How to Price a Freemium AI Agent Without Giving Away Your Margin

Freemium works for software that costs nothing extra to serve. An AI agent is not that software, because every free query burns real tokens, and founders who copy the old playbook find that out the hard way.

Quick Summary - TLDR:

• Token costs scale per query, so a free tier with unlimited usage can lose money on every single user before anyone converts
• Anysphere's Cursor and Notion AI both cap free usage by a hard unit (completions, queries) rather than time, because time-based limits don't track inference cost
• A useful free-tier ceiling is set from your own cost-per-query number, not from a competitor's pricing page or a round number that feels generous
• Usage-based add-ons, the model Replit and Linear's AI features use, convert better than a hard paywall because they charge only the heavy users who are already getting value

Here's the number that should scare you before you ship a free tier: GPT-5 class inference runs somewhere between half a cent and several cents per complex query, depending on context length and tool calls. Multiply that by ten thousand free signups a month doing five queries each, and you've burned real money before a single one of them has paid you anything. That's not a hypothetical. It's arithmetic any founder running an agent on OpenAI's or Anthropic's API can do in about thirty seconds, and most skip it.

Traditional SaaS freemium worked because the marginal cost of one more free user was close to zero. Dropbox giving away 2GB of storage cost them a sliver of a hard drive. Slack letting a five-person team chat for free cost them almost nothing in server time. The entire freemium playbook, acquire cheaply, convert slowly, was built on that assumption. AI agents break it, because the free tier isn't storage sitting idle, it's compute running on demand, and the agent doing real reasoning, calling tools, holding context, costs you money every time someone hits send.

The mechanism matters here, not the abstraction. A freemium SaaS tool has a fixed cost structure: build it once, serve it to a million users for pennies each. An AI agent has a variable cost structure that tracks usage almost one to one. A user who sends 200 messages a day costs you roughly 200 times what a user who sends one message costs you. There is no equivalent in a Trello or a Dropbox free tier, where usage patterns barely move your bill.

This changes what "free" should mean. You're not giving away a feature, you're giving away metered inference, and metered inference has a price tag attached in real time by your model provider. Anthropic and OpenAI bill per token, not per user, so your free tier's cost is a direct function of how many tokens your free users burn, not how many free users you have. Two users and two thousand users can cost you the same amount if the two are power users and the two thousand barely touch the product.

Set the ceiling from your cost, not from habit

Start with your actual cost per query, not a guess. Pull your API billing for a representative week, divide total token spend by total queries served, and you have a real number. If that number is $0.008 per query, a free tier of 50 queries a month costs you $0.40 per free user before you've made a cent. That's a manageable number. A free tier of 50 queries a day costs you $12 a month per free user, and if your free-to-paid conversion rate sits anywhere near the industry norm, most of those users never convert.

Chaitanya Prakash, writing about Cursor's pricing evolution on the Replit engineering blog last year, pointed out that AI coding tools converged on hard unit caps rather than time-based trials for exactly this reason. Cursor doesn't give new users "14 days free," it gives them a fixed number of fast completions, then throttles to a slower model. Notion AI did something similar, capping free users at a limited number of AI responses per workspace rather than a calendar window. Both approaches tie the limit to the thing that actually costs money, the number of model calls, rather than to a clock that has nothing to do with cost.

That's the mechanical fix: price your free tier in units of inference, not units of time. A monthly subscription trial measures the wrong variable entirely. Someone who signs up and uses your agent twice in 30 days costs you almost nothing and gets almost no value, so they churn. Someone who signs up and hammers it 400 times in three days gets the most value and costs you the most, and a time-based trial charges both of them identically, zero, for completely different amounts of your money.

Where the free tier should actually end

Frankly, most founders set the ceiling too high because they're afraid of looking stingy next to a competitor's splashy "unlimited free tier" marketing page. Don't bother competing on that number. Replit's own pricing shifted from a generous unlimited AI tier toward metered credits once usage data showed a small fraction of free users were consuming a disproportionate share of compute, the same pattern Linear's engineering team described when they moved AI features behind usage-based add-ons rather than bundling them into the flat subscription price. The lesson from both is the same: a flat, generous free tier attracts exactly the heaviest users first, because power users are the ones most likely to find and stress-test a new tool.

Set your ceiling at the point where a typical evaluating user can genuinely assess whether your agent solves their problem, and no further. That's usually a small number of full task completions, not messages. If your agent drafts a legal contract, three completed drafts tells someone whether it's useful. Thirty completed drafts just gives away the product. The specific number depends on your task complexity and your margin tolerance, but the method is fixed: work backward from "how many times does someone need to see this work before they trust it," not forward from "how generous can we afford to sound."

Converting without a hard wall

The free-tier-to-paid conversion rate for AI products tends to run lower than classic SaaS benchmarks, because the thing being gated, inference, is also the thing that makes the product good. Gate it too hard and the free user never experiences enough value to want more. Gate it too loosely and you're subsidizing people who'll never pay.

The middle path that's worked for usage-heavy AI tools is a soft overage rather than a hard wall: let someone hit their free ceiling, then offer pay-as-you-go top-ups at a visible per-query price before pushing them to a full subscription. This is effectively what OpenAI itself does with ChatGPT's free tier, capping message volume and model access rather than cutting users off entirely, and letting the limit itself do the selling. Someone who hits a usage wall mid-task and sees exactly what they'd get by paying converts at a meaningfully higher rate than someone who gets a generic "upgrade now" banner on day one.

None of this requires exotic pricing theory. It requires knowing your per-query cost before you set a free-tier number, and resisting the instinct to match a competitor's giveaway instead of your own margin. The founders who get burned aren't the ones who charge too much too early. They're the ones who never ran the arithmetic at all.

Also read: How to Price an Enterprise Pilot Without Giving Away the Product • How to Structure an AI Agent Pilot So Procurement Actually Says Yes • How to Price an AI Agent When It Fails a Task Halfway Through

This article is posted in AI News, check it out for more related stories.