xAI are killing it (Grok 4.6, Image 2.0 and more...)

English
BBetter Stack
Computing/SoftwarePhotography/ArtInternet Technology

Transcript

00:00:00When Grok first came onto the scene, despite being pretty late to the party,
00:00:03they actually ended up competing with Frontier Labs like OpenAI, Google, and Anthropic.
00:00:08But that was all the way back in 2024, and over the following two years,
00:00:12the landscape completely changed.
00:00:14OpenAI and Anthropic became the only labs worth taking seriously,
00:00:18at least for software engineering, and Grok just faded into the background.
00:00:22But something has clearly changed because in the last few months,
00:00:24XAI have been shipping like madmen.
00:00:27We've got a brand new Grok 4.6.
00:00:30Seriously impressive video and image generation with Grok Imagine,
00:00:34a dedicated agent harness with Grok Build,
00:00:36and a voice agent builder for custom text-to-speech agents.
00:00:40Following the merger with SpaceX and the purchase of Cursor,
00:00:44clearly that has been a massive catalyst in moving them forward,
00:00:48and now they seem to have all the ingredients to start competing again on the main stage.
00:00:52So I thought this would be a good opportunity to look again at everything XAI have to offer,
00:00:57because if you're building software, I think they're a serious contender again.
00:01:05So let's start with the newest model release, Grok 4.6.
00:01:09You can see here on Artificial Analysis Index,
00:01:12this is like a composite analysis between multiple different benchmarks.
00:01:16So we've got things like Terminal Bench 2.1.
00:01:19We've got the GDP Val.
00:01:21And you can see that there's been a massive jump here all the way up to fourth place,
00:01:26now competing with GPT 5.6, and not far off Viable and Opus as well.
00:01:32And this is way more impressive considering the price at just $2 in and $6 out.
00:01:37We can come down here as well and look at the individual benchmarks.
00:01:40So if we look at GDP Val, we've come from like 12th place over here all the way up to second place.
00:01:45And this is evaluated on real world tasks, so not just simple tasks.
00:01:50This is now in second place, only falling behind Opus 5.
00:01:54And again, the price difference between these two models is huge.
00:01:58So this is incredibly impressive considering this is just a point increase from 4.5 to 4.6.
00:02:04If we look at Terminal Bench as well, again, we see a massive leap from 4.5 to 4.6,
00:02:10again, coming into second place.
00:02:11And then the most interesting chart, I think, is this one here,
00:02:14which is the intelligence index versus the cost.
00:02:18On the y-axis here, we have intelligence.
00:02:21And on the x-axis, we have a logarithmic scale for cost.
00:02:25And the perfect model would be something right up here,
00:02:27the most intelligent and the cheapest cost.
00:02:30So anything within this box would be classed as the most attractive quadrant.
00:02:34Now, you can see here, interestingly, 4.5 was inside this area.
00:02:38And now we've got a leap in intelligence, so an increase in intelligence here.
00:02:42Interestingly, the model has also become more expensive despite the same price on paper.
00:02:47So this must mean that to gain this new intelligence,
00:02:52the internal thinking steps on the model has changed.
00:02:55So potentially more tokens are being burned to get to certain results.
00:03:00So even though both 4.5 and 4.6 are $2 in and $6 out,
00:03:044.6 burns far more tokens to achieve a task.
00:03:09That's why we've got an increase in intelligence, but also an increase in cost as well.
00:03:13Now, this chart judges things based on the completion of a task.
00:03:16So it gives it a complex task and makes sure it finishes it end-to-end.
00:03:20So that's how it would detect things like models burning more tokens to achieve a certain task.
00:03:25So just something to keep in mind, even though it's the same cost on paper,
00:03:28you'll likely end up spending more money using 4.6 over 4.5.
00:03:33And if you're enjoying this one, guys, we cover the latest in AI constantly on this channel.
00:03:37So you'd do me a massive favor if you could subscribe.
00:03:40So next up, let's look at the Grok Imagine API.
00:03:43This covers multiple things, image narration and video generation.
00:03:47And video is actually an interesting one because if you remember Sora from OpenAI,
00:03:52that was a product that was recently closed down because it was too expensive to run
00:03:56and too distracting because they basically were running too many products at once
00:04:00and they wanted to focus, I believe, more on software engineering and all of the codex features.
00:04:06However, Grok have decided to now take over video.
00:04:09And you can see on their comparison chart,
