Why SaaS Freemium Breaks in AI

Every free prompt burns GPU, so gate usage, outcomes and heavy modalities instead of your best features.

CheatsheetLayout, Section, NumberedItem, ExampleSection, Callout, } from "@/components/CheatsheetLayout";

Why the old playbook breaks

In traditional SaaS, serving one more free user costs roughly nothing. In AI, every time a free user hits Enter your GPUs fire and cash burns. The classic play, give away the basics and gate the best features, assumes the free tier is cheap. It is not. Worse, you cannot gate the magic. As Elena Verna argues, AI products have to give away a large amount of value before a user reaches the aha moment. Someone needs to watch the model write the perfect email before they come back a second time. Time to value has to be instant. That is the paradox Google hit when launching its AI subscriptions. The free Gemini tier could write flawless Python, draft marketing copy and summarise long email threads. Users asked a fair question: why pay $20 a month when the free version is already smarter than me? The SaaS answer would have been to paygate the viral features: the NotebookLM podcast generator, Nano Banana images, Deep Research, the way Slack gated message history and Figma gated team libraries. But that would have stopped users building a daily habit, and it would not have touched the biggest compute costs. So the paywall was rebuilt around three pillars.

Pillar 1: gate usage intensity

Google's first paywall was the familiar premium play: a single $20 Gemini Advanced tier where you pay for the smartest model. Two problems showed up fast. The free tier was already so capable that most users saw no reason to upgrade. The power users who did upgrade consumed so much compute that the unit economics were, in the author's word, terrifying. The real value was not answer quality. It was the volume of work a user could push through the system, and that had to be priced. Each tier maps to a level of usage intensity, with higher limits and larger context windows (up to 1 million tokens) as you move up. Every tier gets a taste of most features, and prepaid tiers avoid the AWS style bill shock consumers hate. The lesson: gating usage intensity was a stronger lever than gating model intelligence. Fast Mode gives instant GPU access from a limited pool of monthly hours. Once Fast Hours run out, users on Standard plans or higher drop into unlimited Relax Mode, a queue that moves with demand. People pay for priority access and more images faster, not better images, and the company keeps its unit economics tight.

Pillar 2: gate outcomes

Pillar 1 monetises raw compute. Pillar 2 monetises productivity. The free tier can give the right answer but still leave the user copying, pasting, re-prompting and formatting. Put the paywall in front of features that collapse a multi step task into a single click. Pro users will justify a subscription on hours saved alone. Gemini Advanced was originally sold on the best and fastest answer. As models became agents, Google made Chrome auto browse, an agent that navigates sites, fills forms, compares products and books appointments, exclusive to higher tiers. The paywall was framed around labour saved. In the author's words, "We stopped selling answers and started selling hours." The post calls Fin the gold standard of outcome based pricing. Letting the AI try is free; you pay only when the customer confirms the problem was solved. Sierra follows the same model.

Pillar 3: gate the heaviest compute modalities

Some things cannot be free, physically. When Google planned the launch of Genie 3, its real time interactive world model, the team joked that TPUs melted on every prompt. There were not enough TPUs to serve the 100K+ QPS hundreds of millions of daily users would need. The useful discovery: consumers already understand this. Nobody expects a cinematic, photorealistic 3D world to be free the way they expect a 500 word email to be. So Genie 3 went to the highest tier only, which turned a compute liability into the strongest incentive to climb the pricing ladder, and let Google show off a frontier model without hitting capacity limits. The rule: keep text and basic image generation open to everyone to drive top of funnel growth, then set a hard paywall the moment someone wants to render video, run a real time simulation or enter a persistent 3D environment.

Conversion catalysts

Well built tiers capture a user's experimental budget. They do not keep it. AI subscriptions churn harder than traditional SaaS because usage habits are still forming. Deciding what goes in which tier is half the job; the other half is when and how you prompt the upgrade. The post names three kinds of catalyst:

Behavioural triggers

The moment a highly engaged free user crosses a line that the pillars define: Fast Hours exhausted, a multi step chore the agent could have done, a request for a heavy modality.

Contextual UX nudges

The upgrade prompt shown in the flow, at the point of high intent, rather than as a generic banner. The author calls this more art than science.

Strategic packaging

Bundling tiers so the roadblock reads as an obvious upgrade rather than a frustration. The goal in every case is the same: the right nudge, to the right user, at the exact moment of intent.

How to decide for your product

Where does compute actually go? Rank features by cost per use before gating anything. The viral feature is often not the expensive one. What must a user feel before day two? That stays free, however good it is. Can you meter intensity? If yes, sell volume and speed in prepaid tiers before model quality. Which features remove hours of manual work? Those are your outcome gates. If you can verify the outcome, consider charging per result. Which modality would melt the hardware if free? That is your top tier trigger. Where is the moment of intent? Put the upgrade prompt there, not on a pricing page. For choosing between add-on, bundle and standalone, and for price points, see How to Monetize AI Features. Do not gate your best features. Gate the usage, the outcomes and the compute heavy modalities, and keep the magic free so the habit forms before the bill arrives. Source: Vikas Kansal, "Why SaaS freemium playbooks don't work in AI, and what to do instead", Lenny's Newsletter (May 2026). );