan to usage-based billing. Instead of counting requests, each plan now ships with a monthly budget of AI Credits, and usage spends those credits by token count, input, output, and cached, at each model's published API rate (GitHub). Pro includes $10 a month in credits, Pro+ $39, Business $19 per user, Enterprise $39 per user. Code completions and next-edit suggestions stay free and unlimited. GitHub was plain about why: Copilot "evolved into an agentic platform capable of running long, multi-step coding sessions" with "significantly higher compute and inference demands," and under the old flat price "a quick chat question and a multi-hour autonomous coding session can cost the user the same amount."

Cursor. Same day, Cursor reworked its Teams plan (Cursor). It split each seat's usage into two pools, first-party models and third-party APIs, and added a Premium seat that gives 5x the included usage at 3x the cost. Standard runs $32 a month billed annually, Premium $96. Cursor built the Premium seat for developers "running agents all day" and expects its Composer pool to "cover a full month of heavy agent usage for 99% of users." The company says outright that power users "tend to drive the majority of spend." Cursor had already traded fixed request limits for usage credits a year earlier, in June 2025; the 2026 change is the cleanup. Renewing customers move over on July 1, 2026.

Zapier. From June 15, 2026, AI by Zapier prices each AI step by model tier (Zapier). A Standard model spends 1x your task budget per run, Advanced spends 3x, Premium 5x, and every tool the model calls adds more at the same rate. If a single step hits 75 tasks in one run, Zapier pauses it and asks you to approve before it continues.

Notion. Notion retired its flat per-seat AI add-on, folded core AI into the Business plan at $20 a seat, and put a meter on its new autonomous feature, Custom Agents, billing them by credit on top of the seat (Notion).

Four products, four different mechanics, one direction.

Why agents broke the flat seat

A flat seat is a cross-subsidy. Everyone pays the same $20, the light users cost the vendor a few cents, the heavy users cost far more, and the average comes out fine. That arithmetic holds as long as the gap between light and heavy stays small.

Agents blew the gap open. An autonomous coding session can run for hours, reading files, writing code, calling tools, looping on its own output, and each step spends tokens the vendor pays for. GitHub's own line names the break: a chat question and a multi-hour agent run used to cost the same. The heavy user stopped being a rounding error and became the cost center. Once a handful of operators consume what used to be the budget for hundreds of casual users, the flat seat can't average them out anymore. The vendor either raises the price for everyone or meters the people doing the spending. They chose the meter.

This is the buyer's side of a story we already told from the provider's side. The math that loses money on heavy users is the same math that pushed Meta to price its AI for the light user and eat the rest. What's new in June is that the dev and productivity tools stopped eating it.

Who pays less, who pays more

The meter is not a blanket price hike. It is a sort.

Cursor expects its changes to lower costs for most teams and frames the Premium seat as the place heavy users go. Copilot keeps autocomplete free, which is most of what most developers use. So the casual user, the one who accepts a few suggestions and asks the odd question, often comes out flat or cheaper. That is real, and the vendors lead with it.

The person on the other side of the sort is the operator. If you run agents to ship features, generate content, or drive automations, you were the most profitable customer to acquire on a flat seat and the least profitable to keep. The new pricing finds you. On Copilot you spend past your included credits into metered tokens. On Cursor you move to the Premium seat or pay overages. On Zapier your agentic steps cost 3x or 5x per run. The tool you adopted because it automated the most work now bills you in proportion to exactly that.

It is honest pricing. It still moves the risk to you.

The fair reading is that the meter is correct. Flat all-you-can-eat pricing on token-metered compute was never sustainable, and GitHub said as much. Charging for what you use is how every other compute bill already works, from cloud servers to bandwidth. There is no villain here.

But correct for the vendor means the budget risk shifts to you. A flat seat was a number you could plan a year around. A meter is a number that moves with your usage, and usage is the thing that grows when a tool works. Two forces pull against each other now. Per-token prices keep falling as model providers undercut each other, which we walked through in the agent-loop math. Cheaper tokens pull the bill down. Heavier agent use pulls it up, and at 2026 adoption levels it usually pulls harder. The meter only bites if you actually run the heavy workloads, so the question is no longer what a seat costs. It is how much you run, and whether you can see it before the invoice does.

What to do if you build on these tools

• Budget on usage, not seats. Pull last month's actual token or task consumption per tool and price the next month off that, not off the sticker. The seat is now the floor, not the bill.
• Find your heavy workflows. A few agents or automations usually drive most of the spend, the same way a few users drive most of a vendor's cost. Know which ones, because those are where a tier change or overage lands first.
• Watch the caps and the defaults. Zapier's 75-task pause and Cursor's pool splits are guardrails, but they also reshape behavior. Read what your tools throttle, what they bill as an overage, and what tier they put you on by default.
• Keep your pipeline portable. When the cost is metered per provider, the freedom to move matters more. Hard lock-in to one vendor's pricing is the same trap as a cheap subscription with an expensive exit, which is the lock-in math behind falling token prices.

The flat seat is ending for the tools that automate real work, and it is ending because they got good enough to be worth metering. The sticker price is no longer the thing to read. The usage is. Whoever measures it first, you or the vendor, sets the budget.

Sources

• GitHub: GitHub Copilot is moving to usage-based billing (June 1, 2026)
• Cursor: Improvements to Teams pricing (June 1, 2026)
• Zapier: AI by Zapier, new model-based pricing starting June 15, 2026
• Notion: Pricing

This is not financial advice.