Grok 4.7 Released: 2.1 Trillion Parameters, SpaceX Data, $2/M Pricing, Competitive Benchmarks

The Grok 4.7 release came on September 21, 2026 from xAI — 2.1 trillion parameters, a 40 percent increase over Grok 4.6, trained partly on SpaceX engineering data, priced unchanged at $2 per million input tokens, and scoring competitively on CursorBench and AA-Briefcase.

The Numbers

xAI released Grok 4.7 on September 21, 2026. The headline figure is 2.1 trillion parameters — a 40 percent increase over Grok 4.6's 1.5 trillion. The pricing is unchanged: $2 per million input tokens, $6 per million output tokens, with cached input at $0.50 per million.

The model is available on the xAI API, through Cursor, and through Grok Build. The consumer-facing Grok app at the $30-per-month tier still serves Grok 4.6.

Elon Musk, announcing the Grok 4.7 release, acknowledged one tradeoff openly: Grok 4.7 is "slightly slower to serve" than 4.6, but with "even better token efficiency."

That combination — bigger, slower per request, more efficient per token — is the shape of a model that is optimized for workloads where the cost of a single call matters less than the total cost of getting the job done.

Where the 2.1 Trillion Figure Comes From

The 2.1 trillion parameter count traces back to Musk's claims about the model. It was flagged in pre-release reporting as early as early September, and it has been widely repeated across tech outlets since the Grok 4.7 release.

It is worth noting that Musk's parameter claims have historically been aggressive. The figure should be treated as a Musk-sourced number — useful, widely cited, but not independently verified in the way a third-party benchmark score is. That said, the release is real, the pricing is real, and the benchmarks are real.

The SpaceX Data Angle

The distinctive element of Grok 4.7's training is the use of SpaceX engineering data. Musk's company operates both xAI and SpaceX, and the Grok 4.7 release marks the clearest public signal yet that the two entities' data are being used to train xAI's models.

What exactly that data is has not been spelled out in detail by xAI. The reporting around the Grok 4.7 release has pointed to SpaceX engineering datasets generally — the kind of data generated by a company that builds and launches rockets, runs satellite operations, and maintains a large manufacturing and telemetry pipeline. The specifics are not public.

What is reportable is the structure of the claim: Grok 4.7 is larger than its predecessor, and a meaningful part of that increase is tied to data that only one company in the industry has access to. That is a competitive data point. It is also, depending on your view, a reason to be interested in what a model trained partly on aerospace engineering data can do.

The Benchmarks

Grok 4.7's benchmark scores place it just behind the leading models — not at the front of the pack, but close enough to matter.

On CursorBench 4.0, a code-engineering benchmark, Grok 4.7 scores 46.3 percent. On AA-Briefcase v1.1, it scores 1,657. EEBench is the only engineering row where Grok 4.7 leads.

The CursorBench and AA-Briefcase numbers are specific and they are reportable. They place Grok 4.7 in the conversation with the other frontier models without claiming the top spot. That is an honest position for a fourth lab's flagship.

The Five-Lab Landscape

With the Grok 4.7 release, the five-lab picture is now complete for the moment.

OpenAI has GPT-6 Astra — rated "Critical" on ExploitBench — and is previewing GPT-6 Cyber at DevDay on September 29. GPT-6 Sol and Luna are shipping. Anthropic has Claude Opus 5.5 and Claude Fable 5.1, with Fable 5.1 being the cost leader in its class. Google DeepMind has Gemini 3.8 and is pushing Gemini 4 through post-training with an "as soon as possible" timeline from its new chief, Koray Kavukcuoglu. And now xAI has Grok 4.7.

Grok 4.7 is not the biggest model. It is not the cheapest. It is not the fastest. But it is 40 percent larger than its predecessor at the same price, trained in part on data no competitor has, and it is competitively placed on the engineering benchmarks that matter to the Cursor and developer audience.

That is a coherent position for a model in a competitive market.

The Pricing Signal

The unchanged $2-per-million-input-token pricing at 40 percent more parameters is a concrete data point in the pricing conversation.

OpenAI and Anthropic both moved toward cheaper, broader access in mid-September with their simultaneous model launches. Gemini's pricing has followed its own path. Grok 4.7 holds the line where 4.6 was, which means the per-token economics improve at the same price point — more capability for the same cost, even if each individual request is a little slower.

For developers and enterprises comparing models on cost and capability, that is a useful signal. It says xAI is not using the parameter increase as an excuse to raise prices. It is absorbing the cost of the larger model at the same price.

What Is Not Yet Clear

There are open questions around the Grok 4.7 release that the current reporting does not resolve.

The exact composition of the SpaceX engineering data is not public. We know it is there; we do not know precisely what it consists of — satellite telemetry, manufacturing data, launch data, or some combination.

The full benchmark picture relative to GPT-6 Astra, Claude Opus 5.5, and Gemini 3.8 is not as complete as the CursorBench and AA-Briefcase numbers. The model sits "just behind" the leading models on the available scores, but the precise ranking across all relevant benchmarks is not settled.

And the Musk-sourced parameter claim, while widely repeated, has not been independently verified in the way a third-party evaluation would verify it.

None of that is unusual for a major model release. It is the normal state of uncertainty around the first 48 to 72 hours of a launch.

What Grok 4.7 Is and Is Not

The Grok 4.7 release is a significant model launch from a fourth major AI lab that AIPress has not covered in depth until now. It completes the five-lab picture and fills a coverage gap.

It is not a model that leads the field on benchmarks. It is not a model that breaks new ground on safety. It is, instead, a model that makes a coherent competitive argument: bigger than its predecessor at the same price, trained on data no one else has, and positioned for the engineering workloads where CursorBench and AA-Briefcase matter.

For a lab that has historically made its name on personality, candor, and a distinct point of view, Grok 4.7 is a notably conventional release. Bigger. Same price. Slower per call. Better per token. Trained on proprietary data. Competitively placed.

That is a solid model release. It is not a surprise. It is a statement.

The Bottom Line

The Grok 4.7 release delivered 2.1 trillion parameters — a 40 percent increase over Grok 4.6 — trained in part on SpaceX engineering data, priced unchanged at $2 per million input tokens, and slightly slower to serve with better token efficiency.

Its CursorBench 4.0 score of 46.3 percent and AA-Briefcase v1.1 score of 1,657 place it just behind the leading models — not at the front, but close enough to matter.

In a five-lab landscape now populated by GPT-6 Astra, Claude Opus 5.5 and Fable 5.1, Gemini 3.8, and Grok 4.7, xAI's latest is a solid, conventional, competitively reasoned release. It does not try to be everything. It is, instead, a claim on a specific position in the market: bigger, same price, proprietary data, engineering-focused.