Leaked OpenAI table shows GPT-6.2 Astra saves 42.87% training time with NanoGPT

scaling01 · x · 2026-11-18

A leaked OpenAI internal eval table (Table 5) shared by scaling01 suggests OpenAI is heavily using NanoGPT for research engineering and model training. GPT-6.2 Astra saves 42.87% of baseline training time on NanoGPT, far above GPT-5.6 Sol's 9.69% and GPT-6 Astra's 7.73%. The table also covers Internal Research Debugging, KernelGen 1P, PostTrainBench Lite, and MLE-Bench Revised, where GPT-6.2 Astra leads on most metrics (96.06 percentile on MLE-Bench Revised, 81.85% mean rubric on research debugging). The poster quips that OpenAI is 'NanoGPT maxxing,' implying NanoGPT-style training has become its internal standard toolchain.

Original post →

More from Models

Models channel →