Toby Ord argues in “Swarm Scaling” that AI swarms trade efficiency for speed

Toby Ord argues in “Swarm Scaling” that AI swarms trade efficiency for speed

In an essay published September 21, 2026, Toby Ord argues that adding AI agents to a swarm can increase capability, but is less compute-efficient than giving one agent more time to reason. Ord estimates that, across three benchmarks, scaling the swarm by a factor of ten yields the performance gain of scaling one agent’s tokens by three to five. He describes this as diminishing returns: the more agents added, the further swarms fall behind single-agent duration scaling in performance for a given compute increase. Ord’s case for swarms is speed. Agents can run in parallel, so a swarm may finish sooner at greater total cost; he says the measured results make speed the main demonstrated reason to use them. He also raises a possible higher performance plateau for very large swarms, but says this remains unshown. His broader concern is that the measured values of the “stepping on toes” parameter, λ, are close to estimates used in models of recursive self-improvement. In those models, Ord says, higher λ makes an intelligence explosion more likely. He had hoped the values would be lower, and concludes that people should keep tracking them, especially as new orchestration methods emerge. The estimates come from limited benchmark data; whether they hold at much larger swarm sizes remains unknown.

Publication: 803ddfa4-72d4-3939-413f-3e3c316c9869

attribution — pass: The source is a single essay by Toby Ord on his own site. The reading consistently attributes the arguments, estimates and concerns to Ord rather than presenting them as established fact or as Escape Meridian's view.

Swarm Scaling September 21, 2026Toby Ord

summary — uncertain: The reading tracks Ord's argument faithfully, but a non-specialist may not follow it. λ is named only as the 'stepping on toes' parameter; the reading never says it measures how parallelisable a task is. 'Tokens', 'duration scaling', 'intelligence explosion' and 'orchestration methods' are also undefined and have no definitions entry. The reading omits that the estimates are inferred from OpenAI's GPT 5.6 Sol launch charts. Smallest repair: add one plain-language clause explaining λ, add 'tokens' and 'intelligence explosion' to the term library, and name the OpenAI chart data.

Here $\lambda$ is a parameter measuring how parallelisable the task is. They call it the ‘stepping on toes’ parameter.

date — pass: originalPublished is 2026-09-21T18:53:19+01:00, which is September 21, 2026 in UTC. This matches the source dateline and the reading's day-level date. The page was edited on 27/09/2026 to remove a closing Navier-Stokes section. The reading neither uses nor relies on that edit, and its cited passages are plainly part of the original body. There is no hindsight framing, and this is a publisher page, so no archive disclosure is needed.

Edit (27/09/2026): I originally closed this essay with a speculation on swarm scaling

support — pass: Each claim matches its excerpt: the efficiency shortfall, the 10x-to-3x–5x conversion, speed as the main demonstrated reason, the possible higher plateau, the link to recursive self-improvement, Ord's hope for lower values and his call for tracking. The λ values come from regressions Ord asked Claude Opus 5 to run, but Ord adopts them, so 'Ord estimates' is acceptable. Despite the later edit, the cited text is plainly original body text.

I’d hoped that the value of $\lambda$ for AI agents would be lower, making an intelligence explosion less likely, but that appears to not be the case.

media — pass: No media is selected, and no verified candidates were supplied, so a null media choice is appropriate.