T he upcoming AI model from xAI, which has been described as a 2 trillion-parameter model, is expected to be known as Grok 4.6 and will undergo the initial training phase in one week’s time. The reveal was made by Elon Musk in response to a tweet on X by Min Choi, a fellow AI commentator.
In response to a chart that shows the cost-effectiveness and intelligence of top AI models, Musk said that the xAI's 2T model, which is superior to its 1.5T model in all ways, will be done with training next week, and that the 2T model may be able to outperform Kimi but with speeds and token efficiency similar to that of the 1.5T Grok 4.5. The statement, unlike many statements by Musk, lacked any sort of marketing blabber.
It must be noted that the background of this discussion is relevant.
The discussion started after Min Choi provided a chart from Artificial Analysis that compared the cost of running various AI models with their task completion capabilities. According to this chart, Grok 4.5 was in a good place, maintaining low costs and speed, despite being lighter than some of OpenAI and Google's heavier models.
The new model is quite an upgrade from the current production model at xAI. As per Elon Musk, the model is much better than the 1.5T in all aspects, which is not a very ambiguous statement, considering that it is meant to be a direct upgrade rather than a specific version. Grok 4.5, built using the V9 foundation by the company, is three times bigger than the current line of production models, Grok 4.x, which has a 500 billion parameter configuration.
To understand how well xAI does compared to other companies: Kimi K3, which is the largest open-source AI model in the world with 2.8 trillion parameters, though, according to Min Choi of xAI, Musk's models consume fewer resources – studies have indicated that the cost of tasks on Kimi is three times the cost of xAI's models.
There is a critical point that is being overlooked in some reporting, and that is the fact that training is completed does not mean that the product is launched.
The Grok 4.6 version of the model was announced by Musk on July 18, 2026, as the direct successor to the previous version of the Grok 4.5. The pre-training process of this model was expected to be completed around the first week of July 2026.
However, at the time of the publication, the product has yet to be released—there are no official details on pricing, benchmarking, context window, or even the date of the release, but community projections place the actual date four to six weeks following the completion of the training period.
This actually aligns with past reporting. A 2-trillion parameter model had been previously reported in late June as being in the final stages of the training process, with a launch date expected for August, although the company had yet to officially commit to any exact dates.
Early July reporting had also identified that the 2-trillion parameter model was set to be completed and launched in August following its completion in.
There are several unknowns still left, and these are precisely the factors that will decide whether or not Grok 4.6 will live up to the narrative established by Musk.
This includes the date of the release itself because training completion does not always equate to a launch during the following week, with reinforcement learning and safety tests usually taking between four and six weeks at xAI's pace.
There are no benchmark scores provided in terms of general reasoning, code generation, or agency abilities, as well. The context window size of the model is still unknown, as well as its architecture, which is yet unclear in terms of density and mixture of experts, and the number of parameters that will be used.
Lastly, there are no details concerning pricing or the addition of new capabilities, including visual, auditory, or video processing.
It should be noted that Elon Musk has had a precedent of aggressive time frames for the xAI product launch, with Grok-1.5 being released "next week" at the time of its first introduction, following Grok 2 being advertised as exceeding the current AI "on all metrics." This is not to say that this precedent.
Photo: @Screenshot/@elonmusk on X
