Grok 4.20 went live in March 2026 as the first model in xAI's 4.20 branch, and it is the version that set the shape of the line: a one-million-token context window, image and text input, and two-tier billing in which a prompt that reaches 200,000 tokens reprices the entire request at double rates. xAI describes it as a high-speed model built for agentic tool calling, and claims for it the lowest hallucination rate on the market together with strict prompt adherence - a positioning the company has not repeated for later versions. It is also the cut-off point for a quiet API change: logprobs and top_logprobs are silently ignored by Grok 4.20 and everything released after it. Six months on, the model is still sold, still cheaper than the 4.5 and 4.6 flagships that followed, and still takes twice their context.