GPT-6 rolls out amid teen-safety test regressions

OpenAI on Wednesday began moving all of ChatGPT to GPT-6. The same day, it published a safety document showing that the new models do worse than the ones they replace on several of its tests for users under 18. The company says additional safeguards, which those tests leave out, make the answers teens actually receive safer. Its report does not measure by how much.

According to OpenAI’s announcement, as reported by MacRumors, GPT-6 started reaching ChatGPT Plus, Pro, Business and Enterprise subscribers on Oct. 7 and expands to the Free and Go tiers on Oct. 8. Paid tiers get GPT-6 Sol; Free and Go get GPT-6 Luna. The change applies to the Chat tab; the models used for Work and Codex are not changing.

The headline feature is what OpenAI calls Intelligent UI. Instead of only text, answers can be built from graphics, buttons, forms, charts and other elements that the user can operate inside the conversation, and the model can start answering before it has finished thinking. The safety document says the new models replace GPT-5.6 Sol and GPT-5.6 Luna in ChatGPT on free and paid plans worldwide.

The report measures the October GPT-6 models against the August GPT-5.6 versions they replace. On OpenAI’s general production benchmarks, GPT-6 Sol shows a statistically significant regression on self-harm, and GPT-6 Luna on self-harm, gore and sexual content. OpenAI says it reviewed the failing responses by hand and judged them borderline but generally safe.

OpenAI also runs separate under-18 evaluations, which test the stricter rules it applies to users it believes may be teens. Its key finding there reads: “Relative to their respective GPT‑5.6 (August) counterparts, GPT‑6 Sol (October) and GPT‑6 Luna (October) show statistically significant regressions on age-restricted content, sexual content, and emotional reliance.” The document adds that Luna also regressed on gore. Its text does not put a size on these regressions.

The report also records gains. Both models improved on factuality across nearly all of OpenAI’s measures and on its mental-health benchmark. On health questions the picture is mixed: both scored higher than their predecessors on HealthBench Professional but lower on the standard HealthBench. On multi-turn jailbreaks, results were broadly in line with earlier versions.

OpenAI’s answer to the teen regressions rests on three points, all in the same document. For teens, it adds a classifier that blocks responses that may contain self-harm, sexual content or gore; OpenAI says this block is not reflected in the evaluation results and makes responses safer. The evaluations also leave out system-level protections such as Trusted Contact, local crisis helplines and Parental Controls. And on emotional reliance, OpenAI says its test overreacts when the model uses harmless nicknames such as “bro” or “bestie,” which it does not count as a violation when a user asks for them.

In a separate announcement on teens published the same day, OpenAI said accounts identified as under 18 get a teen experience by default, including reminders to take breaks, and that it expects to seek feedback on teen safety defaults, parental controls and notifications, and AI literacy. The same post previews a College Planner tool that is not yet available.

Set the regressions next to the safeguards, and two gaps appear.

The extra classifier for teens covers self-harm, sexual content and gore. That matches some of the teen regressions — sexual content in both models, gore in Luna. It does not cover the other two that both models share: age-restricted content and emotional reliance. OpenAI’s own teen policy lists access to age-restricted goods and services among the areas where it sets stricter limits for teens. For emotional reliance, OpenAI offers an explanation of the test, not a fix. For age-restricted content, the document offers neither an explanation nor a specific safeguard.

The second gap is in the measurement itself. For every safeguard it names, the document says the effect is outside the numbers. It does not publish results for the teen experience as actually deployed — model plus classifier plus product controls — which is the one teens will meet.

OpenAI published these regressions on the day of launch, in its own document, and that deserves credit. But the report leaves readers to trust that safeguards it did not measure make up for regressions it did. The natural next disclosure is the deployed system’s results on the same teen tests. Until then, whether those safeguards close the gap remains an open question, not a finding.