Free trial or free plan in 2026: what converts for a small SaaS

Every small SaaS makes this choice before it has the data to make it well: let people in free for 14 days, let them stay free forever on a smaller plan, or ask for a card before they see anything. Most founders copy whatever the product they admire does. That is a reasonable start, but the numbers behind each model have moved, and in 2026 there is a new cost on the table: a free user who calls an AI model costs real money every time they click.

This piece puts the public benchmarks side by side, works through what each model gives you from the same 10,000 visitors, adds the cost of free users when the product runs on a model, and shows how many sign-ups you need before a test can tell two models apart. The short answer for most small products is below, but the useful part is the sums.

The four models

- Free trial, no card (opt-in). Full product for a fixed time, usually 14 days, then pay or lose access.
- Free trial, card up front (opt-out). Same, but the card is taken at sign-up and charged when the trial ends unless the user cancels.
- Freemium. A free plan with limits, forever. People upgrade when they hit a limit or need a paid feature.
- Reverse trial. New accounts get the paid plan free for a while, then drop to the free plan instead of losing access. Calendly does this: a 14-day trial of its Teams plan, then the free tier, no card asked.

In January 2026 ChartMogul, ProductLed and Kyle Poyar's Growth Unhinged surveyed 200 software products. 57% use a free trial as the main way in, 26% freemium, 7% a reverse trial, 7% an interactive demo and 4% a paid trial. Of the trials, 62% last 14 days, 14% last 7 and 14% last 30, and only 20% ask for a card.

What the benchmarks say

Conversion here is the share of sign-ups who pay within six months. The same report gives a "good" and "great" range for each model:

A second data set, from the marketing agency First Page Sage (86 SaaS companies, 2022 to 2025, mostly its own clients), has higher conversion figures but the same shape: opt-out trials convert 48.8% of trial users, opt-in trials 18.2%, freemium 2.6%. The reverse-trial sample in the ChartMogul report was small and the report itself says the difference from the other models is not statistically significant, so treat reverse trials as promising rather than proven.

A conversion rate on its own tells you almost nothing. Count paying customers per visitor.

Same 10,000 visitors, four models

Conversion rates flatter whichever model has the fewest sign-ups. Asking for a card scares off most people at the door, so the ones who come in convert well. The fair comparison starts from visitors. Here is a month of 10,000 visitors to the pricing page of a product at $29 a month, using the ChartMogul funnel numbers:

The two data sets disagree on the most important question. In ChartMogul's numbers a card up front wins by a wide margin per visitor; in First Page Sage's, the no-card trial wins and the card costs you about a fifth of customers. Both agree freemium brings the most sign-ups and the fewest payers per visitor. That disagreement is the honest state of the evidence, and the reason to run your own test rather than pick a side from a blog post.

Two things the table leaves out. Card-up-front trials carry more refunds, disputes and angry "I forgot to cancel" emails, and some of those customers churn in month two. Freemium sign-ups keep paying off later: people who stayed free for months, and the colleagues they invite, upgrade long after the six-month window these reports measure.

What a free user costs when the product calls a model

For a classic SaaS a free user costs a few cents of hosting. For a product built on a language model each free user's request has a price. The ChartMogul report says it directly: supporting free users has become much more expensive because AI token costs have stayed high.

A worked example, with assumptions you should replace with your own. One request sends 2,000 tokens in and gets 500 tokens back, on a model priced at $3 per million input tokens and $15 per million output tokens:

The cap is the whole game. At 30 requests a month the free plan costs about 5% of revenue, a fair marketing expense. At 300 it eats half the revenue, before you pay for the paying customers' own usage. A no-card trial has the same per-request cost, but only for 14 days per user, so the bill stops. This is why AI products set tight free limits (GitHub's Copilot Free stops at 2,000 code completions a month), and why infrastructure companies with real costs have closed free plans outright: Heroku ended free dynos and databases on 28 November 2022, and PlanetScale stopped new free Hobby databases on 6 March 2024, saying it wanted to prioritise profitability.

If your product calls a model, our piece on gross margin for a SaaS that calls a model shows how to measure the cost per paying customer, and per seat, per use or per result covers how the paid plan itself can charge for usage.

Which to start with

Without your own data, these are sensible defaults:

- Product that shows its value in one sitting (a converter, a generator, a checker): freemium with a hard usage cap, or no sign-up wall at all for the first use. Sign-ups are cheap and the cap tells people when to pay.
- Product that needs setup, data or a team (a CRM, a scheduling or helpdesk tool): a 14-day trial without a card. Spend the 14 days getting them to the moment the product is useful, then ask.
- Product where each user costs you real money (heavy AI use, phone minutes, storage): a card up front, or a short reverse trial that drops to a tightly capped free plan.
- Reverse trial when the paid features are the ones people miss: they feel the loss on day 15 but stay in the product.

Trial length matters less than what happens inside it. Most products use 14 days; choose 7 if the value shows in a day, 30 only if users need a month-end or a payroll run to see it.

Testing it with a few hundred sign-ups

A small product cannot A/B test small differences. The numbers below are the sign-ups you need in each group to tell two conversion rates apart with the usual 95% confidence and 80% power:

So a few hundred sign-ups can settle card versus no card, because the gap is huge, but not freemium versus a no-card trial, where the medians are a couple of points apart. A practical way to do it:

1. Decide what you will count before you start: paying customers per 1,000 pricing-page visitors after 30 days, not sign-up conversion.
2. Run the two options one after the other if you can't split traffic: four weeks each, same channels, no launches or price changes in between. Note anything unusual that happened in either period.
3. For card versus no card, about 50 sign-ups in each period is enough to see a difference as large as the benchmarks suggest. For anything subtler, test something bigger first (price, onboarding) and come back to the model later.
4. Count the cost side too: refunds and disputes for the card trial, and the model and support cost of free users for freemium.
5. Keep whichever brings more paying customers per visitor at a cost you can carry, and look at it again in six months, when the later upgrades have come in.

Sources

ChartMogul with Growth Unhinged and ProductLed, The SaaS Conversion Report (study January 2026, 200 products; funnel, good and great ranges, model shares, trial lengths, card share, the note on AI token costs) and Growth Unhinged, Free-to-paid conversion report (4 Feb 2026); Userpilot, Free trial conversion rate and Fungies, Free trial vs freemium 2026 (First Page Sage figures: 86 companies, opt-out 48.8%, opt-in 18.2%, freemium 2.6%, visitor to paid 1.22%, 1.55%, 0.35%); Calendly, Pricing (14-day Teams trial, then the free tier); GitHub Docs, Plans for GitHub Copilot (Copilot Free: 2,000 completions a month); Heroku, Free dynos, Postgres and Redis no longer available (28 Nov 2022); PlanetScale, Hobby tier deprecation FAQ (6 Mar 2024, 8 Apr 2024) and LogRocket, PlanetScale alternatives (the profitability quote). Model prices in the worked example are assumptions, not a quote for a named model. Sample sizes: standard two-proportion formula, z = 1.96 and 0.84. All read October 2026.