00:04:10they also compare themselves to Sora despite it being a closed product.
00:04:15And according to their own benchmarks, it's cheaper on price, just 7 cents per second.
00:04:20And it's also much faster at just 45 seconds compared to Sora 2 Pro coming in at 2, 3, 6 seconds.
00:04:27So a massive improvement there.
00:04:29So the first thing I wanted to try out was to ask three different models
00:04:32to generate a flyer for a local food festival.
00:04:35I asked OpenAI, Gemini and Grok.
00:04:38So we look at ChatGBT's output.
00:04:40It's pretty decent.
00:04:41This is actually usable, I would say.
00:04:43All the text is completely legible.
00:04:45Looks really nice.
00:04:46If we look at Gemini's, pretty bad.
00:04:48It's gone with this kind of tune-ish effect.
00:04:51This looks super AI generated.
00:04:53So it's a little bit of a weird one.
00:04:54And then let's look at what Grok did.
00:04:56Grok's looks really good.
00:04:58It also looks very similar to the style that OpenAI or ChatGBT went with.
00:05:02All of the text is completely legible.
00:05:04We have clear icons, images of real looking food.
00:05:07It doesn't look AI generated at all.
00:05:09Incorporated a lot of the elements I mentioned.
00:05:11I mentioned it was local food.
00:05:13I mentioned it was organic.
00:05:15I mentioned there'll be local vendors as well.
00:05:17So support local vendors here.
00:05:18So overall, I think this output is pretty impressive.
00:05:21But the more impressive thing is if we look back over at the Grok image UI.
00:05:25So here again, we've got the same image.
00:05:27But interestingly, we have this thing at the side.
00:05:29Once you've generated an image called Segments.
00:05:31Each individual segment of this image is listed on the right here.
00:05:35All of the leaves down here.
00:05:37All of the individual icons here.
00:05:39This bowl of hummus.
00:05:40The brushstroke here.
00:05:42So all of these elements individually can be changed.
00:05:45So if, for example, I was to select this leaf here.
00:05:48I can then put in a separate prompt just to change this one element.
00:05:51So I could say, make this flower instead.
00:05:54So you can go in and just change individual elements by clicking them on the right here.
00:05:59And then describing your edits.
00:06:00Just clicking on the elements themselves within the image.
00:06:03So this makes customization to the generated images really, really flexible.
00:06:07It just makes the whole feature super powerful in my opinion.
00:06:09Next day, advertise a feature called Virtual Try-On where you can place your own products within images.
00:06:15So we look at this reference here, multi-image editing.
00:06:17You can see with the main example, we've got a woman here and then an outfit here.
00:06:22And this would produce this result.
00:06:24Give this a go.
00:06:25Use template.
00:06:26See if it works.
00:06:27Okay, nice.
00:06:27So that looks pretty good.
00:06:28You can see the outfit mentioned here is now on the lady.
00:06:32Okay, that looks pretty good.
00:06:33But I'd like to try this using my own images to see if it works just as well.
00:06:37So let's swap out this.
00:06:39We'll add another image here.
00:06:41Got one here of me earlier.
00:06:43Use this template.
00:06:45Okay, let's see how this looks.
00:06:46Nice!
00:06:48There you go.
00:06:49I don't know why I'm suddenly stood on the sofa, but you can see it's put me in the actual outfit, so it looks pretty good.
00:06:55Now let's try an even harder version.
00:06:58Here's me doing a marathon earlier this year.
00:07:00Let's see if that works.
00:07:02Okay, pretty solid.
00:07:06It's actually really good.
00:07:07I'm a big fan.
00:07:08They advertise a bunch of these features on the main homepage.
00:07:11You can also turn these into videos as well.
00:07:13So you can see here you can turn both the product and then the person into a moving video, and then they can do things like product placement here as well.
00:07:21So you can see it's a single video that you can then change the outfit, and you get a very consistent look between the movements and the character, just changing the one element that you want to be different.
00:07:33We can also do things like turn sketches into 3D models, into more realistic prototypes.
00:07:38And all of this is available through their API as well.
00:07:41So if you want to build services on top of this, that is now completely possible.
00:07:45Now let's look at the harness.
00:07:46So this one's a little bit strange.
00:07:48They've got the CLI, Grok Build, which is the same as like Claude or Codex CLI, purely terminal-based.
00:07:54Then they have Grok Bots, which is closer to Claude Cowork, where bots can drive your machine and browse the web.
00:07:59Then, of course, they now have Cursor after the acquisition, and Grok 4.6 is the model being pushed inside Cursor.
