Grok 4.20 Multi-Agent launched alongside the standard Grok 4.20 in March 2026 and is xAI's answer to deep research: instead of one model working through a question, several agents work on it in parallel and their findings are combined. It shares everything measurable with its sibling - the same one-million-token context window, the same text and image input, the same per-token price down to the cached-input rate - and differs in what the operator is allowed to draw from it. The multi-agent endpoint is capped at 9 requests per second and 2.5 million tokens per minute, roughly a quarter of what the single-agent model may consume, which is the practical cost of running several agents behind one call. xAI still labels the model beta.