AI Radar: xAI Grok 4.3 — Fast, Closed, Frontier

xAI is moving fast. With Grok 4.3 and the multi-agent offshoot 4.20, the lab shows where proprietary frontier models are heading, which is exactly why it belongs on the radar, even though it does not fit our self-hosting approach.

Grok 4.3 is xAI's entry in the upper model tier — tightly coupled to X, and to be judged accordingly. (Quelle: xAI / Wikimedia, public domain)

Without reasoning Grok 4.3 drops eleven index points — the effort level here is not a detail, it is the product. (Quelle: Artificial Analysis — model leaderboard)

The State of Play

Grok 4.3: available since April, with native video input, document generation (PDF, spreadsheets, slides), and noticeably better tool-calling.
Grok 4.20: a multi-agent architecture with four cooperating agents and a 2M-token context, plus real-time access to X data.
Grok 5: training on the Colossus 2 cluster, in the multi-trillion-parameter range, expected later this year.
Image and video: with grok-imagine, generative video models are in the API too.

Closed and Tightly Bound to X

Grok is a cloud product through and through. There are no open weights, no download, no self-hosting: the only point of contact is the xAI API. This is compounded by the interweaving with the X platform: inputs and responses run over the same infrastructure that handles the social-media data streams. That is no side effect but strategy: real-time access to X is one of the few genuine differentiators over competitors. For users it means: confidentiality is only as large as the trust placed in a US platform.

One detail from the recent past is worth noting: xAI released Grok 2.5 under Apache 2.0 in August 2025, a one-off turn toward openness. The current Grok 3 and Grok 4 models do not follow that path; they remain closed. So anyone hoping for a self-hostable Grok variant currently gets only the older 2.5 generation.

The Caveat

Grok is consistently closed: no open weights, no self-hosting, data runs through xAI and is tightly interwoven with the X platform. For confidential business data, and for anyone who takes data sovereignty seriously, that is a disqualifier, no matter how fast the model is.

What We Read Into It

Even so, we do not lose sight of Grok. The architectural decisions xAI makes (multi-agent setups, context windows of two million tokens, native multimodality) chart the direction the whole field is moving in. What sits behind a closed API today typically turns up a quarter or two later as an open implementation. For us, then, Grok is less a tool than a compass: it shows which capabilities will soon arrive in the open space. For concrete SME work, though, a different yardstick counts: the model must run under GDPR control in your own server room. Speed alone is not enough.

Our Take

We track Grok as an indicator of pace and features: what xAI demonstrates often shows up months later as an open alternative elsewhere. It still does not belong on our self-hosting path. Speed is no substitute for data sovereignty — and in the SME world, sovereignty is the yardstick.