OpenAI: GPT-6.2 Astra is first external model to hit "High" AI self-improvement
ChrisGPT · x · 2026-11-18
OpenAI says GPT-6.2 Astra is its first externally released model to reach the "High" threshold for AI self-improvement.
How it was evaluated
OpenAI tested whether the model can "accelerate AI research" across five research-engineering evaluations: debugging experiments, writing faster GPU kernels, speeding up pretraining, and improving an existing model through further training.
Highlight number
In one test, Astra cut a small language model's baseline training time by 42.87% on average while hitting target prediction quality — with two hours and a single H100.
Caveats
Astra 6.2 still fell below OpenAI's stated cutoffs on both the pretraining and post-training tests. Per the released document, OpenAI also improved research-environment isolation, significantly expanded frontier-agent monitoring, and introduced "Control Room," a central system that tracks model workloads and can stop them quickly on safety or security concerns.
More from Models
- Domingos on iWSP paper: just use the program instead of writing it into LLM weights — pmddomingos · 2026-11-18
- GPT-6.2 Astra and Sol post highest error rates ever on VoxelBench 2 — legit_api · 2026-11-18
- GPT-6.2 Astra hits new harmonic milestones on the Bach Benchmark — deanwball · 2026-11-18
- Early hands-on: GPT-6.2 wins many benchmarks and is surprisingly cheap — ramos_casals · 2026-11-18
- Chutes partners with Alibaba Cloud to offer GLM-5.3 and DeepSeek V4.1 at 35% off — markjeffrey · 2026-11-18
- Meta launches Muse Spark Open: 816B MoE coding model with 42B active params and 1M token context — PyTorch · 2026-11-18