How to tell whether you still need a paid AI subscription

Somewhere on your card statement there is a line for an AI subscription you set up in a hurry. You probably had a specific reason at the time. A free tier cut you off in the middle of a job, or a client asked for something the free model would not do, or a trial ended and cancelling was more effort than paying. That reason may still hold. It may also have quietly expired, because the free tiers of the major AI products get rewritten far more often than your billing cycle does, and nobody emails you when the thing you were paying to avoid stops existing.

This guide is a method for checking, not a verdict. It will not tell you to cancel, because for a lot of people the paid tier is still the right buy. It will tell you how to find out which of the two you are, in about 20 minutes of looking plus two weeks of ordinary work. It is not for teams whose tier is already dictated by a compliance or procurement policy, and it is not for anyone running production traffic through an API, where the cost question is metered per token and needs different arithmetic entirely.

The cap you were paying to avoid may already be gone

The most common reason a solo operator or a small team buys an AI seat is not a feature. It is a counter. You were 40 minutes into a piece of work, the model stopped, and you paid $20 rather than lose the thread. That purchase logic has a shelf life.

On 6 August 2026, OpenAI made GPT-5.6 Luna the default model for Free and Go users and expanded free access with unlimited text chats, plus a “Think” button for harder questions [1]. The current help-centre page states the policy plainly: “Free users have unlimited everyday text chats, subject to abuse-prevention safeguards” [2]. The same announcement claims that responses containing at least one factual error were about 62% less common with GPT-5.6 Luna and 68% less common with GPT-5.6 Sol than with GPT-5.5 Instant, on financial, medical and legal prompts [1]. Those are OpenAI’s own internal evaluations, not an independent audit, so read them as a direction of travel rather than a measurement.

The other free tiers moved too. Google’s free plan costs “$0 / month with a Google Account” and includes Gemini 3.6 Flash, varying access to 3.1 Pro, image generation and editing, Deep Research, Gemini Live, Canvas and Gems [7]. Claude’s free plan includes web search, memory across conversations, file creation, code execution and extended thinking [8]. None of that was true of a free tier two years ago. If your mental model of “free AI” is the one you formed when you signed up, it is out of date, and the seat you bought to escape it is being renewed on a stale premise.

What gates you now is a capability, not a volume

Once the message counter goes away, the honest question changes shape. It is no longer how much you use the tool. It is which specific thing you cannot do.

The limits that remain are attached to features, not to conversation length. OpenAI’s usage-limits page lists file uploads, image generation, voice and data analysis as separately capped for free users [2]. That is a useful list, because it tells you what a downgrade would actually cost you. If your week is drafting, summarising, rewriting, thinking out loud and asking questions, none of those caps touch you. If your week involves feeding the model 30 client PDFs, or generating images for every post, or living in voice mode on the commute, you will hit a wall on day two, and the wall will be a feature wall rather than a volume wall.

One capability people still assume is paywalled is reasoning depth. It is not, at least not on ChatGPT: the same August change gave free users a “Think” button for questions that need more than a fast answer [1]. If you have been holding a seat so that hard questions get the careful model, check whether that is still what you are paying for, because on this product it has moved below the paywall.

This distinction matters because it changes what you are buying. A volume problem is solved by any bigger plan. A capability problem is solved by one specific plan, often not the most expensive one, and sometimes by a different vendor’s free tier instead. Before you renew, write down the last three times the tool stopped being useful and what exactly stopped. “I ran out” is not an answer you can shop with. “It refused the fourth spreadsheet of the day” is.

Price the tier you need, not the one at the top of the page

Pricing pages are built to move you to the middle-right column. The middle-left columns are where most small operators actually belong, and almost nobody reads them.

ChatGPT Plus is “$20/month (billed monthly)” [3]. Below it sits ChatGPT Go, which OpenAI launched worldwide at “$8 per month” in the US, pitched at the time as “10x more messages, file uploads and image creation than the free tier” [4]. Google’s ladder runs from the free plan through Google AI Plus at “$4.99 / month” with “2x higher usage limits than Free”, then Google AI Pro at “$19.99 / month” with “4x higher usage limits than Free”, then AI Ultra at “Starting at: $99.99/month” [7]. Claude runs Free, then Pro at “$17 per month with annual subscription discount ($200 billed up front)” or “$20 if billed monthly”, then Max at “From $100” per month [8].

Notice what the Google rungs are actually selling. The free plan already lists varying access to 3.1 Pro, so the strong model is not strictly behind the paywall here; what the Pro rung adds is 4x the usage limits plus features such as video generation [7]. If you assumed you were paying for the model itself, the ladder is not built the way you think it is.

