SpaceXAI is expected to ship its next flagship model, Grok 4.7, in mid-September. Elon Musk said on X on September 2 that the model was roughly 10 days away. The stated scale is about 2.1 trillion parameters, roughly 40 percent larger than the current Grok 4.6, and the training run reportedly drew on rocket and satellite engineering data. The developer documentation, however, still lists no model ID and no pricing for it.
A Release Window Described Only as About 10 Days
The entire basis for the timing is a single line Musk posted on X: about 10 days out. Counting from September 2, that lands around September 11 or 12. No dated press release has followed, and that post remains the only firm signal.
The gap from the previous generation is short. Grok 4.5 arrived in July and Grok 4.6 on August 12, which puts the next release roughly a month later. The cadence has tightened as competition over long-running agent workloads has intensified.
Meanwhile, the developer documentation was last updated on August 21 and still shows grok-4.6 as the newest entry. There is no grok-4.7 model ID, no pricing, no context length, and no benchmark card. For anyone planning an integration, nothing usable has been published yet.
How to Read the 2.1 Trillion Parameter Figure
Scale is an easy number to quote and a hard one to interpret. By Musk's account, Grok 4.6 sits at roughly 1.5 trillion parameters and Grok 4.7 at roughly 2.1 trillion, a gap of about 40 percent.
That figure alone does not reveal the parameters that actually run. In a mixture-of-experts design, the total parameter count and the number used per token are different things. Active parameters, sparsity, expert routing, and inference cost have not been disclosed.
Parameter count also does not map linearly onto performance. Architecture, training data quality, post-training work, and inference-time efficiency all contribute. A 40 percent increase in scale does not imply a 40 percent improvement in what users feel.
The Bet on SpaceX Engineering Data
The other headline is the training data. A supplemental training phase reportedly incorporated SpaceX engineering records, telemetry, internal documents, and Starlink satellite data.
That fits the company's position. Grok is developed by SpaceXAI LLC, and the brand guidelines state plainly that it is a separate company from X, formerly Twitter. Few organizations can route operational data from rockets and satellites into language model training.
The hoped-for payoff is that the model handles real engineering problems differently from text-trained systems. That assumption has not been independently validated at scale. Adding design data may sound like it should produce better design work, but intuition and evidence are separate matters.
Musk has said Grok 4.7 outperforms Grok 4.6 across the board and improves token efficiency, while noting that inference speed may drop slightly given the larger size.
Where Grok 4.6 Stands Today
Grok 4.6, released on August 12, is the benchmark the new model will be measured against. It offers a 500,000-token context, with API pricing of 2 USD (about 310 yen) per million input tokens and 6 USD (about 940 yen) per million output tokens, and a fast variant at twice that rate. Its knowledge cutoff is February 1, 2026.
1 USD = 156 JPY (as of September 5, 2026)
On the published evaluations, Grok 4.6 scored 61 on the Artificial Analysis Intelligence Index, a composite of nine benchmarks, matching GPT-5.6 Sol Max. Fable 5 Max scored 62. Broken out, it recorded 1753 on GDPVal-AA v2, 69.9 percent on CursorBench v3.2, 65.9 percent on DeepSWE v1.1, and 26 percent on Terminal-Bench v3.0, with rival models clearly ahead on some of those. It is aimed at long-running agent work and is available through Cursor, Grok Build, the API, OpenRouter, Vercel, and Cloudflare.
If Grok 4.7 truly improves on every axis, these are the numbers that will show it.
Judgment Should Wait for the Numbers
History is worth noting here. When Grok 4.6 launched, observers pointed out that a complete set of benchmarks was not published immediately. If Grok 4.7 arrives the same way, the claim that it beats every other model belongs in marketing territory for the time being.
For anyone weighing adoption, the checklist is clear: active parameter count, context length, input and output pricing, and benchmarks a third party can reproduce. Only once those are filled in does replacing an existing workflow become a decision rather than a guess.
Summary
Grok 4.7 is expected in mid-September at roughly 2.1 trillion parameters, about 40 percent larger than Grok 4.6. The use of SpaceX engineering data in supplemental training is the distinguishing claim, but active parameters, pricing, context length, and independently verified benchmarks all remain unpublished. The developer documentation still tops out at grok-4.6, so the only firm facts are an approximate release window and a claim about scale. Waiting for the numbers is the safer position.
