AI Business Models.

In this model, a monthly plan buys an allowance of credits, and customers buy top-ups or a bigger plan when they run out. It has spread fast among AI-heavy tools, though it's far from universal. Rob Litterst, co-founder of the pricing research firm PricingSaaS, found that 79 of the 500 software companies his firm tracks offered credits by the end of 2025, up from 35 a year earlier¹. That's still only about one company in six, but the number more than doubled in a year. The model has produced some of the fastest-growing companies in AI, including Cursor, the AI code editor, and Lovable, the AI app builder. The catch is that credits measure usage, not results. Customers can't tell how much work a month of credits will buy, and vendors can't always tell which customers they're losing money on.

Builders pay in the middle of their first project

Free credits are deliberately small. Lovable gives free users 5 credits a day, capped at 30 a month⁹, and one credit buys one message to the AI¹⁴. A first session can use up a day's allowance before the app is finished. That makes running out the upgrade trigger, and it usually hits in the middle of the first build, when the builder is most excited. They can pay to keep going, or wait for tomorrow's credits and build slowly. This sequence is our reading of how the pricing works; we haven't found published data on when builders convert.

Credits split the difference between seats and pure usage

Kyle Poyar, who writes Growth Unhinged, a newsletter on how software companies grow and price, explains why credits have spread so fast. Customers get a predictable monthly bill, like a license, and vendors get a usage component that protects margins as usage grows^1,2. Lovable's plans show how it works in practice. The Pro plan starts at $25 a month for 100 credits, a dial raises the allowance as high as 10,000 credits, free users get 5 credits a day, and extra credits cost $15 for 50⁹. Lovable told the news site Observer it reached $400 million in annualized revenue by March 2026¹⁰. Cursor sells a $20 Pro plan that includes $20 of frontier model usage at API prices³, and it passed $4 billion in annualized revenue by June 2026, according to an unnamed source cited by Dealroom, a startup data platform¹³.

Nobody knows what a credit buys

Elena Verna, who leads growth at Lovable, admits the problem with her own company's plan: "$25/mo gives you 100 Pro credits, but no one knows what that means"¹⁶. She argues that credits, tokens and compute units are an "imaginary currency" that puts mental distance between customers and the money they're spending. That makes a plan easy to buy and its cost hard to predict, which feeds bill shock. Her fix is to translate credits into results people understand. Firecrawl, a web scraping API, describes its credits in pages scraped and searches run. Verna also says Lovable would like to charge for outcomes one day, such as hosting or a share of what its users sell, which would move it toward models 4 and 6¹⁶.

More people use more credits

Expansion in this model doesn't come from selling seats. It comes from more people drawing on the same credits. Lovable's paid plans aren't priced per seat: one workspace's credits are shared across unlimited users⁹. So when a builder invites a colleague, the price doesn't change, but more people use more credits and the allowance runs out sooner. The workspace then buys top-ups or moves up the credit dial, which on Lovable's Pro plan runs from 100 credits for $25 a month to 10,000 credits for $2,250⁹. That's the team loop on the map. It applies to other models too, but here it shows up as burn, not seats.

Teams grow into enterprise contracts

The biggest accounts end up on a negotiated contract, and that's where much of the money is. Dealroom reported in March 2026 that about 60% of Cursor's revenue came from corporate customers, though it didn't name its source¹⁷. Cursor's Enterprise plan is sold through its sales team, with pooled usage across the company, invoice billing, single sign-on, seat management, spend limits and audit logs¹⁸. At Lovable, Enterprise customers buy a committed credit balance for the workspace, and usage beyond it shows up as overage¹⁹. Lovable says more than half of Fortune 500 companies use it in some way, and a multiyear deal with Google Cloud makes it easier for large companies to buy it through their existing cloud contract²⁰. Sacra, a research firm that tracks private tech companies, describes a shift at Replit, an AI app-building platform, "from students & prosumers to B2B & large enterprises," helped by listings on the Azure and Google Cloud marketplaces²¹.

At this level, credits become a committed annual balance with overage, much like an API contract in model 1. The trigger is our inference, not documented: a team's spend gets big enough to need a discount and one bill, or IT and security need controls the self-serve plans don't offer, such as single sign-on, spend limits, and a guarantee that company code isn't used for training¹⁸. Some vendors skip the free tier altogether and sell credit plans this way from day one, through a demo and a paid plan, which is closer to traditional B2B software.

Heavy users burn the margin

Poyar has found that about 10% of users can account for 70–80% of token consumption, which makes heavy users very unprofitable without guardrails². The evidence is everywhere. Newcomer, a tech industry newsletter, reported in August 2025 that Cursor had negative gross margins overall, driven mostly by individual developers on subscription plans, whose compute bills Cursor paid to Anthropic⁵. Sacra estimates that Replit's gross margin swung between 36% and negative 14% during 2025 because of model costs for its coding agent⁷. Anthropic itself added weekly limits to its Claude subscriptions after one user consumed tens of thousands of dollars of model usage on a $200 plan⁶. That's why the map has a margin loop. When heavy users cost more than they pay, the vendor tightens the plan: it caps usage, charges overages, or moves expensive models into pricier tiers.

Re-pricing is where trust breaks

Every one of those fixes landed on customers. In June 2025, Cursor replaced 500 monthly requests with a usage allowance. Some users ran out after a few prompts and others got bills they didn't expect⁴. Cursor CEO Michael Truell apologized, explaining that "the hardest requests cost an order of magnitude more than simple ones," and refunded the surprise charges³. GitHub Copilot made a similar shift when it started enforcing a monthly cap of 300 premium requests on its Pro plan. Each request to a pricier model counts several times over, and requests beyond the cap cost from $0.04 each¹¹. Clay, the sales data platform, went the other way in March 2026. It split its single credit into data credits and platform actions, cut data prices 50–90%, and expects the change to cost it about 10% of revenue⁸.

Projects end, so churn comes in waves

Many customers in this model buy for a project, not a job. When the app ships or the idea dies, the subscription lapses. In September 2025, Business Insider reported data from analysts at the investment bank Barclays showing that web traffic had fallen 40% at Lovable from its peak, 64% at v0, Vercel's app builder, since May and 27% at the app builder Bolt since June¹². Making unused credits roll over is one way to keep a quiet month from turning into a cancellation. Lovable's monthly credits roll over for two months⁹, and Clay's data credits roll over up to twice the monthly amount⁸.