ChatGPT just went unlimited for free users, and most businesses have not thought it through

On 6 August, OpenAI announced a change that will reach more people than most technology news ever does. Free ChatGPT accounts now default to its GPT-5.6 Luna model and come with unlimited text chats, plus a button that lets anyone reach for heavier reasoning on a harder question. Paying subscribers got their own update, with a slider that controls how much thought the model puts into each answer.

OpenAI says around a billion people use ChatGPT every week. So this is not a niche product tweak. It is a very large number of people who previously hit a wall partway through the day, and who now do not.

For anyone running a business, that is worth five minutes of attention, though probably not for the reason the announcement suggests.

The number worth noticing

Buried in the announcement is the most useful figure OpenAI published. In an internal evaluation using financial, medical and legal prompts that required factual detail, answers 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 the older GPT-5.5 Instant.

That is a genuine improvement, and it deserves to be recognised as one. It is also, read carefully, an admission worth sitting with.

A 68% reduction in answers containing a factual error is not the same as an accurate model. It means errors have become roughly three times rarer. It does not mean they announce themselves. The tone of a wrong answer is identical to the tone of a right one, which is the single most important thing to understand about this technology and the thing least likely to appear in a product launch.

In practice, the improvement makes the failure mode harder to catch, not easier. When a tool is wrong one time in three, people check it. When it is wrong one time in ten, they stop checking, and the one gets through.

What actually changes on Monday morning

Here is the part that has not been thought through in most workplaces.

Until now, the free tier had a natural brake on it. Staff who used ChatGPT for work hit their limit, got mildly annoyed, and went back to doing the task themselves. That limit was doing quiet policy work that nobody had written down.

Remove it, and the behaviour changes. The person in accounts who used to paste in one tricky email now pastes in twenty. The casual who was drafting social copy is now drafting customer responses. None of this is authorised, none of it is logged, and almost none of it is malicious. It is people trying to get through their day with a tool that is now free and no longer runs out.

The industry has a name for this, shadow AI, and the numbers behind it are consistently higher than management expects, for the obvious reason that nobody puts their hand up to report it.

The risk is not that the tool is bad. It is that customer details, supplier pricing, draft contracts and staff information are going into a system that the business has no agreement with, no record of, and no ability to audit. If a client asked tomorrow where their data had been, most Australian businesses could not answer honestly.

The wrong response, and the right one

The instinctive reaction is to ban it. That reaction fails every time, and it fails in a specific way: usage does not stop, it moves to personal phones where there is no visibility at all. A ban converts a manageable problem into an invisible one.

The response that works is duller. Decide what may go in and what may not, write it in plain language on one page, and tell people. Most staff follow a rule they understand. Very few follow a rule they have never been given.

A workable version of that page fits in three lines. Nothing that identifies a customer, a patient or an employee. Nothing under a confidentiality agreement, which includes supplier pricing and unsigned contracts. Nothing that goes out to a customer without a person reading it first.

Then provide a sanctioned option. A business account with the same vendor keeps the work inside something the organisation controls, and costs less than the first incident would. The point is not to stop people using the tool. It is to give them a version of it that does not create an exposure nobody knew about.

Where it genuinely earns its place

None of the above is an argument against using this technology. The tools have become properly useful, and businesses that avoid them entirely are giving up real time for no benefit.

The pattern that works is narrow and unglamorous. It handles the repetitive, high-volume work with a knowable right answer: turning notes into a structured summary, sorting incoming enquiries so the urgent ones surface first, drafting the first version of something a person will then edit, pulling numbers out of documents that somebody currently retypes.

The pattern that fails is anything involving judgement, money leaving the business, or a commitment to a customer, unless a named person signs it off before it goes anywhere. That line has not moved despite the improvements, and the error statistics above are the reason.

Businesses that get value out of this tend to do one process properly and leave it running for a month before touching anything else. The ones that buy a platform and try to automate six things at once usually have nothing running by Christmas.

Customer enquiries are the usual place to start, because the volume is high and the correct answers already sit on the business’s own website. The tools differ enormously on the one feature that actually matters there, which is whether a person can read a response before the customer does, so a comparison of the AI chatbots available in Australia is a better starting point than any vendor’s own feature list.

The honest summary

An unlimited free tier on the world’s most used AI assistant is genuinely good news for the people using it. It is also, for any organisation employing those people, a policy question that arrived without warning and will not announce itself.

Nothing about this requires a project, a budget or a consultant. It requires somebody to spend an hour this week writing down what may and may not be pasted into a chatbot, and then telling the team. That hour is the highest return work available to most small businesses right now, and it will stay undone in most of them.

The tools got better this week. The habits did not. That gap is where the trouble lives.