What a free AI tier actually costs you

You have been on the free plan for months and it has been fine. ChatGPT’s Free tier is $0 and includes unlimited text chats with GPT-5.6 Luna, plus limited uploads, limited image generation, limited voice, limited deep research and limited memory [1]. Most days that covers the work. The bill you are not paying is the whole point, and it feels like the one line in your software spend that cannot go wrong.

It can go wrong, in a specific and boring way. A free tier is a price the vendor set, and the vendor reviews it on a schedule you do not see. What you pay instead of money is ads, data, or a ceiling on how much you can do in a day, and all three of those can move. This guide is for a solo operator or small team running real work through a free AI plan and wondering how much to lean on it. It is not for anyone whose vendor is fixed by procurement, and it is not a comparison of which free tier is best. Figures below were read from vendor pages on 4 September 2026.

A free tier is priced off the cheapest model in the lineup

Free stops looking mysterious once you look at what a query costs the vendor. On OpenAI’s API, at short context on standard processing, GPT-5.6 Luna is $0.10 per million input tokens and $0.60 per million output tokens, while GPT-6 Astra is $5.00 and $25.00 [5]. Astra is 50 times Luna on input and about 42 times on output. A cheap model is not a lesser version of the free tier’s economics. It is the reason the free tier exists.

That is why the Free plan’s headline entitlement names the cheap model specifically. Unlimited text chats are unlimited on Luna [1]. The things that are metered on Free are the expensive things: uploads, image generation, voice, deep research, memory and context, Codex and desktop Work access [1]. Paid plans are where the reasoning models sit, with Plus listing advanced reasoning models with GPT-6 Astra and GPT-5.6 [1]. Nobody is giving away the model that bills $25 per million output tokens without a limit.

The practical read is that generosity on a free tier tracks the cost curve of the cheapest serving model, not the vendor’s opinion of you. When you plan around free capacity, you are planning around someone else’s unit economics. That is a reasonable thing to do with your eyes open and a poor thing to do by accident.

The exchange rate is published, and it is ads or fewer messages

The clearest statement of what a free tier costs comes from the ads documentation rather than the pricing page. The test runs for logged-in adult users on the Free and Go tiers, and the Plus, Pro, Business, Enterprise and Education tiers do not have ads [2]. Ads are matched using the topic of your conversation, your past chats and your past interactions with ads [2]. Advertisers get aggregate performance figures such as views and clicks, and do not have access to your chats, chat history, memories or personal details [2].

Then comes the sentence worth reading twice. Among the controls listed, you can dismiss ads, manage personalisation, delete ad data, upgrade to Plus or Pro, or opt out of ads in the Free tier in exchange for fewer daily free messages [2]. That is the exchange rate stated in plain terms by the vendor: on the free tier you pay in attention, and if you decline to pay in attention you pay in throughput instead. There is no version where you pay in nothing.

The scale of that business is not hypothetical. Testing began with a United States pilot in February, the expansion to 31 European markets was announced on 18 August 2026, and an update dated 31 August 2026 says self-service access through Ads Manager is available across those markets [3]. Six months from pilot to a self-service ad product across 31 markets is the shape of a line item a company intends to grow, and free tiers are the inventory it grows on.

The other price is written in the data terms, not the price table

The second thing you pay with is what you type. OpenAI’s own explanation is direct: ChatGPT improves by further training on the conversations people have with it unless you opt out, and you opt out through the privacy portal by clicking “do not train on my content” [6]. The same page states that by default OpenAI does not train on any inputs or outputs from ChatGPT Business, ChatGPT Enterprise or the API, and that organisations are opted out of data sharing unless they explicitly opt in [6]. The default flips depending on which side of the paywall you are standing on.

Google draws the same line and is blunter about the consequences. The Gemini API pricing page marks free-tier usage as content used to improve Google products and paid-tier usage as content not used that way [7]. The terms spell out what that means in practice: on the unpaid services Google uses the content you submit and the responses to provide, improve and develop Google products, human reviewers may read, annotate and process your API input and output, and you are told plainly not to submit sensitive, confidential, or personal information to the unpaid services [8]. On the paid services, Google states it does not use your prompts or responses to improve its products [8].

Read those two together and you get a working rule that does not depend on trusting anyone. A free tier is a place where a human employed by the vendor may read what you wrote, and where the training default is set for the vendor’s benefit rather than yours. That is not a scandal. It is the deal, written down, on a page you can open in a minute.

