Usage-Based Pricing for AI: What Cursor, Copilot and Meta Muse Show

Meta prices Muse by weekly token allowances, Muse Code by requests per five-hour window, and Cursor switched Pro to a pool of included usage in 2025. Here is what those choices teach founders about building a usage meter customers will accept.

Meta sells its Muse agent in plans priced “depending on tokens,” CNBC reported on Oct. 7, 2026, and gives its developer tool Muse Code a request cap that resets every five hours. In mid-2025, Cursor had to apologize and offer refunds after moving its $20 Pro plan to metered usage. Between them, these cases are a practical guide to usage based pricing AI founders can borrow from, and to the mistakes that make customers angry.

Where it stands: Muse launched on Sept. 8, 2026 with two paid plans, “Power at $20/month and Maximum at $100/month,” according to TechCrunch. Meta’s own docs describe Muse Code subscriptions as an alternative to pay-as-you-go API billing.

Why do AI products move to usage based pricing?

Because one customer’s request can cost many times another’s. Cursor said so plainly in a company blog post: “the hardest requests cost an order of magnitude more than simple ones. API-based pricing is the best way to reflect that.”

In traditional SaaS, a heavy user and a light user cost about the same to serve, so a flat seat price works. In AI, every request triggers inference, the model computation that produces an answer, and longer or harder tasks burn more tokens. A token is the small unit of text a model reads or writes, and it is the unit most vendors bill on.

Usage-based pricing ties what a customer pays to what the customer consumes. The founder’s job is to pick a unit people understand and a structure that does not shock them.

How does Meta meter Muse?

With weekly token allowances that step up by plan. Meta’s newsroom post says only that Muse is “free for most of what people need, with subscription plans for people who want to do more.” The numbers come from outside analysts and reporters.

Goldman Sachs, as summarized by Insider Monkey on Yahoo Finance, reported that “the free version is capped at 100 million weekly tokens, the $20 per-month power plan comes with 500 million weekly tokens, and $100 Maximum tier offers 3 billion weekly tokens and priority access.”

CNBC described tokens as “a metric the company uses to calculate the power AI agents need to accomplish tasks.” TechCrunch added a design detail founders should note: “the app includes a usage meter that shows users what percentage of their usage they have left.”

Product Unit metered Plans reported
Muse (consumer) Weekly tokens Free, $20, $100
Muse Code Requests per five hours $5, $15, $50
Cursor Pro Dollars of included usage $20 of usage
GitHub Copilot Premium requests per month Allowance plus spending limit

How is Muse Code priced differently from the Muse app?

By requests in a rolling window, not by tokens. Subscription Insider reported that Muse Code exited beta with three plans: “Everyday Usage: $5 per month High Usage: $15 per month Power Usage: $50 per month.” The entry plan “includes a usage allowance of 10 to 50 requests to the AI tool every five hours.”

The New Stack described the ladder in multiples rather than raw numbers: High Usage “promises three times as much usage; and the $50 Power Usage plan offers 10 times as much usage.” Expressing tiers as multiples of a base plan is a useful trick. It tells customers what they get without forcing them to understand tokens.

Meta also keeps a pay-as-you-go option. Its developer docs say: “Use our models with pay-as-you-go billing on Meta Model API and in Muse Code , or in Muse Code with a monthly subscription .” Offering both lets price-sensitive teams cap their bill and lets heavy users pay for exactly what they consume.

One caution from The New Stack’s own testing: “While Muse Code cost much less, it consumed substantially more tokens and produced a weaker result on a refactoring task.” A low price per token does not guarantee a low price per finished job.

What usage based pricing AI lesson came from Cursor’s switch?

The switch surprised customers. Cursor’s post says “we moved from request-based pricing to $20 of included usage on Pro,” and the company followed up with an offer: “We are offering usage refunds for those with unexpected usage between June 16 and July 4.”

TechCrunch later framed the move from the outside: Anysphere “changed the way it priced usage for its $20-per-month Pro plan to limit power users from abusing the plan.” It added that “Another AI coding tool provider, Replit, made similar pricing changes in June as well.”

Cursor’s own post opened with an admission: “Our recent pricing changes for individual plans were not communicated clearly, and we take full responsibility.”

The lesson is about communication more than math. Cursor’s new structure was defensible on cost, but users who did not see the change coming were billed for usage they had not planned for. Refunds were the price of a rollout that outpaced explanation.

How did GitHub Copilot handle overages?

With an allowance and an opt-in spending limit. GitHub’s changelog announced that “the monthly allowance of premium requests per user for Copilot Pro, Pro+, Business, and Enterprise plans is now enforced.” For more, “you can set a spending limit in your billing settings.”

GitHub’s note does not say compute cost drove the change, so we will not claim it did. The structure is still worth studying: a predictable included amount, a clearly named premium unit, and an overage path the customer turns on deliberately rather than by accident.

How should a founder design an AI usage meter?

Pull the cases together and a few rules emerge.

Pick a unit customers can picture. Requests and multiples (“three times as much usage”) are easier to grasp than raw tokens. Tokens track your cost more precisely, so you can bill internally in tokens and present requests or percentages.

Show the meter. Muse shows remaining usage as a percentage. Customers who can see their balance are less likely to feel ambushed.

Make overage opt-in. Copilot’s spending limit and Muse Code’s subscription-or-pay-as-you-go choice both put the customer in control of the bill.

Announce changes before they bill. Cursor’s refunds show what happens otherwise.

For the broader math behind these decisions, see our explainer on AI unit economics after inference costs, and our companion piece on freemium for compute-heavy products. If your costs are falling quickly, our article on pricing against falling inference costs covers how to plan allowances that can grow.

Meta’s developer docs list one more pricing choice that removes a common surprise: “No long-context premium : you pay the same rate whether your context window is mostly empty or almost full.”

Frequently Asked Questions

What is usage-based pricing for AI products?

Charging customers according to how much they use, measured in tokens, requests or dollars of compute, rather than a flat fee for unlimited access. Meta sells Muse plans by token allowance, while Cursor now gives Pro subscribers a set dollar amount of included usage.

How does Meta price Muse by tokens?

Muse is free with a weekly token allowance. Goldman Sachs, as summarized by Insider Monkey, put the free tier at 100 million weekly tokens, the $20 Power plan at 500 million and the $100 Maximum plan at 3 billion. Meta’s newsroom post does not list these allowances.

Why did Cursor change its pricing in 2025?

Cursor said its toughest requests had become far more expensive to serve than easy ones, so billing closer to API costs made sense. It moved Pro from request-based pricing to $20 of included usage and offered refunds for unexpected usage between June 16 and July 4.

Should an AI startup charge per request or per token?

Requests are easier for customers to understand; tokens track cost more closely. Muse Code counts requests in five-hour windows, while the consumer Muse app meters tokens. Many teams show a usage meter so customers can see what is left either way.

Sources

This article is general information for founders and is not accounting, legal or investment advice. Figures are as reported by the sources cited on the dates shown; analyst and third-party estimates are not company disclosures. Found Founders is not affiliated with Meta.