We gave 18 frontier AIs the same multi-day coding task. The leader closed 81.7% of the gap; one produced no record.

Eighteen frontier AI models took the same multi-day engineering job in Prime Intellect's NanoGPT Speedrun Frontier benchmark. The leader closed 81.7% of the gap to the human record; one model produced no record at all.

The job: rewrite the training recipe for nanoGPT, a small open-source language model, until it hits a target validation loss of 3.28. Each model got up to 8 days, eight H200 GPUs, and a coding-agent "harness" (software such as Claude Code or Codex that edits files, runs training, and iterates on its own).

Fable 5, the leader, reached 2,726 steps, closing 81.7% of the gap to the 2,600-step human record. The next cluster, Opus 5 and the Kimi K3 variants, closed roughly half the gap. Opus 4.8 closed about 40%, and GLM 5.3 produced no qualifying record. The full leaderboard, traces, and record PRs are on GitHub.

The signal is the spread, not the rank. Prime Intellect's methodology post is explicit: no model invented a fundamentally new training method. The leaders leaned on classical optimizer work (preconditioning, caps/floors, learning-rate schedules, weight averaging), packaged in harnesses that could sustain the loop for hours or days. Different harnesses measure different things, so the leaderboard mixes model capability with tooling ergonomics.

The spread is the reader's check on vendor "AI can do research" claims: some can, some cannot, and the gap is wide.