Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-23 · high
Prime Intellect ran 153 autonomous runs across 18 models in eight-day sandboxed sessions, and Fable 5 topped the leaderboard at 81.7% closed.
Why now: Prime Intellect published the run data and code together, turning a claim about autonomous AI research into one specific, checkable score.
Prime Intellect ran 153 autonomous research runs across 18 frontier models, closing 82% of the gap to a nanoGPT speed record built by humans over months, per the company's benchmark. That's the clearest number yet on how much ground autonomous agents cover inside a narrow, verifiable target, and it says nothing about whether they can choose the target themselves.
Each run got exclusive use of 8xH200 GPUs for up to eight days with no internet access, tasked with cutting training time for a 124M-parameter GPT recipe by tuning optimizer hyperparameters against a shared baseline. Nothing else about the recipe was open to change.
The top leaderboard entry, Fable 5, closed 81.7% of the gap in 8.7 days on 800M tokens, for a score of 2,726. Prime Intellect published the code behind the runs on GitHub, under PrimeIntellect-ai/experiments-autonomous-speedrunning.
Closing 82% of a known, checkable gap in eight days is real progress on search over a bounded hyperparameter space. It's a different claim than being able to find a problem worth solving without a baseline already sitting there. Worth watching is whether Prime Intellect or anyone else runs the same setup on a target that doesn't come with a built-in scoreboard.
Each link below shares sources, entities, or timing with this story.
Claude Code benchmarked against GPT / Shared entities / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Claude Code benchmarked against GPT); both cover Fable, GPT; overlapping topics (best, code).
Anthropic released Fable / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Anthropic released Fable); both cover Fable, GPT; overlapping topics (best, days).
Anthropic released Fable / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Anthropic released Fable); both cover Fable, GPT; earlier Fable coverage from 2026-07-19.
Simon Willison uses Fable / Shared entities / What happened next
Linked by a graph relationship (Simon Willison uses Fable); both cover Fable, GPT; picks up the Fable thread on 2026-08-24.
Anthropic released Fable / Shared entity: Fable / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Anthropic released Fable); both cover Fable; overlapping topics (clean, each).
Linked by a graph relationship (Anthropic released Fable); both cover Fable; overlapping topics (actually, code).
Anthropic released Fable / Shared entities / Earlier coverage
Linked by a graph relationship (Anthropic released Fable); both cover Fable, GPT; earlier Fable coverage from 2026-08-15.
Kimi K3 benchmarked against Fable / Shared entities / Earlier coverage
Linked by a graph relationship (Kimi K3 benchmarked against Fable); both cover Fable, GPT; earlier Fable coverage from 2026-07-21.