IMPLICATOR
OpenAI began rolling out GPT-6 with Intelligent UI to ChatGPT Free and Go users Thursday, and its own October 7 system card recorded statistically significant regressions on tests for users under 18. GPT-6 Luna, the model Free and Go users get, scored 0.904 on the sexual-content test against 0.984 for August’s GPT-5.6 Luna, on a scale where higher means safer responses. OpenAI says more than 1.2 billion people use ChatGPT each week, and the card says ChatGPT applies stricter under-18 provisions to users it believes may be under 18.
What Changed
- GPT-6 began reaching ChatGPT Free and Go users Thursday, a day after Plus, Pro, Business and Enterprise users got GPT-6 Sol.
- OpenAI's October 7 system card records statistically significant regressions on its tests for users under 18; GPT-6 Luna scored 0.904 on sexual content against 0.984 for August's GPT-5.6 Luna.
- OpenAI says its emotional-reliance test overreacts to benign nicknames and that a teen classifier block is not captured in the results.
- Intelligent UI adds charts, buttons and small tools to answers; OpenAI has not said whether it raises token use.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
What the safety tests show
Paid Plus, Pro, Business and Enterprise users get GPT-6 Sol from Wednesday, October 7; Free and Go users get the lighter GPT-6 Luna.
The October 7 system card records statistically significant declines for both October models against their respective August GPT-5.6 versions on tests covering age-restricted content, sexual content and emotional reliance. GPT-6 Luna also regressed on gore.
In the same August-to-October comparison, the emotional-reliance score for users under 18 fell to 0.770 for GPT-6 Sol from 0.921 for GPT-5.6 Sol, and to 0.734 for GPT-6 Luna from 0.927 for GPT-5.6 Luna. In the adult production benchmarks, Luna regressed on self-harm, gore and sexual content; Sol regressed on self-harm. These are OpenAI’s own measurements from its system card.
OpenAI’s explanation
OpenAI says its emotional-reliance evaluation is overly sensitive to benign nicknames such as “bro” and “bestie,” which are allowed when users explicitly request them.
OpenAI says an additional classifier blocks responses for teens that may contain self-harm, sexual content or gore, and that this protection is absent from the reported evaluation results. The card’s only specific explanation for a regression concerns emotional reliance; for the sexual-content and gore declines, it points to the classifier block and the difficulty of the test set.
FREE · ABOUT FIVE MINUTES
Track what OpenAI's safety reports actually say.
Implicator cuts the flood to the developments worth your attention, with concise reporting on what changed and why it matters. Delivered every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
From San Francisco. No spam. Unsubscribe anytime.
The tests deliberately use difficult cases. OpenAI says their scores do not estimate how often these behaviors occur in ordinary use. Its review of adult benchmark violations found them borderline but generally safe, and the models are less likely to refuse harmless requests.
The card says both October models had higher observed rates of blocking attempts to bypass safeguards across multiple exchanges than GPT-5.6 Sol at every tested attacker budget. Both October models scored slightly below their September counterparts, with broadly overlapping 95% confidence intervals. The card reports fewer factual errors on OpenAI’s test sets across nearly all metrics.
What Intelligent UI does
Intelligent UI builds answers from a library of ready-made interface parts that the model assembles. A compiler makes the interface appear progressively as the answer is generated. ChatGPT chooses when to add visuals or controls and when to use plain text.
Examples include a lamb-roast planner whose ingredient quantities change with the guest count.
“We have a design team that spent a lot of time figuring out when adding a diagram, chart, or buttons adds value versus when it's starting to feel cluttered,” said Aarush Selvan, an OpenAI product manager.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
Holger Mueller, an analyst at Constellation Research, said: “This could lead to usability and not just the quality of the model and its answers becoming the deciding factor in which chatbots people want to use.”
Google introduced a similar interface with Gemini 3 last year.
Users can request fewer visuals or switch off “Layout and visuals” on the web, although some visual elements may still appear. OpenAI says the model’s design judgment needs work.
Speed and unanswered questions
In its October announcement, OpenAI said GPT-6 Instant starts answering web-search questions 44% sooner on average than GPT-5.6 Instant, measuring when an answer starts rather than when it finishes.
OpenAI has not said whether Intelligent UI increases token consumption. Its announcement also leaves unspecified how sources will appear when an answer takes the form of a chart or generated tool.
Frequently Asked Questions
Which GPT-6 model do free ChatGPT users get?
Free and Go users get GPT-6 Luna, the lighter model, starting Thursday. Paid Plus, Pro, Business and Enterprise users got GPT-6 Sol from Wednesday, October 7.
What did OpenAI's system card find on teen safety?
Compared with the August GPT-5.6 models, both October GPT-6 models showed statistically significant regressions on under-18 tests for age-restricted content, sexual content and emotional reliance. GPT-6 Luna also regressed on gore. Emotional reliance fell to 0.734 from 0.927 for Luna.
How does OpenAI explain the declines?
OpenAI says its emotional-reliance test is overly sensitive to benign nicknames such as "bro" and "bestie." It also says an additional classifier blocks teen responses on self-harm, sexual content and gore, and that this protection is not reflected in the reported results.
What is Intelligent UI?
It lets ChatGPT answer with charts, buttons, forms and small tools, assembled from a library of ready-made interface parts that appear progressively as the answer is generated. Users can request fewer visuals or switch off Layout and visuals on the web.
Who measured these results?
OpenAI did. The safety scores and the 44% faster-start figure for GPT-6 Instant are OpenAI's own measurements from its system card and announcement.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Free AI briefing · Weekdays
The AI stories that matter, sourced and explained.
Join the Morning Briefing. It goes out every weekday at 4:45 a.m. Pacific, with later sends for the East Coast, Berlin and Tokyo.
Free when you sign up: The Paperclip Compendium, our tested guide to running AI agents.
Free. Unsubscribe in one click.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
