Multi-Agent AI Teams Burn Up to 5x Tokens for Little Gain

What changed

Research from Vals AI and Anthropic has quantified the inefficiency of AI agent teams, revealing a stark mismatch between token consumption and quality output. The findings challenge the popular assumption that deploying multiple agents improves results.

Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the “Vibe Code Bench,” comparing solo agents to teams at two reasoning levels: medium and maximum reasoning effort. The teams consumed between 1.8x and 5.1x more tokens than single agents across the tests.

The quality gains did not match the cost. Out of four comparisons, only one showed a statistically significant improvement: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher than the solo agent. At maximum reasoning, neither GPT-6 Sol nor Claude Opus 5.5 benefited from the team structure.

Anthropic conducted its own tests with Opus 5.5 to measure how performance scales with agent count. Teams reached a given performance level faster, but the gains diminished sharply. On a knowledge base task, scores climbed from 0.53 with one agent to 0.70 with ten agents. Adding 90 more agents only pushed the score to 0.74. On Lean theorem proving, the pattern repeated: performance jumped from 0.39 to 0.66 between one and ten agents, then barely moved to 0.68 with 100 agents.

OpenAI researcher Noam Brown confirmed the pattern on the Dwarkesh Podcast. He noted that four agents solved tasks twice as fast but also cost twice as much. The efficiency declined further at 16 agents. Brown emphasized that the gains depend heavily on task type. Web research and math parallelize well. Writing a novel does not, and neither does most creative or sequential work.

OpenAI developer Eric Provencher recently warned against agent swarms, calling them mostly wasted money. He attributed this to a “coordination tax,” the overhead cost of managing communication between agents.

Why it matters

The research directly addresses a critical gap between vendor marketing and real-world economics for professionals deploying AI agents. Most teams pursuing AI agent efficiency assume more agents mean better outcomes. This research shows that assumption is false for most cases, especially when models already run at full compute capacity.

The practical implication is stark. Teams spending on agent orchestration platforms, managing multiple agent deployments, and scaling infrastructure may be burning resources for negligible quality improvements. A single well-configured agent often delivers comparable results at a fraction of the token cost.

This matters most for organizations operating at scale. Token costs accumulate quickly. A workflow that consumes 5x more tokens per task will see that multiplier compound across thousands of requests. The financial exposure grows even larger when using high-end models like Opus 5.5 or Sol at maximum reasoning settings.

The findings also reshape how teams should think about speed versus quality. Agent teams do buy speed. But speed alone does not justify the cost unless latency directly blocks business objectives. In most knowledge work, a slower single agent that delivers the same result costs significantly less.

Tasks unsuitable for parallelization suffer worst from the agent team penalty. Writing, complex reasoning, and sequential workflows see coordination overhead without speed or quality gains.

What to test

Before committing to AI agent teams, validate these claims against your specific use case.

• Run head-to-head benchmarks. Test your exact task with one agent versus multiple agents on the models you plan to deploy. Measure token consumption and output quality side by side. The Vals AI data is directional, not universal.

• Isolate task parallelizability. Be honest about whether your workflow can split into independent parallel streams. If each step depends on the previous one, agent teams will not deliver speed gains. If coordination between agents is necessary, that coordination tax will compound.

• Track actual token consumption. Measure real token usage on your specific tasks before scaling. The difference between theoretical and observed token usage often surprises teams.

• Set a cost threshold. Define the maximum acceptable cost increase for any quality gain you observe. If your test shows a 5x token increase for a 2 percent quality bump, decide whether that is worth it for your business.

• Measure coordination overhead. Time how long agents spend communicating with each other versus executing work. High coordination overhead signals the task is not a good fit for multi-agent approaches.

• Test at your target reasoning level. Anthropic and Vals AI tested different reasoning settings. Maximum reasoning effort showed different results than medium. Run your tests at the exact configuration you plan to deploy.

The conclusion

The evidence from Vals AI, Anthropic, and OpenAI researchers points in one direction: AI agent teams waste massive tokens for barely measurable quality gains in most cases. The coordination tax is real, and the speed gains do not offset the cost for sequential or complex reasoning tasks.

This does not mean agent teams never make sense. They fit specific use cases well: embarrassingly parallel tasks like processing thousands of support tickets, independent web searches, or batch math problems where coordination overhead is minimal. For those cases, speed gains can justify the cost if latency matters.

For general knowledge work, writing, reasoning, and creative tasks, a single well-configured agent remains the more economical choice. Teams pursuing efficiency should default to solo agents and only deploy teams when task structure and business requirements explicitly justify the token and coordination costs.

The research shows that scaling from ten to 100 agents produces marginal returns at best. Organizations should test their specific use cases before assuming that more agents deliver better results or justify higher spending.

Related reading

• ChatGPT Data Agent: What Business Teams Should Test First
• Sakana AI MLR Catches 73% of Core-Claim Errors in Papers
• Claude Dynamic Workflows: 1,000 Parallel Agents Now Available