GPT-6 Goes Free, Safety Scores Slip
OpenAI deployed GPT-6 to all ChatGPT tiers with interactive UI, but its own safety report shows regressions in self-harm, gore, sexual content and teen emotional dependency benchmarks.
ImportanceStructuralEvidenceE3 inspectableWrite-upDeep
GPT-6 began rolling out to paid ChatGPT users on October 7 and to free and Go users on October 8, replacing GPT-5.6 Luna. OpenAI's announcement is dated October 7.
Previously GPT-6 was paid-only since last month's launch. The update introduces Intelligent UI: the model dynamically generates charts, calculators and interactive components, turning the chat box from static text into a variable interface. OpenAI says GPT-6 Instant starts answering 44% sooner on average for search-dependent questions.
The same-day 24-page safety report shows improved jailbreak and instruction-hierarchy defense (Sol resistance at 99.99%), but statistically significant regressions in self-harm, gore, sexual content and teen emotional dependency evaluations. Luna's emotional-dependency score dropped from 0.927 to 0.734. OpenAI attributes the drops to increased willingness to answer informational questions on sensitive topics, notes the benchmarks use deliberately difficult samples, and states product-level classifiers are not included in the reported numbers.
All safety data comes from OpenAI's own system card with no independent reproduction yet. Work and Codex models do not switch with this release.