NanoGPT Speedrun Frontier

The nanoGPT optimizer speedrun executed 153 autonomous runs across 18 frontier models, reporting each model’s top validated result and closure rate, such as Fable achieving 52,726 with an 81.7% closure and GPT-5.6 Sol reaching 3,042 with a 35.9% closure. An equal-budget comparison showed the best final runs attaining a record of 2,600 versus a baseline of 3,290, with 41 curated agent trajectories available for detailed analysis.