Overview
- OpenAI released GPT‑6.2 Astra and a lighter variant, GPT‑6.2 Sol, on Tuesday, Nov. 17, and Microsoft began rolling both models into Copilot for eligible customers.
- OpenAI disclosed that its Codex coding assistant contributed concrete ideas used in Astra, which the company says represents the first released model to cross its internal “High” level for self‑improvement.
- Independent evaluators reported clear gains for Astra on 3D and robotics benchmarks, including a top AA‑3D score, 79/100 on AA‑Robotics v0.1, and 41.2% on Ember‑0, while noting the model still lags on some spatial‑reasoning tests and in real‑time speed.
- On Nov. 20 OpenAI detailed new API‑level protections that tie a model’s protected reasoning to chat history and said internal red‑team tests reduced successful extraction of reasoning traces from 54.0% to 0.2%.
- OpenAI said researchers now lean heavily on agents — agents completed about 46% of 2–4 hour tasks in October — and the company’s push to automate more research raises fresh questions about oversight, staffing and the cost of running physical and long‑running tasks.