MANY MINDEDAI Team Coordination Costs More Than It Gains in Knowledge Production
Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench at medium and maximum reasoning effort levels. Team configurations consumed 1.8 to 5.1 times more tokens than single-agent setups for both models. Only one comparison—GPT-6 Sol at medium reasoning—showed statistically significant quality gains (a 7.3-point score increase). At maximum reasoning effort, team configurations provided no quality advantage for either model. The study observed diminishing returns when scaling teams: Anthropic noted Opus 5.5 teams lost quality gains as they grew from 10 to 100 agents, while Fable 5.1 scores declined slightly when scaling from 30 to 100 agents on knowledge tasks. OpenAI researcher Noam Brown and Eric Provencher identified the core issue as coordination breakdown, calling it the 'coordination tax' that increases computational cost without proportional value.
This friction directly impacts knowledge affordability. When AI teams waste tokens without meaningful quality gains, the computational cost of producing knowledge services rises. For users, this means higher prices for AI tools that could otherwise be cheaper and more accessible. The coordination tax creates a barrier where knowledge production becomes less affordable as teams scale—especially for tasks where quality gains are minimal.
What to watch: How real-world constraints affect scaling patterns. Vals AI’s benchmark doesn’t account for all deployment scenarios, and quality gains vary significantly by task type. The study was published October 11, 2026, but results are specific to the Vibe Code Bench and two reasoning levels. This means the friction may not apply equally to all AI services or all users. The next step is whether teams can reduce coordination costs without sacrificing quality—something that would lower barriers to affordable knowledge services.
Source: The Decoder