SAN FRANCISCO — xAI is closing in on the finish line for Grok 4.6, with Elon Musk revealing that the model — built with two trillion parameters — will complete its initial training run as soon as next week.
Musk shared the timeline on July 18, adding that Grok 4.6 is designed to surpass Moonshot AI's newly released Kimi K3 while holding inference speed and token efficiency close to the current Grok 4.5. It is the clearest sign yet of how quickly xAI, now developed under the SpaceXAI division, is scaling its models.
A Bigger Brain, Similar Speed
Grok 4.5, known internally as V9, runs on roughly 1.5 trillion parameters. Grok 4.6 nearly doubles that count, and Musk says the new model is superior in all aspects to its predecessor while keeping its inference speed and token efficiency largely on par. That balance matters: bigger models often become slower and pricier to run, and xAI is signaling it can add capability without sacrificing responsiveness.
The company has been on a rapid release cadence, layering new capabilities onto Grok at a pace few rivals can match, including its steady stream of Grok Imagine media features. A two-trillion-parameter jump extends that momentum into the core reasoning model.
Chasing Kimi K3 on Cost and Quality
The timing is competitive. Moonshot AI launched Kimi K3, a 2.8-trillion-parameter model, on July 16, and Musk framed Grok 4.6 as the response. The pitch is efficiency: Grok 4.5's per-task cost sits near $0.31, well below Kimi K3's roughly $0.94, and xAI wants Grok 4.6 to close the performance gap while preserving that cost advantage.
xAI has also worked to keep users and creators on its side, including recent moves to protect creators and return their payouts. Pairing a stronger model with a friendlier platform is central to how the company hopes to win share from larger AI labs.
The Road to Grok 5
Grok 4.6 is only one step in a larger plan. xAI is already training models it refers to as Grok 5 at six-trillion and ten-trillion parameters on the Colossus 2 supercluster, the kind of scaling that requires the enormous compute footprint SpaceX has been assembling. According to xAI, the roadmap runs well beyond the current generation.
If Grok 4.6 finishes training on schedule and lands its efficiency targets, xAI will have answered its newest rival within days of its debut, then set its sights on models several times larger. For a company that measures progress in trillions of parameters, next week is only the beginning.
