AI models benchmarked themselves to squeeze maximum speed from nanoGPT

AI agents can now optimize neural network training code entirely on their own.

Researchers tested 18 frontier models across 153 autonomous runs on the nanoGPT optimizer speedrun benchmark. Equipped with coding harnesses like claude-code and codex, the models modified training code to maximize efficiency. Fable 5 posted the fastest validated result at 2,726, closely followed by Opus 5 and Kimi K3.

Why it matters: Squeezing efficiency out of code usually requires engineers to manually hunt down bottleneck operations. Automating that loop means models can iteratively tune their own training scripts on low-cost setups before scaling up.

Know this: The benchmark dataset includes 41 open traces detailing tool calls, subagent actions, and scratchpad notes. The most aggressive runs burned through up to 4.6 million tokens while testing code variations.

If AI models keep finding new ways to streamline their own training code, standard consumer hardware could stretch a lot further.