Single-agent runs cost 25% of a 16-agent swarm's tokens, per GPT-5.6 release data

tobyordoxford · x · 2026-09-22

Toby Ord shares data behind his Swarm Scaling analysis: per graphs in OpenAI's GPT-5.6 release, reaching any given capability level takes the fewest total tokens (and thus compute) with the smallest 1-agent swarm — roughly 50% of what a 4-agent swarm needs and 25% of a 16-agent swarm. A quantitative look at the cost structure of scaling inference via agent swarm size.

Related event: Oxford's Toby Ord quantifies swarm scaling: far less compute-efficient than longer thinking(17 posts)→

Original post →

More from Infra

Infra channel →