Grok 4.7: what’s been said, what’s still missing

Grok 4.7 has a September 12 target if Elon Musk’s September 2 estimate holds. Pricing, access, and how well it handles ordinary work are still open questions in the sources we checked.

That makes this a good moment to separate the statements from the assumptions.

Where the date comes from

On September 2, Musk posted a release estimate that points to September 12. It is his forecast. It does not establish that the model has launched or that everyone will get access on that date. Read Musk's post.

Mark Kretschmann’s September 7 post says SpaceXAI is preparing Grok 4.7 and hints at another announcement. The post does not identify that other announcement. Guessing a product name would add speculation to a story that already has plenty of it. Read Mark's post.

When we checked the official model documentation on September 7, it still listed Grok 4.6 as the latest. We did not find a Grok 4.7 launch announcement in the company’s news index. These pages are the places to recheck before treating a forecast as a release.

Model documentation · [Company announcements]

Model documentation · [Company announcements]

Model documentation · [Company announcements]

What has been said about training

On August 12, Musk said initial training for Grok 4.7 was complete and that supplemental training would add SpaceX company data. He also described an improvement over 4.6. Those are attributed company claims. They are not independent measurements of how the model will perform on your documents, emails, or research. Read the original statement.

The practical question is what the change produces. Does the model keep the important details in a summary? Does it follow a correction? Can it finish a task without quietly skipping a step? A description of the training data cannot answer those questions by itself.

What we still need to see

We did not find confirmed 4.7 pricing, subscription access, usage limits, or independent test results in the sources reviewed for this article. That is the limit of this check, rather than a claim that no information could exist elsewhere.

Keep version numbers attached to the numbers you read. Grok 4.6 has its own documentation. Its prices and limits should not be copied into a 4.7 buying decision. Check the model list.

Keep version numbers attached to the numbers you read. Grok 4.6 has its own documentation. Its prices and limits should not be copied into a 4.7 buying decision. Check the model list.

It also helps to separate the model from the products around it. Grok Bot is a product for agents working with tools and apps. The September 3 enterprise announcement describes access controls and administration for that product. It does not establish the capabilities of an unreleased 4.7 model. Grok Bot for Enterprise.

It also helps to separate the model from the products around it. Grok Bot is a product for agents working with tools and apps. The September 3 enterprise announcement describes access controls and administration for that product. It does not establish the capabilities of an unreleased 4.7 model. Grok Bot for Enterprise.

A simple test you can reuse

You do not need a public leaderboard to compare two tools for your own work. Choose a short, non-sensitive document you already understand. Give both tools the same instructions and keep the relevant settings as similar as possible.

First, ask for a summary with three required facts. Check that all three survive. A fluent paragraph can still leave out the one detail you needed.

Next, correct one detail and ask for a revision. Compare the versions. Did the tool make the requested change while preserving everything else? Count any new errors it introduced.

Then remove a piece of information needed for the task. Does the tool identify the gap and ask a useful question, or invent an answer?

For a task with several steps, interrupt it and ask for a progress report. You should be able to tell what is complete, what failed, and what needs your input.

Keep a small record: corrections needed, facts lost, unfinished steps, and time spent checking. Repeat the task before drawing a conclusion from one result. If you are comparing paid access, include what it actually costs you to complete the task.

Our view at Coursiv is that fewer corrections would be a meaningful upgrade for a beginner. Grok 4.7 still has to demonstrate that. Until it does, the useful question is specific: which part of your work would you want it to handle better?