Elon Musk has made some large claims about Grok 5: a 6 trillion parameter model, native understanding of text, images, video and audio, and a 10% chance of reaching true AGI. Almost none of it has been confirmed by xAI itself. Here’s what’s actually been verified, and what’s still just a prediction.
What xAI Has Actually Confirmed
In January 2026, xAI’s Series E announcement stated that Grok 5 was in training, and that the company had over one million H100 GPU equivalents across its two Colossus data centres. Musk later confirmed training was running on Colossus 2. That’s the entire confirmed record. There’s no official parameter count, no training compute figure, no context window, no price, and no architecture paper.
A Release Date That Kept Moving
Musk first said Grok 5 would launch before the end of 2025. That slipped to “next month”, then “a few weeks”, then the first quarter of 2026. That deadline also passed. By September 2026, Musk had shifted his framing entirely, describing a staircase of smaller releases, 4.7, 4.8, 4.9, with Grok 5 positioned as a longer-term goal rather than the next big thing. That’s a meaningful change in tone from the confident predictions made a year earlier.
The Claims Still Unverified
The 6 trillion parameter figure, doubling Grok 3 and Grok 4’s claimed 3 trillion, comes from a Musk interview, not a published spec sheet. Nobody has confirmed whether the model is dense or a mixture of experts, or how many parameters are actually active during use. Musk’s claim of a 10% chance of reaching AGI is, by his own description, a personal estimate, with no published test or definition behind it.
What Grok 4.7 Tells Us Instead
Since Grok 5 can’t be tested, Grok 4.7, released in September 2026, is the clearest evidence of where xAI is actually heading. It uses a larger base model and a longer reinforcement learning phase focused on tasks that take many hours to complete. The results were real but narrow: xAI’s own benchmarks show large jumps in coding and terminal-based agent tasks, while independent testing found its general intelligence score barely moved and some measures actually slipped. The gains came from reasoning for roughly twice as many tokens per task, which affects real-world cost.
Where It Still Trails The Competition
By xAI’s own published figures, Grok 4.7 trails Claude Opus 5.5 on coding benchmarks and sits behind GPT-6 Astra on others. Independent testing puts it below both flagship rivals on general intelligence. Pricing looks competitive on paper, but because Grok 4.7 uses more tokens to reason, the real cost per completed task narrows that advantage considerably.
What Would Actually Prove The Claims
A model card, a confirmed architecture, independently verified results across standard benchmarks, a genuine working demonstration of real-time video understanding, and a published safety report. Until any of that appears, Grok 5 remains an unverified claim, and Grok 4.7 is the only real evidence of xAI’s direction so far.
If you’d rather judge AI tools on what they actually deliver than on headlines, Claude LaunchPad is the free starting point for setting up Claude properly.
