Meta Muse Spark ran 60+ agent loops in 2 hours for under $1, wowing Scale CEO

alexandr_wang · x · 2026-09-03

Scale AI CEO Alexandr Wang amplified a user test of Meta's Muse Spark 1.3 Ultra Contributor vs Fable 5.1 xHigh: with a self-improvement rule (redo if judges score below 9.5/10), Muse Spark ran 20 improvement loops with 3 agents each—60+ agent runs over 2 hours—for under $1 total, far exceeding expectations.

Related event: Meta's Muse Spark Runs 2-Hour Self-Improvement Loops for Under $1, Stunning Scale CEO(3 posts)→

Original post →

More from coding & agent

coding & agent channel →