Agent Teams Cost Up to 5.1x More and Almost Never Win: Vals AI Puts a Number on the Coordination Tax

The most fashionable idea in agent engineering - that pointing several AI agents at one problem beats using one - just got a price tag. Evals company Vals AI ran GPT-6 Sol and Claude Opus 5.5 both solo and in teams on its Vibe Code Bench, which asks a model to build a complete web application from a product spec and then scores the deployed app with browser agents clicking through it. Teams cost 1.8x to 5.1x more per app, and only one of four comparisons produced a statistically significant quality gain, according to The Decoder.

The numbers are more pointed than the headline. In the single significant result, GPT-6 Sol at medium reasoning effort scored 84.9 percent in team configuration against 77.6 percent solo - a 7.3-point gain that came at 2.5 times the cost. Claude Opus 5.5 at medium effort scored 91.2 percent as a team and 91.5 percent alone, while paying $9.10 per app versus $4.08. At maximum effort, an Opus team spent $122 per app, 5.1 times the solo figure, for a 3.4-point gain that was not statistically significant.

The cheapest lever beat the expensive one. For Sol, raising a single agent from medium to maximum reasoning effort added 11.4 points - more than the 7.3 points the medium-effort team bought. For Opus, a lone agent at medium effort performed about as well as the most expensive team, whose 93.2 percent was only about 1.6 points better and not significant. Where teams did help, they caught a specific failure: on the nine apps where the Sol team gained 20 or more points, the solo agents had claimed passing tests while skipping features, and the lead agent browser-tested the deployed app in 46 of 50 runs.

Most of the extra spend was not intelligence. Vals found the added cost came mainly from cached input tokens, because each subagent re-reads its own context on every call - at maximum effort, Opus subagents read a median of 224 million cached tokens per app. Time did not improve either: Opus delegated in sequential waves of builders, feature agents and testers, so its teams ran up to 2.3 times longer than a single agent while each wave waited for its slowest member.

The finding lands on ground Anthropic has already mapped. The Claude Opus 5.5 system card shows the same plateau: scaling from 10 to 100 agents nudged knowledge-base scores from 0.70 to 0.74 after 24 hours and Lean theorem proving from 0.66 to 0.68, while a single agent started at 0.53 and 0.39 respectively. OpenAI researcher Noam Brown made the same point on the Dwarkesh Podcast: four agents solve tasks about twice as fast at twice the cost, and the pattern holds at 16 with slightly worse efficiency. Tasks that parallelise - web research, math - benefit; writing a novel does not, any more than assigning 10,000 people to it.

The caveats are real: each configuration ran once, the two models ran through different harnesses, and one benchmark with two models is a narrow sample. But the direction is hard to argue with. If most of what a team buys is wall-clock time and a verification pass, builders may capture most of the benefit by raising a single agent's reasoning effort and adding an explicit end-to-end test - a cheaper configuration Vals did not test.