Elon Musk dropped a quiet bombshell on X on July 18. xAI's next model — expected to be named Grok 4.6 — will complete initial training next week. 2 trillion parameters. Better than the 1.5T version in every way. Speed and token efficiency on par with Grok 4.5.
His phrasing: "It might be able to exceed Kimi." "Might" is doing a lot of work.
Three days earlier, Kimi K3 launched at 2.8 trillion parameters. It topped Arena AI's front-end coding leaderboard with 1,679 points, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). Independent benchmark score: 57 — ahead of Grok 4.5's 54. Musk saw it. He replied: "Impressive."
For Musk, that was unusual. He has called every OpenAI, Google, and Anthropic model release "overwhelming" or "crushing it." He has never been shy about claiming dominance. Against Kimi, he said "might." That says more about the ground shift than any benchmark number.
Kimi noticed. The official response: "Welcome to the 2T+ club."
Grok 4.5 cost $0.31 per coding task. Kimi K3 cost $0.94. Musk's aim is not just catching up — it is doing it at the same unit economics as before.
Colossus is running. 200,000 GPUs are burning. xAI has revealed it is also training 6T and 10T models designated Grok 5 on Colossus 2.
The timeline is tight. If Grok 4.6 finishes training next week, it could land in early August. That puts it directly against Kimi K3's open-weight release on July 27, DeepSeek V4's stable release, and whatever Anthropic and OpenAI are holding.
For Musk, this matters differently. xAI is not a public company. It does not have to answer to Wall Street about the timeline. But the narrative matters — the founder of the fastest AI lab in the world just admitted "might" about a model from China. That is a new kind of signal.
P.S. A few weeks ago, Musk said xAI's speed was the competitive advantage. Now he is measuring himself against a model that was benchmarked before his next one finishes training — and "might" is not a comfortable word for someone who usually shouts "overwhelming."
