Agent teams cost up to 5.1 times more and rarely help

What this is about

Vals AI tested whether multiple AI agents produce better small web applications than a single agent. The evaluation company ran GPT-6 Sol and Claude Opus 5.5 on Vibe Code Bench, each as a solo agent and as a team, at medium and maximum reasoning effort. The primary analysis was published on October 9, 2026.

The result challenges a common expectation: teams cost between 1.8 and 5.1 times as much as solo agents, according to Vals AI. Of four direct comparisons, only GPT-6 Sol at medium reasoning effort delivered a statistically significant advantage. The gap was 7.3 points. At maximum reasoning effort, neither model showed a reliable team advantage.

What the comparison actually does

Vibe Code Bench tests how well models turn a task description into a working web application. Vals AI compared the same model families under two conditions: one agent completed the task alone, or multiple agents divided the work. The company also varied reasoning effort. This setup was intended to reveal whether coordination creates a quality gain of its own or mainly consumes more compute.

The important distinction is between score and cost. A team can attempt several approaches in parallel, compare solutions, and split work. Every additional agent also makes its own model calls, reads context, and must coordinate results. That coordination overhead counts. Vals AI therefore reports not only achieved quality but also cost per application.

The test did not show that teams are always worse. GPT-6 Sol at medium reasoning effort produced a measurable gain. The more precise conclusion is that adding agents is not an automatic quality control. The model, task, reasoning effort, and budget jointly determine whether the division of work pays off.

Why it matters

Agent teams are increasingly used in development tools and internal automation. The appeal is intuitive: if several people can work faster in parallel, several agents should also accomplish more. Models, however, add overhead because tasks must be separated, intermediate results transferred, and conflicting solutions reconciled.

For companies and individual developers, this is a budget decision. Anyone spending five times as much on model calls should expect more than a small or statistically uncertain quality gain. The comparison suggests testing a strong solo agent with a clear task and sufficient reasoning effort first. A team becomes more attractive when the work can genuinely be parallelized or when speed matters more than the lowest cost.

Independent coverage by The Decoder reaches a similar interpretation: more agents can buy speed without automatically improving final quality. Other published tests also show diminishing gains for very large teams. That makes the Vals findings relevant to real budgets, while still leaving teams to measure their own workflows.

In plain language

An agent team is like a group packing one suitcase. It can be faster if one person handles clothes, another handles electronics, and a third handles documents. If everyone keeps rearranging the same suitcase, questions, duplicated work, and conflicts pile up. More hands help only when the task supports a clean division.

A practical example

A small software team wants to create ten simple internal web applications. Suppose a solo agent costs €20 per application and reaches the required quality on eight of ten tasks. A team at the highest multiplier observed by Vals AI would cost up to €102 per application under the same simplified assumption.

If the team still delivers eight usable applications, the total bill rises from €200 to €1,020 without added value. If it delivers nine instead of eight, that one additional usable application costs €820 more in this example. This is not a study result but a decision aid: before using a team, define how much quality improvement or time saved would justify the extra cost.

Scope and limits

• Vibe Code Bench examines the creation of small web applications. Its results do not automatically transfer to research, mathematics, security analysis, or long software projects.
• The comparison covered GPT-6 Sol and Claude Opus 5.5 under specific settings. Other models, role assignments, or coordination methods may perform differently.
• The study measures a limited task set and prices at one point in time. Models, prices, and tools change.
• A statistically insignificant difference does not prove that teams never help. It means the observed advantage in this test was not reliable enough.

SEO & GEO keywords

AI agents, agent teams, multi agent systems, Vibe Code Bench, Vals AI, GPT-6 Sol, Claude Opus 5.5, coding agents, model costs, software development