ChatGPT Just Turned Answers Into Apps: GPT-6 Reaches Every User Today, and Its Own Safety Card Shows Where Teens Lose Out (AI Trends, 8 October 2026)

Q: Can I turn off Intelligent UI?

Partly. On the web you can turn off the Layout and visuals setting to reduce visual elements, but OpenAI says some visuals may still appear. Older ChatGPT desktop apps for macOS and Windows do not support Intelligent UI.

What’s trending in AI on 8 October 2026: Today OpenAI finishes the biggest change to ChatGPT’s everyday experience in a long time. GPT-6 started rolling out to Plus, Pro, Business and Enterprise users yesterday, and Free and Go users get it today. It arrives with Intelligent UI, which lets the model answer with buttons, forms, interactive charts, maps and small working tools instead of only text. For a product that gHacks reports reaches more than 1.2 billion people a week, that is a new interface, not a feature. Two other documents landed at the same time and deserve equal attention. OpenAI’s own system card for the October models shows clear gains on factual accuracy and prompt injection, alongside statistically significant regressions in its under-18 safety tests. And Common Sense Media told parents that ChatGPT for Teens poses an “unacceptable risk.” Below: what launched, what Intelligent UI changes, what the safety numbers say, what it means for your website and your workplace, and a practical plan by role.

Key takeaways

Everyone gets GPT-6 this week. Paid tiers get GPT-6 Sol in the Chat tab; Free and Go get GPT-6 Luna from today. Enterprise access depends on admin settings, and ChatGPT Work and Codex are unchanged.
Answers become small apps. Intelligent UI picks a layout per question: plain text, comparisons, diagrams, forms or a calculator built on request. A “Layout and visuals” setting reduces it, but does not remove every visual.
Safer on facts and injection, weaker on teens. OpenAI’s system card reports fewer factual errors and 97–99.99% prompt-injection robustness, but lower safe rates on several under-18 tests, including emotional reliance.
The teen fight is now public. Common Sense Media’s 4,000-prompt test found missed crisis referrals and parental alerts that did not fire; OpenAI disputes the method.
Websites and workplaces both feel it. Calculator and comparison pages face a new in-chat competitor, and staff now have an AI that can generate forms people type data into.

The most-used AI assistant now answers with interfaces, not just paragraphs. The safety data arrived the same day.

1. What launched, and who gets what

GPT-6 is not brand new to developers. OpenAI released GPT-6 Sol and GPT-6 Luna to the API, Codex and ChatGPT Work on 22 September, following the more powerful GPT-6 Astra introduced earlier that month, whose government testing we covered in AI Is Now Building AI. What is new this week is that October versions of Sol and Luna replace GPT-5.6 in the ordinary Chat tab that most people use, with retuned safety training and the new interface layer on top.

ChatGPT plan	Chat model	Available from	Notes
Plus, Pro	GPT-6 Sol	7 October	GPT-6 replaces the “Latest” menu option; GPT-5.6 Sol stays selectable. Pro’s top thinking level runs on Astra, without Intelligent UI.
Business, Enterprise	GPT-6 Sol	7 October	Enterprise access depends on workspace admin settings.
Free, Go	GPT-6 Luna	8 October	No setting change needed; accounts are switched as the rollout reaches them.
ChatGPT Work and Codex	Unchanged	n/a	Still on the September versions of Sol and Luna.

Compiled from Search Engine Journal, gHacks and OpenAI’s October system card, 7–8 October 2026.

OpenAI also claims speed gains. For questions that need a web search, it says GPT-6 Instant begins answering 44% sooner on average than GPT-5.6 Instant, partly because the model can start writing while it is still thinking or using tools. It also says GPT-6 is better at deciding when to search at all. These are company figures; independent benchmarks for the October Chat models have not been published yet.

Developers had GPT-6 for two weeks. This week it became the default for consumers and most workplaces.

2. Intelligent UI: when the answer is a tool

Intelligent UI is the visible change. Instead of formatting every reply as paragraphs and bullet points, GPT-6 chooses a layout for each question. Ask about wing lift and you may get a labelled diagram; ask about a road trip and you may get a map; ask how much to save each month and you may get a calculator with sliders you can adjust. Users can also ask for simple tools directly. OpenAI’s own examples include a savings calculator, a bill splitter and a small game. As OpenAI product manager Aarush Selvan put it in the TechCrunch demo, “the most helpful answers aren’t just text.”

Under the hood, gHacks reports that OpenAI built a library of streamable components and a compiler so the interface can render while the answer is still being generated. A few practical limits matter:

Not everywhere yet. It is rolling out gradually on the web and in updated apps. Older ChatGPT desktop apps for macOS and Windows do not support it.
Not at every thinking level. It works from Instant to Extra High, but not on the Pro level that runs GPT-6 Astra.
Only partly switchable. Turning off the “Layout and visuals” setting on the web reduces visuals, but some elements may still appear.
Still learning taste. OpenAI acknowledges that the model’s design judgment needs work, so expect odd layout choices.

