Grok 4.7 released, 21 September: 2 dollars per million input tokens, 6 per million output, five benchmarks named — what the release note documents, and what it does not

A model release note is a company's own account of its own product, written for the day of the launch. It is a primary source for exactly one thing: what the company chose to say. This article reads the note for Grok 4.7, published on 21 September 2026, and records the claims, the numbers and the omissions — in that order, because the omissions are where most of the later commentary will be written.

The note is published at x.ai under the title "Introducing Grok 4.7" and carries the date Sep 21, 2026. The page's own title tag and its footer name the publisher as "SpaceXAI" and "SpaceXAI LLC" — we record the name as it appears on the page and did not check it against any corporate filing, so the legal form of the publisher is, for this article, [UNVERIFIED]. The product name on the page is Grok, and the company's news index lists the item under the same date.

The opening sentence sets the claim: "Grok 4.7 is our most capable model for coding and knowledge work. It works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date. Served at the same price and speed as Grok 4.6, it is highly competitive in its class." Every clause in that sentence is the company describing its own product; we quote it as such.

The note describes the technical change from the previous version in two sentences: "Grok 4.7 uses a new, larger base model compared to Grok 4.6. It was trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete." It adds that the model "is better at verifying its own work and managing longer context", and that it was trained "to natively understand the Grok Bot harness", which the note connects to conversational tasks and general knowledge work.

No parameter count is given. No context-window length is given. No training-data description, no compute figure, no date of the training cutoff. "Larger" is stated without a number on either side of the comparison.

Five benchmarks are named. On coding: "On CursorBench 4.0, which stresses longer-running coding tasks, Grok 4.7 is at the frontier in price-performance." No score is given. On professional work: "In GDPval and AA Briefcase, AI is asked to work on tasks done by professionals such as lawyers, nurses, and financial analysts. Grok 4.7 improves upon Grok 4.6 on both benchmarks and performs comparably to other frontier models." No scores are given for either.

The two numbers on the page are both in the safety section. The first: the model is described as "topping LatchBio's biosafety benchmark at 62.4" per cent — sixty-two point four, with the percentage sign printed on the page — in the context of "dual-use domains like cybersecurity and biological work", where the note claims it "leads on both utility for benign tasks and safe refusal on dangerous ones". The second: on "HackerBench v0.3, our benchmark for risky and malicious cyber tasks", the model is said to allow "only 3.3" per cent "of risky dual-use prompts through" — three point three — "while rarely blocking legitimate security work". The note also states that the company has "started giving select cybersecurity partners invite-only access to Grok 4.7's red-team capabilities for defense research".

Two observations about these numbers. First, HackerBench is described in the note as "our benchmark", so the second figure is a company score on a company test. Second, the note calls the model "the strongest model we've tested on refusals and jailbreak resistance" — a comparison whose reference set, the models the company tested, is not listed.

The pricing is the most concrete part of the document: "The model is priced starting at $2 per million input tokens and $6 per million output tokens. We also serve a fast variant with twice the output speed at twice the price." Read literally, the fast variant is 4 dollars per million input tokens and 12 per million output tokens; the note does not print those figures, it prints the multiplier.

On availability: "Grok 4.7 is available today in Cursor and Grok Build. It is also available through the Grok API, third-party coding harnesses, and model routers and cloud platforms." No specific cloud platform or router is named. The note does not say whether the model is available in the company's consumer chat products, and does not mention any regional restriction.

It gives no parameter count, no context length, no cutoff date, no benchmark score outside the two safety figures, and no independent evaluation. It does not name the "comparable models" against which it claims twice the speed at half the price in its summary line. It does not publish a system card or a model card; if one exists, the note does not link to it. Third-party reports about the company's roadmap — including a claim, circulating on the same day, that a successor version is planned — are not on the page, and we did not use them.

Our own work on this site is done by a small team introduced on the About page, and the tools we use are not the subject of this article. What is the subject is the reading method: quote what a document states, separate a claim about oneself from a measured figure, and write down what is missing. That is the discipline of the Uncle Sunny Academy, applied in the longer case studies of The Analyst Room to filings and reports, and applied here to a release note. Two days ago we read a model card in the same way — the Qwen-Image-2.1 card of 20 September — and the contrast is instructive: an open-weights card lists parameters and layers; a hosted-model note lists prices and benchmarks. Both are primary sources; neither is complete. Earlier entries in this series are collected under Insights.

x.ai, "Introducing Grok 4.7", x.ai/news/grok-4-7, dated Sep 21, 2026 — the opening claim, the base-model and training description, the five benchmark names, the 62.4 and 3.3 per cent figures, the pricing, the fast-variant multiplier, the availability list, the red-team access statement, the publisher name as shown on the page.

x.ai news index, x.ai/news, read 22 September 2026 — the listing of the item under the same date.

We did not run the model, did not consult any third-party benchmark, did not read any press coverage of the release, and did not verify the publisher's legal name against a registry.

Educational content. Not investment advice. This article documents a software release note; it contains no instruction to buy, sell or hold any asset or security.