Is GPT-6 Astra Safe for Kids? What a Parent Can Actually Know Right Now

OpenAI released GPT-6 Astra on 3 September 2026 as a limited preview and opened it to the public a day later. Your child may already be using it, because for most families it did not arrive as a new app — it arrived as the model behind the ChatGPT they already had.

Here is the honest position: nobody, including us, has had this model long enough to tell you it is safe for a nine-year-old. What exists is a set of published facts, one of which matters more for parents than anything in the launch coverage.

This is what is known, what is not, and what to do in the meantime.

Last verified: 5 September 2026. This is a two-day-old model; treat any confident safety verdict — ours or anyone's — with suspicion.

The facts, with sources

OpenAI's president, Greg Brockman, said the model could eventually be seen as the arrival of artificial general intelligence. Axios ran the launch under the headline "Welcome to the AGI era."

That framing is not why parents should pay attention. This is.

The one change that matters for a child

GPT-6 Astra uses a new reasoning technique OpenAI calls recurrent depth. Reported plainly: it works in a way that obscures some or all of the model's reasoning — what the industry calls its "chain of thought."

Researchers have raised concerns about this on monitorability grounds: a model whose reasoning cannot be inspected is harder to supervise.

For a parent, that lands somewhere specific. The single most useful thing you can teach a child about AI is to ask how did you get that? — to treat an answer as a claim with a path behind it, not a fact handed down. A model that shows less of its path makes that harder to practise.

This is not a reason to keep your child away from it. It is a reason to change where the checking happens: if the model shows less of its work, the work has to be checked against something outside the model. A second source. A textbook. You.

What we do not know, and will not pretend to

Anyone publishing a confident safety verdict on a model this new is guessing. Specifically unknown as of today:

• How its content filters behave with a child's phrasing. Filters are tuned per-model and change in the weeks after launch.
• Whether it hallucinates more or less than GPT-5.6 on the questions children actually ask — homework, animals, how things work. Launch benchmarks measure hard adult tasks.
• How the obscured reasoning behaves in practice when a child asks it to explain itself.
• Whether the parental and privacy controls changed. Model swaps sometimes reset per-account settings.

We will update this page as those become checkable. What we will not do is invent a verdict to be first.

What to do this week

Four things, none of which require knowing more than we do.

1 · Find out whether your child is already using it

Most families were switched over without doing anything. Open ChatGPT on their device and look at the model name in the interface. If it says GPT-6 or Astra, the change already happened.

2 · Re-run the settings check

A model change is a good prompt to re-verify what you set months ago: memory, chat history, voice, and whether notifications are off. Settings survive model swaps most of the time, and "most of the time" is why you look.

3 · Change the checking question

If you have taught "ask it how it worked that out", add a second habit: check one answer a week against something that is not an AI. A book, a school worksheet, a search that lands on a real source.

This is worth doing regardless of which model your child uses. It is more important now.

4 · Do not upgrade to a paid tier because of the launch

The capabilities OpenAI is advertising — cybersecurity, agentic computer use, professional software engineering — have nothing to do with a ten-year-old's homework. There is no child-facing reason to pay for this launch.

The age rule, stated plainly

OpenAI's terms require users to be at least 13, or the minimum age required in their country. Under 18s need a parent or guardian's permission.

This is a terms-of-service position, not a developmental one. It tells you who OpenAI will take responsibility for. It does not tell you whether your particular child is ready — that is a different question with different signals, and age is the weakest of them.

So: is it safe?

The honest answer for a two-day-old model is "the safety question has not changed as much as the launch coverage suggests, and one thing did get harder."

The risks that mattered last week — a confident wrong answer, a child treating output as fact, work that gets done without being learned — are the same risks this week. The obscured reasoning makes the second one slightly harder to catch, which is an argument for checking outside the model, not for avoiding it.

If your child was using ChatGPT safely on Tuesday, the same setup is a reasonable starting point today. Watch for what changes over the next month, not what was said at launch.

You now know what is established and what is not. The question underneath it is the one that outlasts this launch and the next one: how do you evaluate any new AI your child gets handed, in the first ten minutes, without waiting for someone else's verdict?