Grok 4.3

SpaceXAI's fast, reliable model with strong tool calling and instruction following. It is the only Grok here whose reasoning can be turned off entirely, and it carries the largest context of the three at a million tokens, which makes it the natural pick for high-volume agentic work and long-document tasks. Text and image input, text output.

xAI never published a launch post for Grok 4.3. The model appeared in the catalogue and the API, and the lab's own model card is the only first-party description of it: a fast, reliable model with strong tool calling and instruction following. That absence is worth knowing while reading this page. For most models the lab's own claims and the independently measured table say different things and are worth comparing; here there are no lab-run benchmark claims at all, so the Benchmarks panel above is the whole measured picture rather than half of it.

The reasoning dial is what separates this model from the newer Grok releases. xAI documents four settings - none, low, medium and high - and none genuinely means no reasoning at all. Grok 4.6 and Grok 4.5 cannot be talked out of thinking; this one can, which makes it the Grok to reach for when a step inside an agent loop needs an answer rather than deliberation. Function calling and structured outputs are both supported, and it takes text and images in and returns text.

The million-token window carries a pricing step that a flat rate card cannot show. xAI bills any request whose prompt reaches 200,000 tokens at double the standard rate, and at that higher rate for every token in the request rather than only the tokens past the threshold; cached input doubles with it. So the first fifth of the window is the cheap part and the rest is not, which is something to design a long-document pipeline around rather than discover in an invoice. Batch requests are discounted against standard rates.

xAI serves the model from three regions and, on its own service, allows 37 requests a second and ten million tokens a minute. Those ceilings, the low rate and the ability to switch reasoning off entirely are what make it a sensible default for high-volume, low-stakes agent traffic - classification, extraction, routing, tool dispatch, first-pass summarisation - where the newer models' mandatory thinking is a tax paid on every call.

Where it stops is equally clear. xAI's own recommendation for code and for chat is its current flagship, not this model; Grok 4.3 is the older, cheaper generation, kept for work where price and latency decide the outcome. And like every Grok it knows nothing about events after its training data unless a server-side search tool is enabled on the request.