Grok capabilities

Grok is xAI's AI assistant available on grok.com, iOS, Android, and embedded in X (Twitter). It runs the Grok 4.x model family — Grok 4.3 is the cost-efficient flagship (1M-token context, always-on reasoning) and Grok 4.20 is the multi-agent variant with dedicated reasoning, non-reasoning, and multi-agent-specific model slugs — offering multimodal text, image, and video inputs.

Grok supports global custom instructions (up to 12,000 characters) that apply to all new conversations, letting users set persistent tone, format, or domain preferences. Skills (launched May 18, 2026) extend this to reusable, named capabilities — beyond plain workflow definitions, Skills can generate Office-format documents (Word, PowerPoint, Excel) and PDFs and run automation tasks; teach one once by describing it or uploading reference files, and Grok activates it automatically whenever relevant.

DeepSearch enables Grok to browse the web and X in real time, verify sources, synthesize conflicting information, and produce a cited research report in response to a query. DeeperSearch is a heavier variant that runs multiple verification agents for peer-review-style synthesis.

xAI provides a commercial API exposing Grok models for text/reasoning, image generation, video generation, and real-time voice — all under a single platform with OpenAI-SDK compatibility, native xAI SDK (Python), and Vercel AI SDK support.

Users can attach documents (PDF, TXT, CSV, JSON, HTML, Excel, code files) and images (JPEG, PNG) directly in chat. Grok uses a server-side attachment_search tool to semantically retrieve relevant passages across multiple uploaded files, and applies vision understanding to images.

Grok Build is xAI's terminal-based agentic coding agent and CLI. It plans tasks, writes and edits files, runs shell commands, and builds whole applications from natural-language prompts. Its standout design is parallel sub-agents: it can spawn multiple sub-agents that each work in an isolated git worktree (separate branch and working directory) so they can refactor, test, and edit in parallel, with an orchestrator reconciling the results. It supports a plan-first mode (review/approve/edit the plan before any edit lands), a headless mode for scripts and CI, AGENTS.md conventions, hooks, skills, and native MCP servers.

Grok generates images from text prompts using the Aurora model — an autoregressive mixture-of-experts transformer that builds images patch by patch, excelling at text rendering, lifelike portraits, and consistent stylistic output up to 1024x1024 pixels. The Grok Imagine interface also supports image editing and generating up to 10 variations per prompt.

Grok remembers facts, preferences, and context from previous conversations and applies them automatically in future sessions. Users can view every stored memory item individually, delete specific items, clear all memory, or disable memory entirely from Data Controls settings.

Grok is available in a free tier with strict rate limits and across five paid subscription tiers — X Premium, SuperGrok Lite, SuperGrok, X Premium+, and SuperGrok Heavy — each unlocking progressively higher usage limits, model access, and advanced features; since June 2026 usage is metered as one flexible weekly pool shared across Grok products rather than fixed per-feature caps.

Think mode activates Grok's extended chain-of-thought reasoning, letting it break down problems step by step and show its reasoning process before producing a final answer. Higher-compute variants (Big Brain / Big Reason mode, and Grok 4.20's multi-agent debate) apply additional orchestration for the most demanding queries.

Grok Tasks lets users schedule automated prompts to run at a future time, with results delivered via push notification or email. Supported frequencies include one-time, daily, specific days of the week, monthly, and annual schedules.

Grok Imagine generates short videos from text prompts or by animating still images, producing clips with natively synchronized audio (dialogue, sound effects, ambience) in a single pass. With Video 1.5 (GA 2026-06-16), single-pass clips run up to 15 seconds at 480p or 720p; an Extend from Frame feature chains clips together for longer sequences.

Grok supports real-time spoken conversation through voice mode in the Grok app, enabling bidirectional voice interaction where users can speak and receive spoken responses. A camera-pointing feature lets users direct Grok to analyze what they see during a voice session.

Workspaces are isolated environments within a Grok account, each carrying its own custom instructions, uploaded files, and separate conversation history. Projects (accessible at grok.com/project) extend this with task management, cloud file integration (e.g., Google Drive), and X-platform collaboration features.

Grok has exclusive native access to the X post firehose, enabling real-time analysis of tweets, trending topics, influencer conversations, and breaking news as they happen — without the crawl-delay gap that affects other AI assistants using web search.