xAI Ships Grok 4.3 Beta With 2M-Token Context and Native Video, Launches Grok Computer Autonomous PC Agent

Elon Musk’s xAI has made two coordinated moves in the past 48 hours that together represent the company’s most ambitious product push since the original Grok launched. On April 17, xAI rolled out Grok 4.3 as an early access beta, exclusively available to SuperGrok Heavy subscribers. Simultaneously, Grok Computer — the company’s autonomous PC agent — entered a wider beta, with Musk confirming large-scale public testing would begin within days.

Taken together, the releases extend xAI’s bet on a particular theory of what AI assistants should become: not tools that respond to individual queries, but persistent agents that understand long-running context, operate across modalities, and ultimately take actions on behalf of users — including controlling their computers.

What Grok 4.3 Delivers

The headline technical upgrade in Grok 4.3 is a 2 million token context window, the largest of any publicly available model at this price tier. At roughly 1.5 million words — or approximately the combined length of the entire Lord of the Rings trilogy repeated seven times — this context window enables tasks that are simply impractical with shorter contexts: analyzing entire software codebases, synthesizing multi-year document archives, or maintaining conversation continuity across months of interaction history.

Native video understanding is the second major addition. Unlike image-capable models that process video as sampled frames, Grok 4.3 ingests continuous video streams and reasons about temporal relationships between events — understanding not just what appears in a frame but how actions and states evolve over time. The practical applications span surveillance footage analysis, instructional video parsing, and meeting transcription enriched with visual context.

The model also introduces native document creation: users can ask Grok 4.3 to generate formatted PDFs, PowerPoint presentations, and spreadsheets directly, without exporting through third-party tools. While this capability exists in other products, xAI’s implementation integrates directly with the 2 million token context — meaning documents can be generated from or about extremely large source corpora.

Elon Musk confirmed on X that Grok 4.3 is a live development build rather than a polished versioned release, with “near-daily improvements” expected. A 1 trillion parameter version remains in training, with Musk indicating it will substantially upgrade coding performance, long-context handling, and reasoning depth.

The $300 Threshold

The SuperGrok Heavy subscription required to access Grok 4.3 costs $300 per month — a price point that is deliberately positioned above OpenAI’s ChatGPT Pro ($200/month) and Anthropic’s Claude Max ($100/month for standard, $200/month for extended). This is not a mistake or an artifact of pricing ambition; it reflects a calculated positioning decision.

xAI appears to be targeting the upper tier of professional users for whom AI is a production tool rather than a productivity supplement. At $300 per month, Grok 4.3 must deliver demonstrable return on investment for engineers, analysts, and operators who process large volumes of complex information daily. The 2 million token context window and native video capabilities are specifically designed to create value that is not achievable at lower price points.

The model is available on iOS, Android, and web, with Musk noting that the iOS version will receive Skills — a configurable workflow automation system — in a near-term rollout. Early reviewers who have published assessments under NDA describe the document creation and video understanding as genuinely differentiated, while noting that core reasoning performance is competitive with but not clearly superior to the current best models from Anthropic and Google.

Grok Computer: The Autonomous Agent Layer

Running in parallel to the Grok 4.3 launch, Grok Computer is a more structurally significant product. Where Grok 4.3 is an upgrade to a conversational model, Grok Computer is an attempt to extend AI agency into the desktop environment — and beyond the boundaries of what any chatbot interface can accomplish.

The core capability is straightforward to describe but technically complex to execute: Grok Computer can see your screen, understand what’s on it, and take actions — clicking, typing, navigating between applications, filling forms, executing terminal commands, and chaining multi-step workflows — without requiring any API integration or application-specific configuration.

It achieves this through pixel-level screen reading combined with a continuous 5-second video buffer that gives the agent an understanding of UI state and application context. Because it operates at the pixel level rather than the API level, it works with any software — including legacy applications from the early 2000s that predate modern API conventions.

The practical implications are significant. Grok Computer can, according to xAI’s documentation: research a topic in a browser, synthesize findings into a structured spreadsheet, format that spreadsheet into a presentation, and email it to a specified recipient — all in a single automated workflow, without human intervention at any step. It can test software by interacting with UI elements and generating test reports. It can operate within enterprise applications where no AI integration exists.

The Macrohard Connection

Grok Computer is not a standalone product. It is the consumer preview of a much larger initiative that Musk unveiled on March 11, 2026 under the name Macrohard — a deliberate provocation in the direction of Microsoft. The project is a joint Tesla-xAI venture, backed by a $2 billion Tesla investment into xAI, aimed at building a full software productivity suite powered by Grok’s underlying models.

The name and framing are characteristically Musk: the provocative naming pattern that reframes a competitor as the legacy, the billionaire’s investment making both companies beneficiaries of the same outcome. But the strategic logic is coherent. Tesla’s Optimus humanoid robot program requires exactly the kind of desktop-to-physical-world agency that Grok Computer is developing. A model that can autonomously operate a computer is one step removed from a model that can autonomously operate a robot.

Musk’s confirmation that Grok Computer will receive a “large-scale public testing” expansion within days suggests the private beta has been used to validate core stability and safety. The current beta runs on Grok 4.20 Beta 2, which is described as deliberately limited in capability — the full capability release, expected to be based on Grok 5, is being positioned as a major product milestone later in 2026.

Competitive Positioning

The releases arrive as the autonomous agent space becomes increasingly competitive. OpenAI’s Agents SDK, updated this week, provides a framework for developers to build agentic workflows; Anthropic’s Claude has computer use capabilities available in API. Google is reportedly working on its own PC agent product. Microsoft, the incumbent the Macrohard name is designed to mock, has Copilot integrations in Windows 11 that provide a different but overlapping set of capabilities.

xAI’s competitive advantage, if it materializes, will likely come from vertical integration: a proprietary model (Grok), a proprietary distribution channel (X/Twitter with 600 million monthly users), and a proprietary physical world application (Optimus robots at Tesla). No other company has that stack. Whether the stack produces a product that is significantly better than its competitors’ outputs, or simply different, remains to be seen. But the architecture of the bet is visible — and the pace at which xAI is shipping in April 2026 suggests it is in an execution phase, not a planning one.