LLMs

xAI’s Next Model Outperforms Grok 4.5 in Every Metric — and It Hasn’t Finished Training Yet

Elon Musk has confirmed that xAI's next model — a 2-trillion-parameter system — already outperforms Grok 4.5 on every internal benchmark, with initial training expected to finish within the week.

By Abhinav Ramaswamy July 18, 2026 5 min read

Grok 4.5 Set a New Bar. xAI Is Already Racing Past It.

Grok 4.5 only recently established itself as one of the most capable and cost-efficient frontier models available — praised for its blend of raw intelligence, blistering inference speed, and remarkably low token consumption. Now, before the AI world has fully digested what Grok 4.5 means for the competitive landscape, Elon Musk has confirmed that xAI's next model has already surpassed it across every single dimension — and it hasn't even completed initial training.

In a post on X, Musk revealed that the upcoming 2-trillion-parameter model from xAI's SpaceX AI division is already outperforming Grok 4.5 on every internal benchmark, with training expected to wrap up within the week. This is not a roadmap promise or a speculative leak — it is a public confirmation from xAI's founder that the next generation of Grok is already more capable, with the data to back it up.

Elon Musk on X:

https://x.com/i/status/2078289996323148076

What Musk Actually Said

The confirmation carries several specific and notable claims. The incoming model has 2 trillion parameters, placing it in a scale category that very few systems have publicly reached. More remarkably, Musk noted it already surpasses Grok 4.5 on every measured dimension — and goes further, suggesting it may even challenge Kimi K3, Moonshot AI's 2.8-trillion-parameter open model that has been dominating leaderboards since its launch.

But the claim that will resonate most with developers and enterprise users is what xAI is not sacrificing in the process. Speed and token efficiency — two of the defining advantages that made Grok 4.5 stand out — are being kept close to parity. The next Grok is being built to be smarter without becoming slower or more expensive to run.

Why Grok 4.5's Efficiency Made This So Hard to Do

To understand why this announcement is significant, it helps to understand what made Grok 4.5 unusual in the first place. Most frontier model scaling follows a familiar pattern: more capability comes at the cost of higher latency and greater token usage, making the model more expensive per task. Grok 4.5 broke that pattern by delivering frontier-level intelligence while maintaining one of the lowest costs per task among leading models. That combination — not just raw benchmark scores, but practical economics — is what put it in a category of its own.

Scaling further without degrading those efficiency properties is genuinely difficult. It requires architectural discipline, not just more compute. The fact that xAI is explicitly signaling that speed and token efficiency remain close to Grok 4.5 levels suggests this is not simply a brute-force parameter expansion — it is a targeted intelligence upgrade that preserves what made the previous model commercially compelling.

The Kimi K3 Benchmark Matters

Musk's suggestion that the new model may surpass Kimi K3 is a meaningful benchmark to invoke. Since its official release, Kimi K3 has been one of the most impressive open-weight models ever released — a 2.8-trillion-parameter system with native vision, deep reasoning, and enough coding capability to top LMArena's Code Arena leaderboard, defeating every major Western model in head-to-head human evaluations. Positioning xAI's next model as potentially competitive with Kimi K3 sets a very specific, very public performance target — and one that the broader developer community will be watching closely.

It also frames the next phase of frontier AI competition in sharper terms. The race is no longer simply about who can deploy the most parameters. It is about who can hit the highest capability ceiling while keeping the model fast and cost-effective enough to be practical at scale. xAI is explicitly claiming to be pursuing both at once.

What "Next Week" Actually Means

Musk's statement that initial training is expected to finish within the week does not mean a public release is imminent. Post-training work — alignment, safety evaluation, fine-tuning, and API infrastructure — typically follows initial training and can itself take weeks or months. However, finishing initial training is a genuine milestone, and the fact that internal benchmarks already show superiority over Grok 4.5 before that process is even complete suggests the model has headroom to improve further through post-training.

For context on where xAI's model roadmap is already headed, the SuperGrok Heavy subscription tier has already been restructured to bundle X Premium+ — a signal that xAI is actively repositioning its product stack ahead of a significant model upgrade.

Bigger, But Not Just Bigger

The framing Musk used is worth noting on its own terms. The next Grok model, he emphasized, is not simply getting bigger — it is getting smarter while staying fast and efficient. That distinction matters because the AI industry has spent years conflating parameter count with capability. The most important competitive frontier right now is the combination of peak intelligence and practical deployability, and that is precisely the combination xAI is claiming to have achieved at 2 trillion parameters.

Whether that claim holds up under independent benchmarking remains to be seen. But the public confirmation, the specific performance comparisons to both Grok 4.5 and Kimi K3, and the imminent training completion timeline make this one of the more concrete and falsifiable AI announcements in recent months. The next few weeks will make clear how much of it holds.