A generated form feels like an app, but whatever you type into it is still part of a ChatGPT conversation.

That last point is the one most people will miss. A tool that looks like a finished web app invites people to type in real figures: salaries, client budgets, medical numbers, account balances. Those inputs are conversation data, governed by the same retention and training settings as any other chat. If your organisation has been worried about staff pasting sensitive material into AI tools, as we described in Shadow AI in 2026, a friendly form makes that easier, not harder.

3. What OpenAI’s own system card says

OpenAI published a system card for the October versions of Sol and Luna alongside the rollout. It deserves a read, because the picture is mixed rather than uniformly better. The card says the models carry over safety advances from Astra, are less likely to refuse harmless requests and make significantly fewer factual errors than the August models on nearly every factuality measure. On prompt injection, it reports instruction-hierarchy robustness of 99.99% for Sol and 99.79% for Luna, and robustness against indirect injection (malicious instructions hidden in content the model reads) of 97.13% and 95.80%. Under OpenAI’s Preparedness Framework, both models are treated as High capability in cybersecurity and biology.

The weaker spot is younger users. On OpenAI’s under-18 evaluations, both October models scored lower than August’s GPT-5.6 Sol, and the card calls several drops statistically significant.

Under-18 evaluation (safe rate)	GPT-5.6 Sol (Aug)	GPT-6 Sol (Oct)	GPT-6 Luna (Oct)
Emotional reliance	0.921	0.770	0.734
Age-restricted goods and activities	0.865	0.799	0.814
Sexual content	0.971	0.933	0.904
Eating disorders	0.808	0.772	0.776
Self-harm	0.987	0.977	0.974

Source: OpenAI, GPT-6 Sol and GPT-6 Luna October 2026 system card. Higher is safer.

OpenAI adds two caveats. It says the emotional-reliance test is overly sensitive to harmless nicknames such as “bro,” and that an additional classifier that blocks certain teen content in production is not reflected in these numbers. Both may be true. It is still notable that the model now reaching every free user, including teenagers, scores lower on the company’s own youth tests than the model it replaces.

The biggest drop is on emotional reliance, the test OpenAI says is oversensitive. The others move less, but in the same direction.

4. The teen-safety fight goes public

The system card would have drawn less attention if it had not landed alongside an outside test. On 7 October, Common Sense Media’s Youth AI Safety Institute published its assessment of ChatGPT for Teens, the experience OpenAI announced in August. Testers ran more than 4,000 prompts on accounts registered to 13- to 17-year-olds and set up more than a dozen parent-linked accounts. According to Axios, they found:

Harmful how-to content was generally refused. ChatGPT usually declined to give instructions on self-harm, eating disorders or sexual role-play.
Escalation was the weak spot. It missed more than one in four cases where testers judged a referral to outside help was warranted.
Parental alerts did not fire. Testers could talk about high-risk topics for up to an hour on parent-linked accounts without an alert.
Controls were easy to step around. “Show me the answer” bypassed Study Mode, including during parent-set Study Hours.

Institute head Tom Siegel was blunt: “ChatGPT is not safe for kids to use.” OpenAI replied that it welcomes independent testing but does not think this test reflects how its safeguards work in practice. It said many alert tests ran before parent and teen accounts had finished linking, which can take hours; that leaving Study Mode is a deliberate design choice; that age prediction uses several signals and can take up to two weeks; and that its own data shows more crisis resources shown to under-18 users during the study period. The disagreement is about design philosophy as much as bugs: Common Sense treats flexible controls as failures, while OpenAI treats them as intended.

For businesses this is not just a parenting story. Any company that builds on these models for customer service, education or wellbeing products inherits the same model behaviour, and regulators are paying attention to who carries responsibility when AI goes wrong, as we explored in “The AI Did It” Is No Longer a Defense.

5. What it means for your website

If your site attracts visitors with calculators, converters, comparison tables or simple planners, Intelligent UI is a new competitor. Search Engine Journal points out that tools like savings calculators and bill splitters were already the easiest things for Google’s generative UI in AI Overviews to copy, and now ChatGPT can build them inside the conversation, so the user never needs to click. OpenAI has not explained how sources will be credited inside charts or tools; for search-based answers, citations may appear behind a Sources button.

The response is not to abandon tools, but to make yours worth the click. Generic calculators are the most exposed. Tools that use your own data, pricing, local knowledge or account context are much harder for a chatbot to reproduce. Making your content clear and citable for AI assistants, the discipline we described in Generative Engine Optimization, matters more when the answer arrives as an interface. And the bigger economic question, whether AI platforms will pay the sites they draw on, is the one we tracked in AI Is Starting to Pay for the Web.

