Vals AI: agent teams cost 1.8x to 5.1x more than one agent, one clear win in four tests

Vals AI built 50 web apps on its Vibe Code Bench under eight setups: GPT 6 Sol and Claude Opus 5.5, medium and max reasoning, single agent or a lead plus up to five subagents. Teams cost about 1.8x to 5.1x as much. Only Sol at medium effort gained significantly (+7.3 points, p=0.005); the other three differences were not significant. Each setup ran once.

HEAT

OUTLETS

1

FIRST SEEN

Oct 11, 2026, 11:39 ET

LAST SIGNAL

3h ago