Anthropic Claude Opus 5.5 swarm scales better than GPT-6.2 Astra
Researchers tested how scaling inference-time compute and agent swarms affect performance engineering tasks using a new benchmark suite called Optimization-Bench. They found that Anthropic's Claude models achieved significant speedups in swarms by coordinating through shared boards rather than direct messaging, whereas OpenAI's models often degraded in performance as their swarms grew larger. The study demonstrates that swarm organization and communication topology are critical factors in determining whether multi-agent systems successfully scale. Open article →