00:08:06So you can see on the Cursor homepage under the Models tab, Grok is now the only model that's mentioned.
00:08:12Clearly, see, they're trying to push this after the acquisition.
00:08:15You'll still have access to other models, including Cursor's Composer model.
00:08:19So listed here, you can see all of the models that are available.
00:08:22We've got Composer 2.5, Grok 4.6.
00:08:25You can still access Claude's models through Cursor as well.
00:08:29And we also have Cursor Router here.
00:08:31This is when you select Auto Mode, and then Cursor will automatically push requests to a different model,
00:08:36depending on the context and depending on the intelligence it believes you need.
00:08:40But between Build and Bots and Cursor, these products are kind of fragmented at the moment.
00:08:45And I think we'll start to see these converge since the acquisition.
00:08:49Potentially, Grok Bot, Grok Build, and Cursor will start to blend into one product.
00:08:53But as of right now, it's still fairly fragmented.
00:08:56I have been using Cursor this week, though, with Grok 4.6.
00:08:58And I have to say, it is genuinely good at software engineering.
00:09:01My usual driver is Opus or Fable.
00:09:04But outside of Anthropik's own harness, you're stuck with API rates.
00:09:08And that can get pretty expensive fast.
00:09:10Cursor is very capable and considerably cheaper with Grok 4.6 in comparison.
00:09:16Okay, next up is Grok Voice or Grok Voice Agent Builder.
00:09:19I'm not sure exactly what they're calling this entire suite of tools.
00:09:22Maybe just Grok Voice.
00:09:23But you can build your own agents, connecting it to tools like Web Search, Gmail, and custom MCPs to get the data it needs.
00:09:30You can select a voice and then even hook it up to a phone number to call it directly.
00:09:34So here I've created a personal voice agent called Personal Assistant.
00:09:38And I've gave it the capability to search the web.
00:09:40And then I can just try it live as well.
00:09:42So let's start speaking.
00:09:44Hello, I would like to go out for dinner tonight to have Italian food somewhere in East London because the weather seems quite nice.
00:09:52Could you recommend a restaurant with outdoor seating?
00:09:54I'd be happy to help you find a great spot for Italian tonight.
00:09:58Emilia's Crafted Pasta at St. Catherine Dock's stands out.
00:10:02Nice.
00:10:02So you can see that's pretty fast to reply.
00:10:05You can also interrupt it as well.
00:10:06And it seems like a much more natural interaction.
00:10:09The whole intent of this is not necessarily for me to just talk to personally online.
00:10:15But it's to create things like agents that can handle customer support or reservations.
00:10:19And you can actually connect it directly to a phone number.
00:10:22So you can set up a phone number here.
00:10:24Unfortunately, this is not available within my region.
00:10:26So I couldn't try it out.
00:10:27But once it's hooked up, it means that then customers can just call that phone number.
00:10:32It can interact with your API, use your tools, connect to your MCP.
00:10:36So it's got all the data it needs and then handle a lot of stuff for you.
00:10:40So again, if you want to develop on top of their API and support custom voice agents, that is now possible as well.
00:10:46They also do things like text-to-speech, speech-to-text.
00:10:49You can see all of that is available within the website itself and again, available through the API.
00:10:54You can also train your own voice as well.
00:10:56So if you wanted an agent that had an English accent, for example, that is also possible.
00:11:01Unfortunately, again, not available in my region, so I couldn't try it out.
00:11:04But if you wanted a bespoke voice that is more relevant to your brand rather than the default American accents, then that is also now possible.
00:11:12But Grok and XAI clearly have a lot more to offer now, particularly after the acquisition of Cursor.
00:11:18I think this is going to be an exciting space to watch and we're going to see a lot more happening within the next six months.
00:11:23So if you're doing software development or want to build services on top of their API, I do think these are a really good one to keep an eye on.
00:11:30I've heard a lot of great things from other engineers as well using Cursor with Grok 4.6, so also worth checking out yourself.
00:11:37But hopefully you enjoyed this one, guys.
00:11:39Thank you so much for watching.
00:11:41And again, you do me a huge favor by subscribing to the channel.
00:11:43We cover tech and AI constantly on this channel.
00:11:46So if you want to stay up to date, then subscribe.
00:11:48Thank you so much and I'll see you in the next one.
00:11:51We'll see you in the next one.

