Grok 4.5: preu, rendiment i què confirma SpaceXAI

Grop 4.5 is the next frontier model released by Space XAI. But unlike previous Grop 4 variants, the new Grop 4.5 model is actually based on a newly pre-trained V9 foundation model that is 1.5 trillion parameters in size, nearly three times bigger than the previous V8 base model. Now the reason why this is worth mentioning is because when we have a newly pre-trained model, it typically requires training the model from scratch with a fresh batch of data. And we know that SpaceX recently acquired Kershurs parent company, Anyspher, back in June, and the amount of high quality coding data

that went towards the 9-based model opens up so much room for newer capabilities. And this is exactly what Elon tweeted on X work. Kershurs data was supplemental to the new V9 Foundation models training. Another fact worth noting is the size of the model being 1.5 trillion parameter model, which is comparable to open models like Kimey K2.6, which is a one-chryl-in-primator model

and deep-seek V4 Pro, which is 1.6-primator model. And despite that, GROP 4.5 pulls ahead in performance when you compare them against all open models, as you can see here. But just a bit shy of state-of-the-art models, like Fable 5 and GPD 5.6. But what's crazy to think about when looking at this index from artificial announces is that this entire section of models

were all announced only in the short span of the past 30 days. except for Opus 4.8. Fable 5 was initially released on June 9, GPD 5.6, announced on June 26, and now we have the Grock 4.5 model, all within the past 30 days, which once again shows a brutal competition among Frontier Labs trying to stay ahead. The last time we had a performance lead like this was back in September 2024, when the first reasoning model or one was released by OpenAI as you can see. And it took the rest of the industry nearly six months to catch up. First by Deepseek during the Deepseek moment back in January 2025, and ever since,

we had small increments here and there, and yet again, we have a new leap. But this time, started by Enthropate with their mythists and fable model, and the resident industry like OpenAI and Grock are now catching up really fast. I'm not sure why this chart doesn't have GBD5.6 sole added here, but for reference, the 5.6 sole should land right around here. So, we are undergoing yet another leap in model performance in how LLMs are progressing. Now, when we compare the Grock 4.5 models side by side against other state of the R models like Vable 5, Opus 4.8, GBD 5.6 sol, and hopefully Google's Gemini

models soon, I hope. The first thing you might have noticed is just how cheap the model is in comparison. Pricing in just at $2 per million in boutokins and $6 per million out boutokins. Now, the phrase that always gets tossed around when it comes to pricing is subsidized plans. Most people don't use credits to buy API costs, but opting to pay monthly subscription plan that subsidizes users to use their models. What's crazy about the GROC 4.5 model though is that GROC 4.5 API costs is comparable to a much lower tier models

from other labs like Sonic 5 from Anthropic, GPD 5.6 Lunar from OpenAI and Gemini 3.5 Flash from Google all lower tier models. They're essentially undercutting the entire close labs when it comes to the cost of intelligence. Now when we compare Grockville.5 against open models, they are actually one of the most expensive models available out there. Looking at models like Mini Max M3, KemiK2.6, Deepzig V4 Pro, GLM 5.2, and Q1.3.7 Max, all noticeably cheaper except for Q1.3.7 Max. So really, Grockville.5 is in this wedge not only in performance, but also in pricing, But there are priced better than front-to-labs in the US, but more expensive compared

to open models, while being more capable in intelligence than open models, but slightly below the front-to-labs top performance. So the next question here is usefulness. How useful is Grat4.5 model? And what are its use cases? But first, quick work from me exponseoring this video. Agents are what everyone is building today. But there's a huge technical gap when it comes to having your agent be more than just a chat bot.

to pull information from databases, make updates to spreadsheets, and update fields in the database. And we rely on MCP to handle all these things, but managing multiple MCPs and making sure they're all working together can get quite complicated. Make offers a super intuitive UI to build your own workflow that supports over 3,000 apps, as you can see, from this drop down list. From here, I can select my super-based database to connect to and also my Google sheets. And now, because these things are added as their own tool inside of make, my agent can vote them through a unified mcp through make on jobs that I give permissions to. As you can see, when I tell my agent to pull information from the database to collect all workup scores, it can do that. And I can also tell my agent to write to my Google Sheets all through

make. You can extend your agent through make for free using one month pro subscription that allows 10,000 operations link in the description below. When we read through the release note for for Grockvopin 5, they're signaling that this model is perfect for coding use cases for developing software and office work for knowledge tasks. You can see in their demonstration of how Grockvopin 5, one shot at the solar system app that simulates our solar system in the super interactive way where we have a planets POV like this as they float around in space. But this is just their demonstration, so we got to pull up our sleeves and test it out ourselves.

using their own harness called Grok build to start with. And I asked the model to put together a major data centers in the US in a visualized site. And I was surprised by just how fast the model actually got things done. It pulled together a lot of information about data centers but also built the site with code in just a few minutes. And the site looked something like this where it visually shows you a map

of all major data centers in the US and I can filter them by companies that are hosting them and go through each one by one. and I can also zoom into them as well. Pretty impressive for how quick it was able to get the job done. Now, to test their office task abilities, I pulled up a different harness since Graph 4.5 is also available on cursor. So I loaded up cursor and picked the model,

Graph 4.5 from the dropdown, and I asked it to convert this website we just created into a word file as a corporate briefing and also create a PowerPoint to do presentations on. And within just a minute or two again, you see that Graph was able to generate two extra files here, one for Word document and another for PowerPoint. The document written here shows a briefing outlining all data centers nicely organized and detailed. And the same thing for PowerPoints, very simple and clean as I prompted specifically for

Grok to follow. Now beyond this, you can keep trying out the model and for me, I was testing something really technical until I ran out of credits. But my honest feedback is that the model is noticeably fast in output. Looking at their 85 tokens per second SLA, it might not seem all that impressive, especially in comparison to OpenAI's recent GPD 5.6 announcement, running 750 tokens per second as an option, but I think the appeal to a model like this is using less tokens to get the job done, which is a different kind of speed. According to their report, they showed that Grog 4.5 solves the Swedish Probenchmark using less tokens where the model's average

output tokens was up to 4.2 times less than Opus. All in all, Grog 4.5 is showing the industry that XAI they're trying to wedge themselves in the market when it comes to pricing and intelligence, but also being a serious competitor in the enterprise adoption of AI models, given its smaller footprint and model size, they're integration with cursor and their large compute availability. And according to Elon, there will be increasing their Context window to 1 million context in just the next few days, which currently the model only offers up to 500,000 tokens.