Development

Agents given idle compute beat the human speedrun baseline

The speedrun the corpus has tracked since December as a human sport is won here by agents, turning the leaderboard from a measure of researchers into a measure of the compute handed to models.

development2026-05-15rsispeedrunautonomous-researchai-r-and-d

Claim

Prime Intellect handed Codex and Claude Code its idle compute to attack the NanoGPT speedrun optimizer track, and after some 14,000 GPU-hours both agents beat the human baseline, with Opus 4.7 holding the record at 2,930 steps.

Reported metric: 14,000 H200 hours / 2,930 steps

Reporting date: (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Verification record

These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.

  • Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z

The speedrun has been in this corpus since December as a human sport. Prime Intellect let Codex and Claude Code iterate autonomously on the NanoGPT speedrun optimizer track for two weeks on idle H200s, roughly 10,000 runs and some 14,000 GPU-hours, after which both agents had beaten the human baseline and Claude Opus 4.7 held the record at 2,930 steps (write-up). The competitors have changed: every earlier record in the series, down to 86.8 seconds in March, was set by people, and the lab’s own report is as much about where autonomous research agents break down as where they win. It sits between Karpathy’s 650-experiment autoresearch loop of March and the rule-lawyering a July speedrun would expose as the next barrier.

Knowledge graphopen in browser →

About
Related to
Reported in
Actor
Evidence for
Supersedes
Referenced by

Cite or correct this record

Suggest a correction — opens a GitHub draft for you to review and submit.