NEW Grok 4.7 is HERE! 🤯

Julian Goldie (0:00)

Today, we're going to be testing out Grok 4.7. So Grok 4.7 just dropped from SpaceX AI, and we're going to see how it performs so far. I've actually built over 20 different things with Grok. If you're wondering how it performs, you can see the benchmarks here. So they've actually compared in the benchmarks versus GPT 5.6 Sol and Fable 5.1.

So on some of the benchmarks, it's actually performing pretty well. For example, at Software Engineering, it scores 46.3% versus 41.7%.

The input tokens are a lot cheaper than Fable and GPT 5.6 Sol. And the output token as well is a lot cheaper than GPT 5.6 Sol and Fable 5.1 Max. So it's cheaper and it's right up there on the benchmarks. Interestingly, this hasn't been compared against Astra 6 So I don't think it's anywhere near the same level as Astra 6 But at the same time, it's a lot cheaper and it has been compared against Fable 5.1 Max. And it's outperforming it on some of the benchmarks, for example, like software engineering. As you can see, it scores 71 versus Fable 5.1 Max is 70%. So how does it perform in reality?

Let's take a look right here. So we've got the tests on Goldie Bench. And we've been testing this inside our Agent OS system as well. So if we go to the Grok build section here, you can see that we can flick between Grok 4.7, 4.6, 4.5, right? And inside the workspace, we've actually got all of our builds with Grok 4.7 ready to preview, which is pretty nice too.

So here's an example of like a 3D world game that we created called Twilight Veil. It's not bad, it's not like the same level or anywhere near the same level with something like Astra 6, I would say. But the gameplay is smooth and it's quite easy to build stuff, as you can see right here. Let's see what else we got. RPG style game here. This is cool. One thing I've noticed in pretty much all of the coding tests is like it lacks detail, if that makes sense. So it doesn't have that same sort of viable feel as something built with Fable 5.1, but it's pretty fast and it's a lot cheaper. So it's and you know, I don't think it's like the perfect model. It's not going to replace any models I've created yet, but it's pretty easy to use. And also you can use it with CLI. So if you're already subscribed to Twitter, for example, you can use the CLI with this as well, which is pretty nice. This is a flight simulator, as you can see, which actually perform like decently well with.

It's not bad.

And also you can see the, we can just use it straight inside here. So if we go to the Agent OS, we've already plugged it in, and we're ready to go on that. Now, also, I would expect Grok 4.8 to come out pretty soon. It seems like these models are getting released, I would say every couple of weeks or every month with Grok, like it's moving really fast now. And the quality of this stuff is getting better and better, as you've seen on the benchmarks already. This is another one called Neon City, as you can see right here. By the way, all of this stuff is on Goldie Bench as well. So if you want to check out the stuff that we built with it, how it compares to other tools, etc. Check out Goldie Bench. You can also see what we've built versus other models. So for example, you want to see how Sakana Fugu performs versus Grok 4.7. You can see them side by side right here, which is super useful. You can also see it's designed more for work as well, so it's better at creating documents and presentations. In GDPVAL, an AI briefcase, AI is asked to work on tasks done by professionals, and Grok 4.7 improves on Grok 4.6, as you can see on the benchmarks right here. So it's just below FAPEL 5.1 with GPT-6 Astra, way behind on the benchmarks right there. And you can actually use this inside Grok build as well as the harness. So they have their Grok build harness, which is super useful too.

But yeah, kind of fun. You can see how it compares versus Grok 4.7 and 4.6 here as well. So similar sort of outputs on that test. Let's have a look at the 3D breakout game. You can see here like FAPEL 5 versus Grok 4.7 versus 4.6. Grok 4.7's output versus 4.6 is massively better. And it's a big step up versus the previous ones. 4.6 was already good, but if we look at, for example, Cursor Bench 4.0, the test, 4.6 scores 40.4% versus Cursor Bench 4 for 4.7, which is 46.3%. And for Deep SWE V11, Grok 4.7 scores 71%, whereas for 4.6, it scored 65%. So it's jumped up massively from the previous models here as well. The same for Artificial Analysis Intelligence Index. Geeks 6 is crushing Grok 4.7. The same with Fable 5.1, right? So for Artificial Intelligence Index, 4.7 is nowhere near. And that's what I mean. I wouldn't replace my main frontier models with this. I'm not going to replace Astra 6 with 4.7 anytime soon, but it is good and it is useful to use. Some cool things I'll show you that you can actually use for this. So for example, inside Hermes Agent here, we have Hermes Oracle, and Hermes Oracle is basically pulling in the most interested news from X using Grok with Hermes Agent. So that runs on a schedule, and that's the only API that's going to give you access to this stuff. So if you have the model, you have the CLI or the API for Grok 4.7, it can be super useful in this way, particularly for getting new news, if that makes sense.