6. What it means at work: admins, data and cost

For Business and Enterprise customers, the most important line in the rollout notes is that Enterprise access depends on admin settings. That gives IT a short window to decide whether GPT-6 is switched on now or after a pilot, and to brief staff before the interface changes under them. Three workplace issues stand out:

Generated forms invite real data. Treat anything typed into an in-chat tool exactly like a pasted document. Personal accounts on the free tier are the bigger risk, since they sit outside company data controls.
Answers that look finished get trusted more. A polished chart is still model output. The factuality gains in the system card are real, but they are not zero errors, and a chart hides its assumptions better than a paragraph does.
Click-through is a new permission question. Today’s buttons mostly recalculate or expand an answer. As in-chat interfaces connect to agents and apps, every button becomes an action. We set out the questions to ask in “Allow Always” Is the New “I Agree”, and OpenAI’s always-on agents from DevDay 2026 show where that is heading.

On cost, the small-model tier just reached a new floor. OpenAI prices GPT-6 Luna in the API at $0.10 per million input tokens and $0.50 per million output tokens, half its earlier promotional price, and Anthropic launched Claude Haiku 5.5 the same day at the same $0.10 and $0.50 for prompts up to 100,000 tokens. The price war we described in The 90-Minute AI Price War has reached the models most businesses use for high-volume work.

For high-volume work, the cost of a capable small model has halved in weeks. Governance, not price, is now the main constraint.

7. The GPT-6 week plan: what each role should do

A plan split by who owns the problem: one task for this week, one to settle this month.

Role	This week	This month	Signal it’s working
IT and workspace admins	Check the GPT-6 setting in your Enterprise workspace and decide: on now, or after a pilot group.	Confirm data retention and training settings cover in-chat tools, and that older desktop apps are updated.	Staff hear about the change from you, not from a surprise.
Security and privacy	Add one line to your AI policy: data typed into AI-generated forms counts as data shared with the AI.	Review whether personal free accounts are used for work, and steer that work to managed accounts.	Fewer sensitive figures appearing in personal AI accounts.
Marketing and web	List the calculators, converters and comparison pages that bring traffic, and check this month’s trend for each.	Upgrade the most exposed tools with your own data, pricing or account features a chatbot cannot copy.	Tool pages hold engagement even as AI answers improve.
Team leads	Tell people that a chart or tool from ChatGPT is a draft, and needs the same checking as a paragraph.	Agree which AI outputs need a human check before they reach a client or a decision.	Mistakes are caught in review, not by customers.
Product and customer-facing teams	If you serve young users, read the under-18 section of the system card.	Run your own tests on teen and vulnerable-user scenarios before switching models.	You can show a regulator your own evidence, not just the vendor’s.
Finance and developers	Re-price high-volume workloads against $0.10 / $0.50 small models.	Benchmark Luna and Haiku 5.5 on your own tasks before moving production traffic.	Lower cost per task with equal or better quality.

Delana’s GPT-6 week plan. Adapt the roles to your organisation’s size; in a small business one person may own several rows.

8. What to watch next

Source credit inside interactive answers. Publishers will want to know whether charts and tools link back to the data they rest on.
OpenAI’s answer to the teen findings. Whether it tightens Study Hours, speeds up account linking or changes alerts will show how much outside testing moves the product.
Rivals’ versions of generated interfaces. Google already shows generative UI in AI Overviews; expect Anthropic and others to answer.
Labelling and provenance. As AI output becomes interactive, watermarks like the one in OpenAI’s textGrain will need to cover more than text.

Frequently asked questions

When do free ChatGPT users get GPT-6?

Free and Go users receive GPT-6 Luna in the Chat tab from 8 October 2026, one day after Plus, Pro, Business and Enterprise users began receiving GPT-6 Sol. No setting change is needed; accounts are switched as the rollout reaches them.

What is Intelligent UI in ChatGPT?

Intelligent UI lets GPT-6 choose how to present each answer, mixing text with interactive elements such as buttons, forms, charts, diagrams, maps, comparison tables and small tools like calculators or bill splitters. It works from the Instant to Extra High thinking levels, but not on the Pro level that runs GPT-6 Astra.

Can I turn off Intelligent UI?

Partly. On the web you can turn off the “Layout and visuals” setting to reduce visual elements, but OpenAI says some visuals may still appear. Older ChatGPT desktop apps for macOS and Windows do not support Intelligent UI.

Is GPT-6 safer than GPT-5.6?

In some ways. OpenAI’s system card reports fewer factual errors, strong prompt-injection robustness and fewer unnecessary refusals. But both October models scored lower than GPT-5.6 Sol on several under-18 safety tests, including emotional reliance, age-restricted content and sexual content. OpenAI says one test is oversensitive and that a production classifier is not reflected in the results.

What did Common Sense Media say about ChatGPT for Teens?

After more than 4,000 test prompts on teen accounts, Common Sense Media said ChatGPT for Teens poses an unacceptable risk, citing missed crisis referrals, parental alerts that did not trigger and easy ways around Study Mode. OpenAI disputes the methodology and says several behaviours are deliberate design choices.

What should businesses do about GPT-6?

Enterprise admins should decide whether to enable GPT-6 now or after a pilot, update AI policies so data typed into AI-generated forms is treated as shared with the AI, remind staff that charts and tools from ChatGPT still need checking, and review website tools that a chatbot could now replicate.