OpenAI upgrades ChatGPT for paying users, hands free accounts unlimited chats

Key points

• OpenAI is rolling out GPT-5.6 Sol to ChatGPT Plus and Pro subscribers, and GPT-5.6 Luna to Free and Go users this week.
• OpenAI's internal tests found GPT-5.6 Sol produced 68% fewer responses containing at least one factual error than GPT-5.5 Instant, with Luna cutting them by 62%.
• Free and Go users will get unlimited text chats starting next week, though file uploads and image generation remain capped.
• A new slider lets users dial reasoning effort from Instant to High, with High potentially taking minutes to answer.
• Stricter content rules will apply for users believed to be under 18, covering romantic roleplay, graphic violence, and several other categories.

OpenAI is pushing a new version of ChatGPT this week, and for once the interesting bit isn't the marketing. Paid users on Plus and Pro get GPT-5.6 Sol. Free and Go users get GPT-5.6 Luna, and starting next week, unlimited text chats with it.

That second part matters more than the model name.

What actually changed?

The headline change is reliability: fewer confidently wrong answers. OpenAI says GPT-5.6 Sol produces 68% fewer responses containing at least one factual error compared to GPT-5.5 Instant, with Luna close behind at 62%, based on the company's own internal evaluations. Apply the usual scepticism you'd give any vendor's self-reported benchmarks.

There's also a new intelligence slider, first observed by BleepingComputer, letting users pick how hard the model thinks before answering. Instant returns a quick reply; High can take several minutes for something more considered. Free users don't get the slider but they do get a Think button that does roughly the same job for one question at a time.

The model is also supposed to stop padding answers. Ask who the US president is and you get one line. Ask about the mechanics of US elections and you get proper context. Sounds obvious in a blog post. Quietly regresses two updates later.

Why should security teams care about a chatbot update?

Because unlimited free chats change staff behaviour, and staff behaviour is where data leakage starts. Rate-limited tools make people cautious about what they paste. Remove the limit and that caution goes with it.

We've been tracking this gap in our AI safety tools piece from 4 August. The failure mode is familiar: an employee pastes a customer list, a snippet of source code, or an internal contract into ChatGPT for a quick summary. That content leaves your environment. Depending on account settings it may or may not feed training, but it's no longer under your control. Samsung learned this the hard way in 2023, and the lesson hasn't stuck.

Fewer factual errors is a real improvement for end users. It's not a data-loss prevention control. A more confident-sounding wrong answer on a legal or medical question is arguably worse than an obviously bad one, because staff are more likely to act on it without checking.

What about younger users?

OpenAI is adding stricter boundaries for accounts it believes belong to under-18s. The new limits cover romantic roleplay, sexual content, eating disorders, age-restricted goods, graphic violence, and related areas. Age assurance on consumer chatbots is still rough science, and the wording "users believed to be under 18" tells you how confident OpenAI is in its own detection.

What should organisations actually do?

Treat ChatGPT as an outbound data channel and log it like one. Put an AI usage policy in front of staff that names specific data types they must not paste in: customer records, credentials, source code, board material, anything under NDA. Then check whether your existing DLP, the data-loss-prevention tooling that spots sensitive information leaving the network, actually inspects traffic to chat.openai.com and the mobile apps. Most default deployments don't.

The post-mortem will say it plainly when the leak happens: we knew staff were using it, we just didn't know what they were pasting.