SAN FRANCISCO, 11 AUG 2026 — OpenAI has uncapped text messages on ChatGPT's free tier and put a better model behind it. The key word here is "text".

Free users and subscribers to the low-cost Go plan get GPT-5.6 Luna as their default model, replacing GPT-5.5 Instant, and unlimited text chats "subject to abuse guardrails". Paying Plus and Pro subscribers get an upgraded GPT-5.6 Sol and a thinking slider that replaces the old split between Instant and Thinking modes.

What is actually uncapped

Only typing and reading. Everything that costs OpenAI more than tokens keeps its limit.

Now unlimitedPure text conversations, for Free and Go, subject to abuse guardrails.
Still cappedFile uploads, images, voice chats, image generation and other tools.
New defaultGPT-5.6 Luna replaces GPT-5.5 Instant on Free and Go.
New controlA Think button, giving the model longer on harder questions.

That distinction matters because the old free-tier experience did not fail at the point people expected. You would use your allowance without noticing, and ChatGPT would quietly drop you to a smaller model. Answers got worse and nothing said so.

This removes the trapdoor for conversation, but not for the features people increasingly use: photographing a document, uploading a spreadsheet, or asking about a picture. Those allowances still run out.

The accuracy claim, and what it does not say

OpenAI says factual errors were 62 per cent less common with GPT-5.6 Luna than with GPT-5.5 Instant, and 68 per cent less common with the new Sol.

Two things about that number: it comes from OpenAI's own evaluation, not an independent benchmark. And 62 per cent fewer errors is not the same as few errors — it is a relative improvement against a baseline the company has not published in absolute terms. If the old model was wrong often, the new one can be 62 per cent better and still be wrong often.

The improvement is almost certainly real. It is simply not a number you can convert into "how likely is this answer to be right", which is the only question a reader actually has.

The test set is the interesting part

Those error rates were measured on prompts in three domains: financial, medical and legal, chosen because they demand factual detail.

To OpenAI's credit, the choice of test domains is both revealing and honest. These are the hardest areas to be wrong in, and precisely where a free-tier user is most exposed. Somebody who cannot afford a subscription is also, on average, somebody less likely to have a lawyer, an accountant or easy access to a doctor.

The people who just got this better, unlimited model are disproportionately the ones using it for questions where errors have consequences. For them, a confident wrong answer is worse than no answer at all, because it stops them from asking someone who knows.

Nothing in this release changes what a chatbot is. It is not a clinician, an adviser or a lawyer, and a 62 per cent reduction in errors on medical prompts still leaves a residue you would not accept from a professional.

The Think button is the useful part

The more useful day-to-day change might be a small one: a new button that lets the model spend longer on a hard question.

This is the first time free users have had an explicit handle on the speed-versus-care trade-off. Until now that decision was made for you, and made in favour of speed, because speed is cheaper to serve.

The practical advice is simple. If a question has a definite right answer that matters — a calculation, a date, a rule, a deadline — press it. If you are drafting, brainstorming or rephrasing, do not bother. The button costs you a few seconds and buys you the model's attention, and most people will under-use it because the default feels fast and fine.

Why give it away

Uncapping text is cheap in a way that uncapping images is not. Text is the least expensive thing a model does, and the marginal cost of another conversation is small next to the cost of generating a picture or processing an uploaded file.

So this is a generous move in the place where generosity is affordable, and it buys two things. It removes the moment where a free user hits a wall and considers a competitor, and it moves the paid pitch away from "you have run out" toward "you get better models and the expensive features". That is a healthier place for a subscription to sit, and a harder one to argue with.

If you use the free tier

Three things worth knowing this week.

Check which model you are on. The upgrade rolls out over days rather than at once, and the improvement described here only applies once Luna is actually serving you.

Expect the wall to have moved rather than gone. Long text conversations are fine now; a session where you upload several documents will still stop.

And treat the accuracy improvement as a reason to check more carefully, not less. A model that is wrong less often is a model whose errors are harder to notice, because the surrounding text has fewer obvious mistakes to put you on guard.

What to watch

Whether "subject to abuse guardrails" turns into a visible limit. Unlimited with an undisclosed throttle is a familiar shape, and the terms of that throttle have not been published.

Whether an independent evaluation reproduces the error-rate improvement. Vendor-run benchmarks on vendor-chosen prompts are a starting point, not a finding.

And whether the Think button survives contact with cost. It is the one part of this that makes free users more expensive to serve, and it is the part most easily quietly rationed.