Adding agents multiplies the bill, not the result
Edition of 2026-10-12
# AI For All ### Monday, October 12, 2026
As of today, anyone assembling an AI agent workflow has to weigh independent evidence that adding teammates multiplies the cost, not the result.
The most expensive way to build is the popular one
Vals AI, an independent evaluator, took GPT 6 Sol from OpenAI and Claude Opus 5.5 from Anthropic and ran each of them on its Vibe Code Bench. The test builds a complete web application from a written specification, then grades it. Each model ran once alone, and once as a team that could spread the build across helper agents, at two levels of effort.
The result is blunt. Teams cost *1.8 to 5.1 times as much as a single agent, and only one of the comparisons produced a score gain that was statistically real: Sol at medium effort, up 7.3 points*. The top-scoring run, Opus at maximum effort with a team, was also the slowest and by far the costliest in the study.
The reason is repeated work rather than skill. Every helper agent keeps its own copy of the specification and its working state, and re-sends both on every model call. For Sol, simply raising the effort of the single agent beat adding teammates, at a fraction of the price. The authors are clear that this is one benchmark of web apps. The caution is theirs, and it is fair.
Weak signal
Britain's AI minister, Kanishka Narayan, told a London audience today that Britain cannot rely on technology companies to police their own models, and that the country needs more of its own AI infrastructure. There is no new rule and no date, so it does not rank. But it lands the same week that Britain's former AI safety chief said the labs could not stop a hostile state from stealing an unreleased model, and that UK intelligence chiefs briefed the cabinet on AI risk. The direction of the regulator is becoming legible.
The next model release will arrive with a score attached. The question worth asking never changes: what did the team cost, and what did it change.