TLDR Tech / Practical
Agent team efficiency tested on Vibe Code Bench
Researchers evaluated native multi-agent functionalities from OpenAI and Anthropic using Vibe Code Bench, assessing GPT-6 Sol and Claude Opus 5.5 in developing full web applications. The study found that using agent teams was 1.8 to 5.1 times more expensive than single agents, with minimal statistically significant score improvements. Sol saw a notable 7.3-point gain at moderate effort, while Opus reached its peak score only at maximum effort, incurring substantially higher costs for runtime and tokens.
Read the full story