Gemini 4 Carbon is Live!

0:00 Okay, so over the past couple of weeks, I've had early access to Gemini 4 Carbon, and by the time you're listening to this, it will be available and you'll be able to try the model out yourself.

0:10 In that time, when the model was announced, it was number one on the artificial analysis intelligence index. But since then, over the past couple of days, Anthropic has put out Fable 5.2

0:21 and actually taken that spot back for themselves. So, the model's no longer the smartest model on the chart. But honestly, that's not really what I wanted to talk about here. Anyway, there

0:31 are a number of things about this model and Fable 5.2 that take things to a whole new level for us as developers here. Okay, so what is Carbon? So, this is Google Deep Mind's new frontier

0:43 model. Google's calling this their strongest model yet. And obviously, that's comparing to Argon, the other Gemini 4 model that's been released. Now, I made a couple of videos about

0:53 Argon already. And it's interesting that Google is actually saying that carbon and argon are actually the same size. Carbon is newer and better, but Argon is the faster one. So, there's probably a

1:06 number of things that we could sort of deduce about what's actually going on under the hood with these two models. Over the past week or so, I've talked about some of the techniques that people

1:14 are using to make these models smarter, and my guess is that we're probably seeing some of those at play here. So a lot of the big improvements here with carbon seems to be that it does better

1:25 thinking and does more thinking on each task. So we can look at that a little bit more as I go through the demos here. But before I get into that, one of the things I want to ask from you is that my

1:35 editors pointed out that the majority of people watching our videos are not subscribed. And nowadays with YouTube's algorithm, it really is basically a game of luck if you're going to see the next

1:44 videos we make if you're not already subscribed to the channel. So, as much as I hate to do the YouTuber thing, if you find the videos useful, please click like and subscribe so the videos

1:53 actually get in front of a lot more people and that hopefully you see them as they come out. The other thing to point out here is that Google is saying that this is not the end of the lineup

2:02 either, right? Gemini 4 has now got two models there, but they're being fairly open that there's a flash size German Gemma Model 4 coming. Keep an eye out for that because I don't think we're

2:13 going to be waiting a very long time for that to come. So what's carbon actually built for? Well, Google lists four areas. The first is multi-day agentic engineering. Basically just longunning

2:25 agentic tasks here that includes things like deeper reasoning, surgical edits to code bases, verification loops, and especially a lot of uses around sort of cyber defense. This is definitely one of

2:38 the models that I would use for checking my codebase for any security issues before going into production, that kind of thing. The second thing up is definitely aimed at Google's probably

2:48 breadandbut customers and this is high precision knowledge work. So this is aimed at being better at things like finance, legal, healthcare, anything that requires a lot of constraint

3:00 checking, exhaustive retrieval and Google's claiming that they have zero fabricated citations here. So that whole sort of element of companies using this with their own data to extract insights

3:12 and useful information out in a way that they can use it and it can become really valuable. Third up is frontier discovery tasks. So this is everything like physics, maths, quantum computing,

3:23 genomics, drug discovery, all those kinds of things. And my guess is for most of us, this is just not what we're going to be focused on using this model for. But I'm sure just like some of the

3:33 other frontier models, we're going to see people using these kind of models to make breakthroughs in those fields. Fourth up is probably the one that I think is most interesting for us as

3:43 software developers. And this is sort of full stack front-end and 3D web development. So, a whole bunch of things related to realistic physics, interactive game mechanics, UIs that the

3:54 model can check itself. And I got to say this is probably the area that I'm most interested in. And I'll show you these in the demos as I go through them. I do think this model has reached a new level

4:04 compared to any of the previous Gemini models in relation to software development here. Now, if we look at the benchmarks actually for this, we can see that on terminal bench 4.0, which is

4:14 aentic coding in a terminal, this is getting 65.2%. That's against about 59 for both Opus 5.5 and GPT6 Astra. On automation bench for knowledge work, it's around 54

4:27 versus the low 40s. And on quite a few of the other benchmarks, it's surpassing Opus 5.5 or similar OpenAI models etc. Now it is important to point out that most of these benchmarks though were put

4:38 together before Fable 5.2 came out. Right? In fact, if we look at the independent side of benchmarks, this model when it was announced actually was number one on the artificial analysis

4:52 intelligence index at 60, a point ahead of Opus 5.5. But since then, Fable 5.2 has come out and it's actually at 65. So, not only is Carbon second, it's second by five points here, which I

5:07 would stress is actually quite a lot. But here's an important element to take into account. Fable 5.2 costs around $26.50 per task on that index. Carbon is a little bit lower than $780 there. So,

5:20 for five points less, you're paying roughly a third per task. And I kind of feel for most of the things that people want to actually run with this every day, that is probably going to be a more

5:32 important number than which of the models is the most intelligent. Now, of course, that may not be the case for scientific discovery where you do want the absolute best of the best for those

5:42 kinds of models. One thing to point out with carbon is that artificial analysis has it at around 26 minutes per task where the median on that benchmark is around 13 minutes. It also generated out

5:54 around 205 million tokens running against their index where the median is around 72 million. So it certainly thinks a lot and you're going to feel that if you're sitting there waiting for

6:04 this. For me in my testing, I generally ran it and then just left it and came back to it most of the time. I don't see it as being the model that I'm going to be using for chat and stuff like that.

