ChatGPT’s Unlimited Free Chats: What You Actually Get

OpenAI announced on 6 August 2026 that ChatGPT’s Free and Go tiers are losing their text message limit, replaced by unlimited text conversations running on GPT‑5.6 Luna.

OpenAI has never published the exact size of the cap it is removing, and outlets differ on it — estimates range from around ten messages every five hours to a somewhat higher allowance before users were downgraded to a lighter model. What is not in dispute is that free users regularly hit it.

That is a real change affecting a lot of people; OpenAI puts ChatGPT’s reach at a billion users a week. But the word doing the work in every headline is unlimited, and it is carrying more weight than it can bear.

What actually changes, and when

Unlimited means text, and only text

OpenAI is explicit that limits still apply to file uploads, images and other tools. So a free user can now talk to ChatGPT indefinitely, but cannot upload documents indefinitely, generate images indefinitely, or use the tool-based features indefinitely.

For a great many people that is fine — typing questions is what they use it for. It is worth knowing before you cancel a subscription on the strength of a headline. If you are consolidating onto one assistant, it is also a sensible moment to check which AI tools still hold access to your files and accounts.

The accuracy gap the coverage split in half

OpenAI published two accuracy figures, and most reports quoted one or the other without mentioning there were two. Here is the full sentence, from an internal evaluation of financial, medical and legal prompts requiring factual detail: 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, measured against GPT‑5.5 Instant.

Both models improved. Both improved substantially. But the gap between them is the point: Luna is what free users get, Sol is what Plus and Pro users get, and Sol is still the more accurate of the two on exactly the categories where accuracy matters most.

The Think button points at Luna’s weak spot

Free users also get a Think button, which OpenAI describes as giving Luna more time to work through harder questions. Plus and Pro users get a slider that does something similar across a range.

There is an awkwardness here worth naming. Luna is the smallest model in the GPT‑5.6 family, sitting below mid-tier Terra and flagship Sol — a hierarchy OpenAI does not spell out in this announcement, though it is consistent across its earlier GPT‑5.6 material and every outlet covering the family. Small models are generally weakest on tasks needing precise retrieval or multi-step reasoning — and “harder questions” is exactly what the Think button invites you to use it for.

More thinking time is not more capability. A model that lacks the underlying reasoning depth does not acquire it by taking longer. The Think button is a genuine improvement over no option at all, but the free tier’s ceiling is set by Luna, not by how long Luna is allowed to think.

The context that explains why

One fact makes the economics legible, and it is missing from almost every write-up: the free tier carries advertising. Multiple pricing trackers date the US rollout to 9 February 2026, with ads appearing beneath responses, and OpenAI’s own help documentation notes ads may appear in certain countries. OpenAI has since opened a self-serve ads manager.

Read the announcement against that and it stops looking purely like generosity. Unlimited conversations mean more sessions, longer sessions, and more ad impressions, delivered on the cheapest model OpenAI runs. Luna is the entry-level model precisely because it is the cheapest to serve at volume — OpenAI cut its API price by around 80% on 30 July, a week before this announcement.

The part almost nobody reported

A substantial section of OpenAI’s announcement, detailed further in its system card, covers new protections for users it believes are under 18. It went largely unmentioned in the coverage.

For those users, OpenAI says the model was trained to avoid romantic roleplay, avoid age-restricted challenges, and avoid presenting itself as a substitute for real-world relationships. It applied age-appropriate boundaries around sexual content, eating disorders and body-image risks, age-restricted goods, dangerous activities and graphic violence. The model is also trained to encourage connection with trusted people when a teenager may need support, reinforced with system-level protections and new evaluations.

Given that this update removes the message ceiling for the tier teenagers are most likely to be using, that section deserved more attention than it got.

My Take — Mr Wangdoo

Removing a rate limit for the free tier of a product a billion people use each week is a significant piece of access policy, and it is worth saying plainly that it is a good thing. Anyone who has hit a message cap mid-problem knows how abruptly it ends the usefulness of the tool.

What I would resist is the framing that free and paid have converged. The tiers are now differentiated by model quality rather than by message count, which is a subtler distinction and a harder one to notice. A free user gets endless conversations with the entry-level model. A paying user gets the flagship, better factual reliability on the prompts that carry consequences, and finer control over reasoning effort. That is arguably a sharper divide than a message cap, because a cap is obvious and a quality gap is not.

The ads detail is what makes the move legible. This is not a company giving away its best product; it is a company removing a friction point on an ad-supported tier running its cheapest model, one week after cutting that model’s serving costs. That is a perfectly rational business decision and it is still good for users. Both things are true, and the coverage that omits the advertising leaves readers with only half the picture.

The number I would most like to see is the one OpenAI has not published: the raw error rate. “62% fewer responses with at least one factual error” tells you the direction of travel but not the destination. If the baseline was poor, a 62% improvement can still leave a model you would not want answering a medical question — and medical prompts were part of the evaluation set.

Mr Wangdoo

Founder and editor-in-chief of Wangdoo.com. Covering AI, cybersecurity, EVs, smart home, and emerging tech from Dublin, Ireland. All opinions are documentation-based; nothing here has been hands-on tested unless explicitly stated.