Grok 4.20 slashes hallucination rate from 12% to 4.2% using multi-agent debate
xAI's Grok 4.20 runs four specialized agents on a single 3-trillion-parameter MoE backbone, cutting hallucinations by 65% while achieving 220.5 tokens/sec and a 2-million-token context window.