GPT-6.2 System Card Adds First Model Spec Compliance Section With 738 Tests
sjgadler · x · 2026-11-20
OpenAI's GPT-6.2 system card includes, for the first time, a section measuring compliance with the Model Spec.
- Targeted Cases (738 compliance tests): GPT-6.2 Astra leads across all five Spec pillars — faithfulness to human intent, minimizing harm, honesty/transparency (94.7% passing), facilitating human agency, and acting with care — mostly above 90%.
- Conversation Replay (596 contexts): models from GPT-6 Astra onward show large generational gains over GPT-5.6, e.g. "express uncertainty" jumping from 53.6 to 91.2 on the rubric.
- The biggest improvements are in the newest human-flourishing-related Spec sections on attending to effects on others and supporting human agency, plus much better uncertainty handling.
More from Models
- LoRA adapter lifts Qwen3.8-27B to 63% on SWE-Bench Pro, 60% fewer tool calls — ben_burtenshaw · 2026-11-20
- GPT Pro veterans are cancelling over 6.2 quality and new Chat quotas — lxfater · 2026-11-20
- Model UX: why Google's writing style keeps users coming back — flavioAd · 2026-11-20
- Users report ChatGPT imagegen is getting noticeably less restrictive — flowersslop · 2026-11-20
- Mistral Claims Best Open Coding Duo, But Its Own Cited Benchmark Ranks Kimi 8 Points Ahead — julsimon · 2026-11-20
- Creating good evals may be the most important post-training job in coming years — joecole · 2026-11-20