Open-source AI model Hermes Agent, launched by Nous Research in February 2026, claims benchmark wins over Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 using Mixture of Agents (MoA) presets, a framework that strings multiple language models together with an aggregator layer to synthesize the best answer. On SWE-bench Pro, a coding-focused evaluation, Opus 4.8 scored 69.2% against GPT-5.5’s 58.6%, and Hermes Agent’s MoA presets claim to clear both bars. The full leaderboard has not been published yet. Hermes Agent also features persistent memory that maintains context across sessions, a built-in learning loop to create new skills from experience, and tool integration for external systems and APIs. Community benchmarking has been active, with meaningful GitHub activity and integration with platforms like OpenRouter. MoA setups run multiple models simultaneously, resulting in higher compute cost per query, but Nous Research frames this as cost-efficient for specialized, task-specific deployments.