Elon Musk announced that Grok 4.7, the next version of xAI's chatbot, will be released in 10 days, around September 12, 2026. Musk claims the model will surpass all current AI systems, citing a jump in parameters to 2.1 trillion from 1.5 trillion in Grok 4.6. The model runs slightly slower but uses tokens more efficiently, and Musk says SpaceX's training corpus makes it the best at real-world engineering.
The new model's specs
Grok 4.7's parameter count is up 40% from its predecessor, which Musk says allows for deeper reasoning. He acknowledged the trade-off: the model is a bit slower in response time, but it compensates by consuming fewer tokens per task. That efficiency could matter for developers who pay per token, though Musk didn't provide specific benchmarks for the new version.
The training run finished in mid-August, according to Musk. He said SpaceX data was folded in through supplemental training, giving the model a grounding in aerospace engineering that other AI labs can't match. Whether that translates into better performance on general tasks remains to be seen, but Musk is betting on it.
How it stacks up against rivals
Grok 4.6, released on August 12, scored 61 on the AA Intelligence Index, tying GPT-5.6 Sol Max and trailing Claude Fable 5 Max's 62. Grok 4.5 had scored 56. On GDPVal-AA v2, Grok 4.6 led with 1753. But on Terminal-Bench v3.0, it managed only 26%, well behind GPT-5.6 Sol Max's 34.6%.
Grok 4.5 topped AutomationBench-AA at 51.4% with a cost of $0.34 per task, but it also logged 0.63 guardrail violations per task, worse than Claude Opus 4.8's 0.55. Those numbers suggest Grok has strengths in some areas and weaknesses in others. Musk's claim that 4.7 will surpass all current models is a high bar, especially with OpenAI's new Astra model on the horizon.
SpaceX data and training
The SpaceX connection is the unusual part. Musk said the training corpus includes proprietary data from SpaceX, which he argues gives Grok 4.7 an edge in real-world engineering problems. That's a differentiator no other lab can copy, but it also raises questions about how much of that data is actually useful for general AI tasks. Musk didn't elaborate on what the supplemental training involved or how it was weighted.
He did say the initial training finished in mid-August, leaving a few weeks for fine-tuning and safety checks before the September 12 release. That timeline is tight, but xAI has shipped on schedule before.
The competitive landscape
OpenAI announced it will release a new Astra model soon, though no date was given. That puts pressure on xAI to deliver on Musk's promises. If Grok 4.7 doesn't beat the competition in independent benchmarks, the hype will backfire. But if it does, the SpaceX data angle could become a template for other specialized training sets.
The release is set for around September 12. Whether Grok 4.7 actually surpasses every model on the market will be tested within days of launch, as third-party evaluators run their standard suites. Musk's claim is bold, but the numbers will tell the real story.
