OpenAI "gpt-next" Leak: Astra Trained on 100k GPUs With Old Models Supervising
imjustnewatai · x · 2026-09-04
A roundup of leaked and confirmed OpenAI details: Astra is the first model whose training was significantly supervised by previous generations, using 100,000+ GPUs. Sam Altman confirmed Astra finished training a while ago; the paused run was "a future model," with large-scale RL restarting August 28. Hiring pages point to a "gpt-next" program whose final RL run integrates memory and multi-agent behavior, while the RSI team trains agents to run research experiments and turn failures into training data. Astra can already do entry-level AI research — likely what Altman means by "much, much, much more capable."
Related event: OpenAI's Next-Gen Model Astra Reportedly Trained on Over 100K GPUs(10 posts)→
More from Models
- GPT-6 Astra tops Andon Labs' Vending-Bench, unseating Fable — Outside-Iron-8242 · 2026-09-09
- OpenAI: training new internal model with 'unprecedented' math benchmark performance — TheZvi · 2026-09-09
- DeepSeek v4.1 Flash flash sale: 58M tokens for $1, available for 2 days only — MicahBerkley · 2026-09-09
- OpenAI's Noam Brown: in a year, everyone will have AI that can tackle Millennium Prize Problems — Neurogence · 2026-09-09
- Aerospace professor pours cold water on 'AI solved Navier-Stokes' claims — Mindrust · 2026-09-09
- Dev loves Google's Astra for coding while others ditch it: model preferences puzzle labs — lucasmeijer · 2026-09-09