According to TrendForce's Beating Monitor, shortly after announcing the start of the Grok 4.5 beta testing, Musk urgently posted to temper his previous extravagant claims. He had initially declared that the performance of the new model "may have surpassed Opus," but quickly backtracked to clarify that the V9 base model is not "amazingly good," but rather a "flagship tool" on par with Opus. This rapid self-denial was evidently aimed at reducing expectations for the model, which had not yet been publicly tested.
Musk further explained the reason for the urgent adjustment to the new model. He revealed that due to the training of the 1.5T version using only the newly acquired Cursor code data as "incremental training" (which was less effective than initial training), the technical performance was compromised. To salvage the situation, xAI initiated the training of a 2T parameter large model several weeks ago, integrating all data into pre-training and scheduling it for release in August.
In addition to the Cursor team's technical contributions to fine-tuning and reinforcement learning, in order to timely unveil the August 2T "new cake," Musk even resorted to cross-domain mobilization. Dozens of top engineers from the Starlink and Starship projects have been seconded to assist in AI training. Faced with a hemorrhage in the core development team, rocket-building experts were brought in to fill the gaps in AI engineering, making this modeling extravaganza appear even more hurried.
Furthermore, Musk has outlined an even larger plan, claiming that the real performance leap will have to wait for another 3 months. At that time, the entire training and inference stack will be refactored in C/C++, drastically eliminating most software layers to achieve ultimate hardware compatibility with the GB300 chip.
