Grok 4.7: What's New in xAI's Smartest Grok Yet
By Chatday Editorial Team ·
Grok 4.7 is the newest and most capable Grok, released on September 21, 2026. It’s aimed at coding and long work tasks, it checks its own work more carefully than Grok 4.6, and it costs exactly the same to run. You can already chat with it in Chatday.
Wait, who makes Grok now?
If the name on the launch post confused you, you’re not alone. Grok is still made by the company you know as xAI, Elon Musk’s AI lab. SpaceX acquired xAI earlier in 2026, and in July the lab started calling itself SpaceXAI. The models are still called Grok, and on Chatday you’ll find them under xAI with the familiar Grok logo.
What matters for you is the model itself, and you can read its specs on the Grok 4.7 model page.
What Grok 4.7 is better at
SpaceXAI calls it its most capable model for coding and knowledge work. In plain terms, three things changed.
It works longer on hard problems. Grok 4.7 was trained with a longer round of practice on difficult tasks, weighted toward the kind that take many hours to finish. The goal is a model you can hand a big job and leave it to grind through, instead of one that gives up or wanders off after a few steps.
It checks itself. SpaceXAI says it double-checks its own work more carefully before answering. That should mean fewer confident mistakes on multi-step problems, like a bug fix that seems to work until you run it.
It’s built on a bigger brain. It’s based on a new, larger base model than Grok 4.6. Musk has claimed it has around 2.1 trillion parameters (the internal settings a model learns), but that figure isn’t in the official announcement, so treat it as his claim, not a confirmed spec.
You’ll also see it handle documents, slides and other office work better, according to the company.
Same price as Grok 4.6, with one catch
| Grok 4.6 | Grok 4.7 | |
|---|---|---|
| Price to send text in (per 1M tokens) | $2 | $2 |
| Price to get text out (per 1M tokens) | $6 | $6 |
| How much it can read at once | about 500K tokens | about 500K tokens |
| Built for | Long-running agents | Coding and long work tasks |
A token is a small chunk of a word. Those prices matter mostly to developers who pay per use, and they double for very long inputs (above 200,000 tokens).
The catch: reviewers noticed Grok 4.7 writes roughly twice as many words per task as Grok 4.6. Cheaper words don’t help if the model uses twice as many. For a quick everyday chat you won’t notice, but on a big automated job the bill can add up.
How good is it, really?
SpaceXAI’s own tests show a clear jump over Grok 4.6 on coding tasks, including a big improvement on a test of command-line work. It led the launch table on a couple of tests, including one on legal tasks.
The company didn’t claim to beat everyone. Musk himself said Grok 4.7 places SpaceXAI “third, after Anthropic & OpenAI, for agentic coding,” meaning AI that writes and runs code on its own over many steps. That’s an unusually modest line for a launch, and it’s a fair framing: a strong, good-value model, not the undisputed leader.
There’s also a timing wrinkle. Grok 4.7 arrived late, after Musk pushed the date back more than once, and he has already started talking up a Grok 4.8. If you’ve been waiting on Grok 4.7 to show up in Grok’s own app, reports a month ago said grok.com still ran Grok 4.6. In Chatday, it’s already there to pick.
Should you use Grok 4.7 or something else?
| If you want to… | Try |
|---|---|
| Fix code or build something step by step | Grok 4.7 |
| Work through a long document or project | Grok 4.7 |
| Get quick, casual answers | A lighter, faster model |
| See how it compares to Grok 4.6 | The side-by-side comparator |
If speed matters more than depth for what you do, our guide to fast vs smart AI models explains when a lighter model is the better pick. And if you’re just starting to explore beyond one assistant, see our tour of AI models beyond ChatGPT.
Try it on the thing you’d normally do yourself in an afternoon: a stubborn bug, a messy spreadsheet, a long report. Then ask it to check its own answer and see what it catches.