Grok 4.20 0309 (Reasoning) is a flagship large language model developed by xAI, released on March 9, 2026. As a specialized reasoning model, it is designed to prioritize deep logical processing and high factual accuracy. Unlike earlier iterations that functioned as general-purpose assistants with optional reasoning toggles, this model is a native reasoning engine, meaning it does not support a standard non-reasoning mode.
Technically, the model features a 4-agent multi-agent architecture where sub-agents—reportedly named Grok, Harper, Benjamin, and Lucas—work in parallel to handle complex problem decomposition. It boasts a significant context window of 2,000,000 tokens, which allows for the ingestion of long-form documents and high-resolution images. The model is multimodal, supporting both text and image inputs while generating text-based responses.
xAI documentation emphasizes the model's low hallucination rates and strict adherence to prompts. It is optimized for agentic workflows, providing advanced tool-calling and structured output capabilities. Performance metrics suggest it maintains high inference speeds despite its intensive reasoning process, with some benchmarks recording speeds up to 247 tokens per second.
Developers should note that the Grok 4.20 0309 (Reasoning) model has specific API constraints. It does not support traditional parameters like presence_penalty, frequency_penalty, or stop sequences in its reasoning state. Furthermore, it does not utilize a reasoning_effort parameter and currently omits support for the logprobs field in its output.