NanoGPT - OpenAI self-improvement percentage
Code
Internal NanoGPT self-improvement result. Baseline, compute budget and formula are unreported; diagnostic rather than an assumed absolute language-model accuracy.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Terra | 14.5 |
| GPT-5.6 Sol | 9.7 |
| GPT-5.6 Luna | 1.7 |