Read those ladders against the wall you wrote down. A person who occasionally needs more file uploads is being sold a $20 plan when an $8 or $4.99 one clears the wall. A person who needs the strongest reasoning model on a daily basis is not going to be rescued by the cheap rung and should stop pretending otherwise. The mistake in both directions is the same: buying a tier by its position in the layout rather than by the constraint it removes.

There is a third option that the ladder hides, which is holding two free tiers instead of one paid one. The caps sit on different features at different vendors. Image generation is capped for free ChatGPT users [2] and included in Google’s free plan, which lists image generation and editing, Deep Research and Gemini Live at $0 with a Google account [7]. Claude’s free plan includes file creation and code execution [8]. A solo operator whose only wall is a weekly image or an occasional long research pass can often clear it by opening a second free account rather than upgrading the first one. The cost is not money, it is the friction of context living in two places, which is a real cost and worth weighing honestly rather than dismissing.

Two of those prices also carry a billing decision. Claude Pro is $3 a month cheaper if you commit to a year and pay $200 up front [8]. Annual billing on a product category that reprices every few months is a bet that the plan you chose in September is still the right plan in March. Sometimes it is. It is still a bet, and it is worth naming as one rather than clicking the discounted option by reflex.

The training default is a reason to pay that has nothing to do with limits

There is a second reason to hold a paid plan, and it survives every free-tier expansion, because it is contractual rather than technical.

OpenAI states that “when you use our services for individuals such as ChatGPT and Codex, we may use your content to train our models”, and says you can “opt out of training through our privacy portal” [6]. The business tiers are the other way round. The ChatGPT Business page lists “No training on your business data by default” as a feature of the plan, priced at “$20 / user / month” billed annually, or $25 per month billed monthly, with a minimum of 2 users [5].

If you handle other people’s material, client files, patient-adjacent notes, unreleased financials, contracts under NDA, then the default matters more than the message cap ever did, and it is the thing you are actually buying. Note that this is a different purchase from the one most people think they are making. It also has a free component: OpenAI says that once you opt out, “new conversations will not be used to train our models” [6], which closes part of the gap without a payment, though it does not give you the administrative controls, the seat management or the written default that a business plan does. Decide which of those three you need before you decide what to spend.

Run the test before the renewal date, not after it

The check is cheap and almost nobody does it, because it requires remembering a date. Put a calendar reminder 2 weeks before each AI renewal, and use those 2 weeks as the test window.

During the window, do your normal work on the free tier of the product you pay for, and keep one line of notes each time you hit something you cannot do. Note the capability, not the frustration. At the end of the fortnight you will have one of three lists. An empty list means you have been paying for a counter that no longer exists, and you can drop to free or to the cheapest rung. A short list clustered on one capability means you should buy the specific tier that removes it, which is often below the one you hold. A long and varied list means the paid seat was correct, you now know why, and the renewal is a decision rather than a default.

Do the test on real jobs. A sample question will always work on a free tier, which is exactly why sample questions tell you nothing. And run it per person rather than per company. If one or two people live in the tool all day while everyone else opens it twice a week to tidy an email, those are two different purchases, and buying the same seat for both is a quiet way to overspend on this category.

What still goes wrong

The accuracy figures that make a free tier look sufficient are almost always the vendor’s own. OpenAI’s 62% and 68% reductions in factual errors come from its internal evaluations [1], and every lab publishes numbers on tests it selected. Those figures are a reasonable input to a purchasing decision and a poor substitute for checking the output on your own work, which is why the two-week test is the part of this guide that actually does the work.

“Unlimited” is a policy, not a guarantee. The wording OpenAI uses is “unlimited everyday text chats, subject to abuse-prevention safeguards” [2], and what counts as a safeguard is not published. Free tiers that expand under competitive pressure can contract when that pressure changes, and none of these vendors owes you notice. A downgrade also has switching costs the pricing page does not mention: on some products your projects, saved chats and custom setups live on the tier you leave, and getting them back means paying again. Check what your own account keeps before you cancel anything, not after.

Finally, every price and limit in this guide was read off a vendor page on 4 September 2026 and will drift. That is the durable point rather than a disclaimer. The tiers are a moving target, the free-versus-paid line moved twice in the past year, and the only defence is a recurring reminder and 20 minutes. Treat the numbers here as an example of how to read a pricing ladder, and go read the current one yourself before you sign anything.