Grok 4.20 Multi-Agent Beta: 4-Specialist Architecture with 65% Hallucination Reduction, 2M Context Window

xAI·high signal

xAI released Grok 4.20 Multi-Agent Beta in Enterprise API with four specialized sub-agents — Grok (coordinator), Harper (real-time X research), Benjamin (logic/coding), and Lucas (creative/contrarian) — running in parallel on every query; a 'Heavy' mode scales to 16 agents. Cross-agent verification drops hallucination rate from ~12% to ~4.2%, a 65% improvement over single-model baselines. The 2M token context window remains the largest among Western frontier models, exceeding GPT-5.4 (128K) and Claude Opus 4.6 (200K).