AI agent teams cost up to 5.1x. One setup got a real gain

AI agent teams cost 1.8x to 5.1x more than a single agent, and in three of four tests they did not score measurably better. That is the result of a Vals AI case study published Oct. 9. It ran GPT-6 Sol and Claude Opus 5.5 as solo agents and as lead-plus-subagent teams on 50 full web-app builds. If a vendor is selling you an "agent swarm," this is the number to ask them about.

What actually happened

Vals used Vibe Code Bench: build a complete web app from a product spec, deploy it with Docker Compose, then let browser agents run the UI test suite. In team mode a lead agent could run up to five subagents at once and was told to split, delegate, integrate and verify.

The four comparisons (mean score, cost per app):

• GPT-6 Sol, medium effort: 77.6% solo vs. 84.9% team, +7.3 points. Statistically significant. Cost went from $1.22 to $3.07.
• GPT-6 Sol, max effort: 89.0% vs. 90.4%. Not significant. $4.82 to $8.54.
• Opus 5.5, medium effort: 91.5% vs. 91.2%. Not significant. $4.08 to $9.10.
• Opus 5.5, max effort: 89.8% vs. 93.2%. Not significant. $23.77 to $122 per app, and median runtime rose from about 75 minutes to 174.

One more detail: for Sol, turning effort to max on a single agent added 11.4 points. That is more than the team added at medium effort. Vals says each setup ran once and the two models used different harnesses, so treat it as a case study, not a law.

The Decoder also pointed to Anthropic's Opus 5.5 system card, where gains got smaller as agent counts grew. In one task, going from 10 to 100 agents moved the score only from 0.70 to 0.74.

Why it matters for your business

Multi-agent is easy to sell because it looks like a team of employees. But each extra agent re-reads context, coordinates and re-checks other agents' work, and you pay for every token. For most small-business automation, like processing invoices, answering tickets or drafting product copy, one well-scoped agent with good tools and a test step beats an orchestra.

How we decide:

• Start with one agent. Change the effort level and the tools before you add agents.
• Add agents for speed, not quality. Parallel research or independent tasks can finish faster. Do not expect better answers.
• Measure cost per finished task. Score the output with your own tests, then divide by the bill.
• Cap the fan-out. A hard subagent limit and a per-run budget stop a $122 run from becoming the default.

Sources: Vals AI, The Decoder.