AI Agent Teams Burn Up to 5.1x More Tokens for a Barely Noticeable Gain

Teams of AI agents barely outperform single agents, yet cost up to 5.1x more. That is the conclusion Vals AI reached. The researchers ran scenarios on two flagship models, GPT-6 Sol and Claude Opus 5.5, and saw a measurable benefit from teamwork in only one of four tests. The other three produced differences that are hard to distinguish from noise. Anthropic's data paints the same picture: past ten agents, quality plateaus while tokens keep burning.

What the tests showed

The idea behind multi-agent systems sounds logical. A single agent struggles to hold a large task in its head, so the work gets split among specialists: one plans, others execute, another checks. Each works in its own context and, in theory, makes fewer mistakes.

Vals AI compared such teams with single agents on identical tasks. The price of teamwork is steep: cost rose by up to 5.1x. Against that, the quality gain is modest. It showed up in one scenario out of four, and the result held on both models, GPT-6 Sol and Claude Opus 5.5, so the approach's weakness cannot be blamed on flaws in either one.

The summary I worked from does not make clear which tasks were in the test set or how the orchestration was built. That is why the numbers are better read as a signal than as a verdict on all architectures.

A plateau after ten agents

Anthropic offers a separate argument. The company's own data shows that once a team has more than ten agents, quality stops improving while token cost keeps rising. The returns curve flattens; the cost curve does not.

The cause is probably overhead. Agents have to pass context to one another, reconcile results, and re-check each other's work. Each of those steps costs tokens and adds almost no new information to the system. This is my own interpretation, since the source data does not disclose the mechanism, but it explains well why adding more participants stops paying off.

Key figures from the research:

• researcher: Vals AI;
• models in the tests: GPT-6 Sol and Claude Opus 5.5;
• cost increase for the team approach: up to 5.1x relative to a single agent;
• measurable quality gain: 1 test out of 4;
• threshold per Anthropic's data: beyond 10 agents, quality plateaus;
• token consumption beyond the threshold keeps growing.

How to test multi-agent setups in your own product

For teams building agent systems, the takeaway is practical: a swarm of agents should not be the default architecture. The baseline should always be a single agent. If one agent with a good prompt and good tools solves the task, add a second only when the gain is proven.

What to do:

1. Measure the cost per solved task, not average quality. A gain of a couple of percentage points at five times the bill almost always loses.

2. Compare the team and the single agent on your own data. Results from someone else's benchmark transfer poorly, and in the Vals AI study the gain appeared in only one scenario out of four.

3. Cap the team size. As the number of agents grows, watch where the gains end.

According to Anthropic's data, the limit sits at around ten. 4. Set a token budget per run and abort the run when it is exceeded.

Team systems make mistakes more expensive: a swarm stuck in a loop burns money faster than a single agent.

Multi-agent setups remain reasonable for tasks where parallelism saves real time or where isolated contexts are needed. But their value now has to be proven with numbers, not with an elegant orchestration diagram.