Every free-tier term carries a review date you do not control

The most useful evidence that free terms move is that vendors sometimes publish the move in advance. Google’s pricing page lists Gemini 3.8 Flash on the standard paid tier at $0.75 per million input tokens through 31 December 2026 and $1.50 from 1 January 2027, with output going from $3.75 to $7.50 on the same date [7]. That is a paid rate rather than a free one, and it is exactly the point: even the published, contractual number has an expiry printed next to it. The free tier sitting above it has no such commitment at all.

Compare what a free plan actually promises. Free gets unlimited text chats on one named cheap model, and the word “limited” as the entitlement for seven other capabilities, with no number attached to it [1]. Go is described as a plan that may include ads [1]. You cannot plan against “limited”, and you are not meant to. It is a level the vendor adjusts as the cost of serving it changes, as the ad business grows, and as the upgrade rate to $20-a-month Plus goes up or down [4].

So the question to ask about any free capability is not whether it works today. It is how much of your week breaks if it halves next quarter. If the answer is “an inconvenience”, carry on. If the answer is “a client deliverable is late”, you are not using a free tier, you are depending on one, and those are different arrangements with the same price.

Decide which work a free tier is allowed to touch

Once the terms are on the table, the sorting is easy and mostly about other people’s material.

Anything covered by a client agreement, anything containing a third party’s personal data, and anything you would not want a vendor’s reviewer reading goes on a paid lane or nowhere. Google’s instruction not to submit sensitive, confidential, or personal information to the unpaid services is not a caution, it is a term you already agreed to [8]. On the consumer side, the equivalent step is checking whether training is on for your account and turning it off if the answer matters to you [6].

Your own drafts, your own numbers, your own half-formed thinking, public material you are summarising, and anything you were going to rewrite anyway can all sit on a free tier without much thought. For most small operators that is the majority of the volume, which is why free tiers are genuinely useful rather than merely cheap.

The third category is the one people get wrong: work that is not confidential but is time-critical. A proposal due at five o’clock does not belong on a plan where uploads, image generation and deep research are quantified only as “limited” [1], and where declining ads costs you daily messages [2]. Free capacity is fine for work that can wait an hour. It is a poor foundation for work that cannot.

Before you decide whether to stay free, price the friction rather than arguing about it.

If the number lands under $20, the free tier is earning its place and you should stop thinking about it. If it lands at three times that, you have been paying for the free plan in the most expensive currency you have, and the upgrade is not a treat but an arithmetic correction [4].

Make leaving cheap before you need to leave

The defence against a free tier changing is not loyalty to a second vendor. It is keeping the switching cost small enough that a change is an afternoon rather than a project.

Keep prompts, instructions and reference material in files you own rather than inside one product’s saved projects and memory. A prompt you can paste into another tool is portable; a prompt that exists only as an account setting is not. This matters more on free tiers than paid ones, because memory and context are among the things the Free plan explicitly meters [1], so the state you build up inside the product is both smaller and less durable than it feels.

Know the paid price of everything you currently get free, before you need it. For ChatGPT that is $20 a month for Plus [4]. For a free API tier, it is the standard rate on the same page as the free one, which is where you find out that the free version of a model and the paid version can carry different data terms as well as different prices [7][8]. A migration you have already costed is a decision. One you have not is an emergency.

Then set a date rather than a habit. Once a quarter, open the pricing page and the data-use page of every vendor you rely on for free, and check three things: whether the free entitlement still covers what you use it for, whether the training default changed, and whether anything you now depend on has moved behind a paid plan. Fifteen minutes, four times a year, is the entire maintenance cost of this position.

What still goes wrong

The terms in this guide are the ones on the vendors’ pages on 4 September 2026, and the whole argument is that they move. Ads reached 31 European markets six months after a United States pilot [3], and a published paid rate can carry a printed increase for 1 January 2027 [7]. Anything specific here should be treated as a snapshot to re-check, not a standing fact. The method survives the changes; the numbers will not.

The quieter failure is drift in what you send. People are careful with free tiers in the first week and casual by the second month, and the pasted document that contains a client’s name arrives on a Tuesday when you are busy, not on the day you were thinking about data policy. No setting prevents this. The only real control is deciding in advance which categories of material never go through a free lane and keeping that list short enough to remember under pressure.

The last one is a matter of framing rather than risk. A free tier is a good deal precisely because a vendor wants your usage, and wanting your usage is not the same as wanting your interests. The advertising documentation says advertisers get aggregate metrics and not your chats [2], and there is no reason to doubt it. It remains true that a plan funded by ads is optimised for time spent and a plan funded by subscriptions is optimised for renewal, and those two produce different products over a few years. Notice which one you are using, and take the free one on purpose.