Key Takeaway

xAI has restructured its AI ecosystem through Grok 4.6, Grok Imagine, and Cursor integration, offering high performance at lower operational costs.

Highlights

  • Grok 4.6 costs $2 in and $6 out while competing closely with GPT 5.6 on intelligence benchmarks.

  • Grok Imagine generates videos at 7 cents per second and renders them in 45 seconds.

  • Grok image editing breaks images down into individual selectable segments for precise prompt-based modifications.

  • Cursor integrates Grok 4.6 as a native model option following the xAI acquisition.

  • Grok Voice agents connect directly to web search, Gmail, and custom Model Context Protocol tools to handle phone interactions.

Timeline

Grok 4.6 Performance and Pricing

  • Grok 4.6 ranks fourth overall on the Artificial Analysis Index and second on GDP Val.
  • Pricing remains fixed at $2 per million input tokens and $6 per million output tokens.
  • Internal thinking steps consume more tokens during tasks, increasing actual usage costs despite flat paper pricing.

After a period of quiet development, xAI released Grok 4.6, dramatically improving its ranking across major software engineering and real-world task benchmarks. Composite analysis places it alongside top-tier frontier models while maintaining competitive API pricing. However, task completion evaluations indicate that the model burns significantly more tokens internally to achieve higher intelligence scores.

Grok Imagine Image and Video Generation

  • Grok Imagine video generation costs 7 cents per second with a 45-second render time.
  • Image outputs feature selectable segment editing for modifying individual elements using text prompts.
  • Virtual try-on capabilities map target outfits onto user-uploaded images and generate consistent character video loops.

The Grok Imagine API expands xAI into multimodal generation, offering competitive video creation speeds compared to defunct models like Sora. When generating complex marketing assets like festival flyers, the model produces legible text and realistic layouts. Advanced features allow users to isolate specific objects within an image, swap clothing using multi-image editing templates, and convert flat sketches into 3D prototypes.

Development Harnesses and Cursor Integration

  • Cursor features Grok 4.6 as its primary default model following the xAI acquisition.
  • Grok Build provides terminal-based CLI tools similar to Codex and Claude CLI.
  • Grok Bots operate automated workflows across local machines and web browsers.

The acquisition of Cursor positioned Grok 4.6 directly inside popular developer tooling, replacing or supplementing existing routing options. While xAI maintains separate development interfaces like Grok Build and Grok Bots, these tools currently exist in a fragmented state. Engineers utilizing Cursor with Grok 4.6 experience capable software engineering performance at lower costs than native Anthropic rate limits.

Grok Voice Agents and Custom Tools

  • Grok Voice allows users to build conversational audio agents connected to external APIs and web search.
  • Agents integrate with phone numbers to handle customer support and automated reservations directly.
  • Bespoke voice training enables custom regional accents and brand-aligned audio profiles.

The voice agent builder lets developers link custom text-to-speech models with external data sources like Gmail and Model Context Protocol servers. Live demonstrations show rapid response times and the ability to handle interruptions naturally. Regional limitations currently restrict direct phone number deployment and custom voice training for certain users, but the underlying API infrastructure supports full production deployment.

Community Posts

No posts yet. Be the first to write about this video!

Write about this video