153 agent runs raced to speed up nanoGPT — Fable 5 crushed it
Prime Intellect ran 153 autonomous coding-agent runs on the nanoGPT optimizer speedrun across 18 frontier models. Harnessed to claude-code, Fable 5 hit an 81.7% pass rate and finished in 8.7 days on one 800M-token training run — doubling the next-best Opus 5. The leaderboard is a rare apples-to-apples test of which agents can actually train models unsupervised, not just emit code.