GPT-6 reaches ChatGPT with visual answers and disclosed safety regressions

Answers can include charts, forms and interactive explanations. The update covers Chat, while Work and Codex retain their earlier models; Enterprise access also depends on administrator settings. [1][3]

OpenAI’s October system card reports statistically significant deterioration against GPT-5.6 on standard self-harm tests for Sol, and self-harm, gore and sexual-content tests for Luna. It also reports stronger resistance to jailbreaks. These are vendor-reported findings, independently reported but not independently reproduced. Difficult tests do not estimate everyday incident rates; OpenAI judged reviewed failures low in severity and describes additional safeguards. [2][6]

Separately, Common Sense Media tested more than 4,000 prompts on teen accounts and rated the teen product an unacceptable risk. Its testers found no parental alerts during prolonged sensitive conversations on newly linked accounts. OpenAI disputes the notification testing, citing account-linking and technical delays. [4][5]

These are different kinds of evidence: a model-level company evaluation and an outside test of a teen product. The latter does not establish GPT-6’s performance. Our recommendation for families is to discuss use directly and review account protections together, rather than assuming that a newer model or a promised alert guarantees safety. For adults, try a useful visual explanation and check its underlying facts as carefully as prose.