Brian Wang Elon Musk confirmed to me, Brian Wang, that the current beta model is the ~500B parameter base model. Overall early consensus from testers, it beats or matches frontier models (GPT-5, Claude 4/Opus 4.5, Gemini 3) in practical coding, simulations, iterative work, and real-world agentic tasks. XAI Grok 4.20 will scale to 16 agents in Heavy mode. Provisional LMSYS/Arena ELO ~1505–1535. XAI Grok 4.20 is projected to take #1 once fully ranked. Grok 4.1 Thinking was 1483). Heavy mode is expected to be +30 to +80 Elo on hard tasks. Realistic range for Heavy is ~1540–1610+ Screenshot xai 4.20 lets you track about 200 active queries. I created my own dashboard and you can easily set it to run the updates for the dashboard at whatever schedule you want. XAI Grok 4.20 made a good flight simulator and passed most of the tests far better than XAI Grok 4.1. Rapid weekly learning — improves every week during beta with public release notes (first model to do this at scale). Dramatically lower hallucinations via internal cross-validation. Much faster inference + better multimodal (especially medical image/file analysis for second opinions). Stronger open-ended engineering reasoning, iterative coding, simulations, and agentic tasks. Unique edges are real-time X data, lower censorship, built-in team intelligence, weekly rapid improvements. Still early beta — no full public benchmark suite yet, but hands-on and trading results are extremely strong. This is the first model that genuinely feels like working with a small expert team instead of one smart assistant. Wes Roth likes the 4 agent system. Completely different paradigm. Multi-agent collab beats single-model reasoning on hard tasks. 00:00 – Intro