Agent teams vs single agent: what Vals AI's test found
2026-10-12 · 4 min read
In a head-to-head test from Vals AI, agent teams cost 1.8 to 5.1 times more than a single agent, and in three of four setups they did not score meaningfully better. The one clear win was GPT-6 Sol at medium reasoning effort, where a team added 7.3 points.
"Use a team of agents" has become standard advice for anyone building with AI, so it's worth seeing a test that prints the bill next to the score.
What Vals AI posted
On October 9, @ValsAI asked whether agent teams pay off. Vals ran GPT-6 Sol and Claude Opus 5.5 on its Vibe Code Bench, where an agent has to build a full web app that is then graded with UI tests. Each model ran alone and as a team, at medium and at max reasoning effort. The headline from the post: teams cost "1.8× to 5.1× more."
The full write-up, by Katrina Drozdov, has the details. A "team" here was a lead agent with up to five subagents. Sol ran through OpenAI's Agents API and Opus through the Claude Code command line tool, so some of the gap between the two models comes from the harness, not the model.
The numbers
Here is cost per app and score for each setup:
| Model | Effort | Single agent | Team |
|---|---|---|---|
| GPT-6 Sol | Medium | 77.6% for $1.22 | 84.9% for $3.07 |
| GPT-6 Sol | Max | 89.0% for $4.82 | 90.4% for $8.54 |
| Claude Opus 5.5 | Medium | 91.5% for $4.08 | 91.2% for $9.10 |
| Claude Opus 5.5 | Max | 89.8% for $23.77 | 93.2% for $122 |
Only Sol's medium-effort jump passed Vals' significance test. The other three differences fall within the noise of a 50-app sample.
The Opus max-effort team posted the best score of the study and also the worst economics: $122 and nearly three hours per app, against under $24 for the single agent. Opus at medium effort, working alone, scored 91.5% for about $4. On this benchmark, the cheapest Opus setup matched the most expensive one.
Why teams cost so much
Vals says most of the extra spend was cached input. Each subagent re-sends its own context, so a team pays to read the same material several times over.
@pvncher, who works on Codex developer experience at OpenAI, made a related point in September. He wrote that beyond two subagents at a time "you're almost certainly burning tokens for 0 quality gain," because agents don't trust each other and keep re-checking each other's work. That was one developer's observation. The Vals numbers mostly back him up.
The two models also behaved differently. Sol's lead agent split the work across parallel subagents early, so its teams ran as fast as or faster than the single agent. Opus teams worked in sequential waves, starting from a shared plan document, and took up to 2.3 times longer.
Our read
You can spend more on a task in two ways: add agents, or let one agent think longer. In this test the second was usually the better buy. For Sol, raising reasoning effort on a single agent added 11.4 points, a bigger gain than any team produced. Vals' own conclusion is careful: a team "can boost performance, but higher reasoning effort is an alternative worth weighing."
The study has limits. Each setup ran once, on one benchmark, and building a web app is a fairly linear job. Vals expects teams to earn their cost on work that splits cleanly into parallel pieces, such as research across many sources or processing a big batch of separate files. We've written before about a 1,393-agent cleanup job where parallel work made sense.
What this means for a business using AI
If a vendor pitches you a "swarm" or a "team of AI agents," ask what a single agent with more reasoning time scores on the same task, and what each option costs per run. A demo that looks impressive can hide a bill that runs five times higher than it needs to.
For most small-business automations, like answering new leads or chasing unpaid invoices, one well-scoped agent with a person reviewing its output is the setup to beat. If you want help deciding where an agent fits in your week, and where it doesn't, New Face Design offers a free process audit.