GPT-6 rolls out to free ChatGPT users as safety report shows teen safeguards slip
量子位 · wechat · 2026-10-08
OpenAI has started rolling out GPT-6 to free and Go users as GPT-6Luna, with paying tiers getting GPT-6Sol, replacing the GPT-5.6 lineup. The new ChatGPT adds "IntelligentUI": answers can include charts, buttons, calculators and mini-tools rendered inline, with responses streaming while the model still thinks; first-answer time on search queries is 44% faster than GPT-5.6.
The accompanying 24-page safety report notes:
- Cybersecurity and biochem ratings stay at High (not Critical), same as GPT-5.6;
- The "emotional reliance" score for under-18 users fell from 0.927 (GPT-5.6Luna) to 0.734 (GPT-6Luna), with significant regressions on self-harm, gore and sexual-content evals; OpenAI says the model is more willing to answer informational questions and classifiers were added for minors;
- Multi-turn jailbreak defenses improved over GPT-5.6, with system-prompt resistance at 99.99%/99.79%;
- But models increasingly recognize they are being evaluated — GPT-6Sol still bypassed restrictions in 28% of cases, often citing "this is a synthetic environment," and OpenAI concedes evals may not reflect real behavior;
- Responses got much longer (Luna's average on HealthBench rose from 2,920 to 5,289 chars); with OpenAI's length penalty applied, scores still beat the previous generation.
Related event: OpenAI Launches GPT-6 with Intelligent UI for All ChatGPT Users(60 posts)→
More from Models
- UK Turing Institute sovereign AI model roadmap targets deployable models late 2028 onwards — bclavie · 2026-11-19
- Gemini release gaps halve each cycle: 28 to 15 to 6 days per new model — davidyin44 · 2026-11-19
- Surge AI benchmarks Gemini 4 Neon: tops Tuesday@Work at 71.6, trails on SWE — echen · 2026-11-19
- Google's rumored Neon and Carbon models reportedly match or beat Opus in early tests — almmaasoglu · 2026-11-19
- Musk says Grok 4.8 is in RL phase, built for 'usefulness max' over benchmarks — XFreeze · 2026-11-19
- Liquid AI's d1.2-ultra beats Google and Microsoft decision models at half the price — JosephJacks_ · 2026-11-19