Musk Announces 2 Trillion Parameter Grok 4.6 to Answer Kimi K3

Musk's answer to China's advance

Elon Musk revealed, in a post on X published on July 18, 2026, that xAI is training Grok 4.6, a new 2 trillion parameter model, and that its first full training cycle should wrap up within the week. According to the post, echoed by Dataconomy, WION, TechBriefly and TradingKey, Musk described the model as better than xAI's current flagship, the 1.5 trillion parameter Grok 4.5, in every way, while keeping speed and token efficiency close to the current model.

The trigger: Moonshot AI's Kimi K3

The announcement comes two days after China's Moonshot AI launched Kimi K3, an open weight model with 2.8 trillion parameters, a 1 million token context window and visual understanding, on July 16, 2026. Shortly after launch, Kimi K3 took the top spot on Arena's blind frontend coding leaderboard, an independent AI model evaluation platform, beating every American flagship model, according to reporting from KuCoin and WION. Musk said Grok 4.6 might be able to exceed Kimi, a direct response to the Chinese model's rise.

Cost per task enters the race too

Beyond raw performance, the race is also about cost. According to data cited by TradingKey and kie.ai, Grok 4.5 completes tasks at an average cost of $0.31 each, versus $0.94 for Kimi K3, a nearly threefold difference. If Grok 4.6 keeps the token efficiency Musk promised even while tripling the model's size, xAI could widen its cost advantage right as the Chinese model dominates capability rankings.

An announcement race that hasn't become a product yet

It's worth stressing that, as of this article's publication, Grok 4.6 remains in training, with no release date for end users or developers. Musk has a track record of announcing optimistic timelines for xAI models that later slip, so finishing training next week does not mean immediate availability through the API or the app.

Why this race matters for whoever picks a model in Brazil

For Brazilian agencies and companies deciding which AI model to use in customer service and automation products, the Grok 4.6 versus Kimi K3 race reinforces an already visible trend: the frontier model market has turned into an ever shorter release cycle, with Chinese open models pressuring the cost of closed Western ones. The tighter that competition gets, the higher the chance that per token API pricing keeps falling for buyers, but also the higher the risk of spending engineering time integrating a model that goes obsolete within weeks.

What's clear

The race between xAI and Moonshot AI shows that the generative AI frontier no longer has a single dominant competitor and has turned into a direct contest between American and Chinese labs, each answering the other's launch within days, not months. For anyone building a product on top of these models, the practical lesson stays the same: an architecture that swaps models easily is worth more than a closed bet on a single vendor, especially at a moment when the market's fastest model can change names every couple of weeks.