Grok 4.6 and Grok 4.7: Inside Musk's Two-Week Frontier Model Cadence

Elon Musk used a Friday post on X to lay out a release schedule that would be aggressive for any AI lab and is close to unprecedented at the frontier: Grok 4.6 lands in two weeks, Grok 4.7 two weeks after that. Two new frontier models inside a single month, from one company.

Do the arithmetic on the July 24 announcement and you get roughly August 7 for Grok 4.6 and roughly August 21 for Grok 4.7. No frontier lab has held that kind of pace at this scale. That's the headline, and it's also the risk — a schedule this tight leaves almost no room for anything to go sideways.

What Grok 4.6 Actually Is

Grok 4.6 is a 2-trillion-parameter model. That's a meaningful jump from Grok 4.5, which runs at 1.5 trillion.

Musk confirmed both the name and the scale on July 18, and his framing was blunt: the new model would beat its predecessor in every way, with performance targeted at or above Moonshot AI's Kimi K3. The model wrapped initial pre-training during the week of July 20, which is what makes the two-week window plausible rather than aspirational.

The pitch isn't purely about raw capability, though. Musk has described Grok 4.6 as matching what Kimi K3 can do while holding onto the traits that defined Grok 4.5 — speed and lower cost per token. That's a specific competitive posture. Not "we're bigger," but "we're comparable and cheaper to run."

SpaceX Data Enters the Training Run

A separate disclosure on July 21 added a detail worth pausing on. Musk said SpaceX engineering data would be folded into supplemental training for the 2-trillion-parameter run, with material restricted under ITAR export controls left out.

That's an unusual input for a general-purpose frontier model. Proprietary aerospace engineering data isn't something a competitor can scrape or license, and the explicit ITAR carve-out signals xAI is drawing the compliance line before regulators do it for them. Whether it shows up as measurable capability gains in engineering or physics tasks is the open question — nobody has published benchmarks on that yet.

The Kimi K3 Pressure

None of this is happening in a vacuum. Moonshot AI shipped Kimi K3 on July 16 — days before Musk's disclosures — at 2.8 trillion parameters, making it the largest open-weight large language model available.

That's the context for the entire Grok 4.6 timeline. A Chinese lab put out an open-weight model larger than anything xAI has released, and the response arrived within a week.

Where Kimi K3 Landed

CNBC's reporting put Kimi K3 in a specific position on the leaderboard, and the nuance matters:

So Kimi K3 isn't the outright leader. It beats the previous generation of American frontier models on coding and agentic work while still trailing the current top tier overall. For an open-weight release, that's a strong showing — and it explains why a competitor's roadmap suddenly compressed.

Grok 4.6 is being positioned to close that specific gap. Match Kimi K3's capability, keep xAI's cost and latency advantages.

Grok 4.7 and the 6-Trillion-Parameter Question

Here's where the picture gets thin. Details on Grok 4.7 are sparse, and xAI hasn't confirmed a parameter count for it.

What's known comes from earlier reporting on the company's roadmap: models at 6 trillion parameters were already training on the Colossus 2 supercluster, running alongside the 2-trillion-parameter work. That's the pipeline. Whether Grok 4.7 is one of those 6-trillion runs, another 2-trillion variant, or something in between hasn't been stated.

The pipeline detail is the interesting part regardless of how 4.7 shakes out. Training multiple model sizes in parallel on the same infrastructure is what makes a two-week release cadence structurally possible. You aren't waiting for one run to finish before starting the next — you're picking which finished run to ship.

What Grok 4.5 Established

Grok 4.5 launched July 8 and remains SpaceXAI's current flagship. Pricing came in at $2 per million input tokens and $6 per million output tokens, available through the xAI API and a Cursor integration.

Musk's own description of Grok 4.5 was that it hit Opus-class performance while being faster, more token-efficient, and cheaper. That's the template Grok 4.6 is meant to extend upward — same value proposition, higher capability ceiling.

The pricing detail is worth holding onto, because it's the clearest read on strategy. xAI isn't competing on being the single most capable model. It's competing on the ratio: near-frontier capability at a cost that makes high-volume agentic workloads viable.

Why the Cadence Is the Real Story

Strip out the parameter counts and what's left is a claim about release velocity.

Shipping two frontier models in four weeks means the constraint has moved. It's no longer compute or training time — it's evaluation, safety review, and deployment. A lab that can compress those steps to two weeks per model is operating on a different clock than its competitors.

The stated timeline is exactly that: stated. Musk announced it, and announced timelines have a way of stretching. But the supporting facts line up better than usual — pre-training finished the week of July 20, parallel runs are already underway on Colossus 2, and the competitive trigger from Kimi K3 is concrete rather than speculative.

If August 7 comes and goes without Grok 4.6, the cadence claim collapses. If it lands, the four-week question becomes whether anyone else can answer at that speed.