GPT-6 Teen Safety Scores Slip as Luna Reaches Free ChatGPT Users

What happened

OpenAI has begun putting GPT-6 in front of its largest audience. The rollout of GPT-6 with Intelligent UI started reaching ChatGPT Free and Go users on Thursday. The model's own documentation gives reason to pause. OpenAI's October 7 system card records statistically significant regressions on the company's safety tests for users under 18. 1

The sharpest figure involves sexual content. GPT-6 Luna is the variant served to Free and Go users. It scored 0.904 on OpenAI's under-18 sexual-content evaluation. GPT-5.6 Luna, released in August, scored 0.984 on the same test. Higher scores mean safer responses. 1 The card also notes regressions on emotional reliance, the measure of whether a model encourages unhealthy dependence on the chatbot. 12

The stakes are large because of reach. OpenAI says more than 1.2 billion people use ChatGPT each week. The system card says ChatGPT applies stricter provisions to users it believes are under 18. 1 Those protections only work as well as the model that enforces them.

One name, several builds

A second account adds an important complication. According to that reporting, the GPT-6 Sol and Luna versions pushed to ChatGPT's Chat tab on October 7 are not the same builds still running in the Work and Codex surfaces. 2

The Chat snapshot carries regressions beyond the teen-specific ones:

• Self-harm: a statistically significant regression on OpenAI's standard self-harm evaluations, not only the under-18 versions. 2
• Teen categories: regressions on age-restricted content, sexual content and emotional reliance for under-18 users. 2
• Health: a slight dip on HealthBench Hard compared with August's GPT-5.6 Sol. 2

The argument from this angle is that a model name can mask different safety profiles depending on where it is deployed. Anyone treating the October 7 ChatGPT launch as a clean capability upgrade may be missing that. 2

Where the accounts overlap and diverge

Both reports rest on the same document, OpenAI's October system card. Both agree it shows measurable backsliding on protections for minors, including emotional reliance and sexual content. 12

The emphasis differs. The first account centers on Luna, the lighter model now reaching free-tier users, and puts the sexual-content score drop in the headline. 1 The second focuses on how builds are split across ChatGPT surfaces. It singles out the self-harm regression on standard evaluations, which affects all users, not just teens. 2

The dates also differ slightly. One cites October 7 as the date of the Chat-tab update and the system card. 21 The other describes the free-tier rollout as beginning Thursday, October 8. 1 Read together, this looks like a staged deployment: paid surfaces first, then a broader push to Free and Go accounts a day later.

Why it matters

The numbers themselves need context. A drop from 0.984 to 0.904 can look small. But on an evaluation meant to catch failures with minors, it means substantially more unsafe responses in the tested set. Applied to a user base above a billion people a week, even a few percentage points can mean very large absolute numbers of exposures. 1

The two choices are hard to separate. OpenAI documented these regressions itself and then shipped the model widely anyway, to the free tier, where younger users are arguably most likely to be. One charitable reading is that the company judged the capability and interface gains worth the trade-off, perhaps paired with product-level safeguards outside the model. A less charitable reading is that the release schedule outran the safety work. The available reporting does not settle which is closer to the truth.

The build-splitting point may matter more over time. If "GPT-6" in Chat behaves differently from "GPT-6" in Work or Codex, then benchmark claims, third-party audits and policy discussions that cite the model name alone become harder to interpret. 2 Regulators and researchers will need to ask which snapshot was tested, on which surface, and on what date.

The takeaway

OpenAI deserves some credit for publishing a system card that plainly shows regressions. Many companies would not. But disclosure is not mitigation. The most defensible reading is this: GPT-6 Luna reached free users carrying weaker teen protections than its predecessor, by OpenAI's own measures. The Chat-tab build also slipped on self-harm handling more broadly. 12 Until OpenAI explains what compensating safeguards exist, and closes the gap between same-named builds, "upgrade" is the wrong word for parents, schools and policymakers to accept at face value.