xai / grok-4-20

## Overview Grok 4.20 is xAI’s flagship text model built for fast, accurate reasoning. It takes long text prompts and can also analyze attached images. It returns text answers, and it can produce schema-locked JSON when you need consistent fields. It also supports agent-style tool calling, so it can search for fresh facts when you enable Live Search. You can run it in a reasoning or non-reasoning variant. Use the reasoning variant for multi-step work. Use the non-reasoning variant when you want short, direct answers at high throughput. ## What you can build - A research assistant that uses Live Search to answer time-sensitive questions - A vision QA tool that explains charts, screenshots, and UI bugs from images - OCR-style extraction from images, followed by a clean summary - Structured extraction pipelines that return JSON for invoices, tickets, or entities - A coding helper that can plan, debug, and explain complex logic - A long-context summarizer for large pasted transcripts or reports ## Inputs - A required instruction or question you write in plain text. - Optional image attachments in JPG/JPEG or PNG format. Each image can be up to 20 MiB. You can attach more than one image. - Optional system-level guidance that sets tone, rules, or role for the assistant. - Optional identifiers for user and session, provided as a string or a number. Keep them stable to maintain conversation history. - Optional choice between a reasoning and non-reasoning variant. Pick reasoning for multi-step tasks. Pick non-reasoning for straightforward Q&A. - Optional Live Search switch. When enabled, the model may fetch recent information during its answer. - Optional output length limit, measured in tokens. Increase it for long reports or large JSON outputs. ## Outputs - A text response that answers your prompt. - When you request a strict structure, the response can be formatted as valid JSON that matches your schema. This helps when you need reliable fields and types. - When Live Search is enabled, the answer can include synthesized findings based on retrieved pages, instead of only training data. ## Recommended settings - Use the reasoning variant for math, multi-step analysis, and high-stakes decisions. - Use the non-reasoning variant for short answers, routing, labeling, and high-volume pipelines. - Turn on Live Search when the question depends on recent events, changing numbers, or new releases. - Raise the output token limit for deep explanations, long summaries, or multi-record JSON. ## Limitations - The total budget for your prompt plus the model’s response must fit within the 2,000,000-token context window. - Image understanding depends on input quality. Blurry photos, tiny text, heavy compression, and skewed scans can reduce accuracy. - Only JPG/JPEG and PNG image types are supported for image attachments. - This model family does not provide token log probabilities. If you request logprobs, they will be ignored. - Without Live Search or other tools enabled, the model can’t reliably answer questions about very recent events. ## Safety & compliance - Follow xAI’s Acceptable Use Policy and all applicable laws. - Don’t use outputs for illegal activity, fraud, spam, doxing, phishing, or hacking. - Don’t request content that sexualizes or exploits children. - Don’t create pornographic depictions of real people or violate privacy or publicity rights. - Don’t seek instructions for weapons of mass destruction, bioweapons, or severe harm. - Don’t try to bypass safety guardrails. Disclose AI use when it matters and don’t misrepresent outputs as human-made.