6:15 It's more the model that I'd sort of launch on a task and then just leave it. All right. So, if we look at the pricing here, the standard price for this is $4 in and $20 out per million tokens.

6:26 That's exactly the same price as Opus 5.5, but currently there's an introductory price of $2 in and $10 out. The exact same price that Argon launched with. That's running through November

6:38 30th there. All your cash input here is 95% off. And in some ways, I kind of feel you could start thinking about these models as being more sort of agents that you call than just an LLM

6:51 model. With Google's API, you no longer have control over things like temperature, top P, top K. They're all preset for you. You can basically determine a reasoning effort level

7:02 there. I would say regarding to the price, and this is me just guessing here, is that I wouldn't be surprised that we either see a newer model that gets introduced with a similar kind of

7:12 pricing or we see that pricing get extended in some way, shape, or form. But I guess time will tell on that. So, we just have to wait and see. All right, let's jump into the demos and have a

7:21 look at what I've actually done here. So, I've actually been throwing a lot of different things at it over the past couple of weeks. I wanted to start out with some 3D things. This is one of the

7:30 areas that Google's making bold claims about this and certainly I've been pretty impressed by what the model can do in regards to this kind of thing. So what I asked it to build here was a

7:43 detailed 3D interactive rendering of marina based sands in Singapore. And I deliberately chose this one because it's probably a lot harder than things like the Eiffel Tower or the common tasks

7:54 that people pick for this kind of thing. and two because I think there's less data online for the model to have learned from for this thing. So, I got to say that I was pretty impressed

8:05 watching this be built. So, this was mostly oneshotted. I did add in two or three prompts afterwards, but it was mostly things like improve the shader, adding clouds and people, and that was

8:17 pretty much it. Not only has it got all the details of Marina Bay Sense, it's also been able to get a lot of the details of the buildings surrounding it, right down to things like the sphere for

8:27 the Apple store, the pyramid for the Louis Vuitton store, and even some of the other buildings here do actually seem to be representative of real buildings and not just made up here. So,

8:39 that's one thing that sort of impressed me. Now, like I mentioned before, for this kind of task, you tend to just kick it off and leave it. And it will take a fair bit of time to actually do this. In

8:49 this case, I didn't ask it to use anything like Blender, though I am sure that the model has definitely been rled a lot for tools like Blender for a lot of these game development things, etc.

9:00 The big thing I was really curious watching this is whether it checks its own work. So, Google talked a lot about self-verified UIs, and I got to say that it does seem like this was checking in

9:11 and adding details over time as it built it out to make it more impressive with each iteration, etc. All right, so the second demo is one that I think is really important to us as developers.

9:23 And this is where I want to say that both this model and the Fable 5.2 2 model are suddenly now I feel in a new league when it comes to software development. This was actually done in

9:35 anti-gravity. I asked it to create a pomodoro app. So if you know what a pomodoro app is, it's basically just a timer sort of app but I asked it to make it a sophisticated pomodoro app but

9:46 actually gave it the prompt build me the ultimate pomodoro native Mac app. I want it to have useful features and to look fantastic. And then I told it the folder I wanted it to put it in. If you look at

9:58 the UI here, it's been able to build not just the sort of standard sort of Pomodoro features in here. It's got a whole bunch of sophisticated features in here. It's got things like different

10:10 kinds of audio to keep you focused when you're in the Pomodoro sort of zone. It's got full tracking of your different projects and how many Pomodoros you've done for here. And the cool thing here

10:21 is that this has features that I didn't even ask for, like a menu bar controller, so that when you shrink the app, it actually just goes up to the menu bar and shows you the key

10:32 information there. Along with that, this has been fully developed in Swift. It's using the standard Mac OS Swift stack to actually build this. And one of the things that impressed me a lot here was

10:42 that it was able to actually do all the signing. Now, I do have Xcode installed. So, I'm not sure if it actually used that to basically do all the signing, but at the end, it was able to give me

10:51 both a full Xcode project, but also a fully compiled Mac app that I could actually run and see it working straight away. The second idea along the same sort of software idea here that I wanted

11:03 to try out was could it take a mashup of two apps and sort of create something creative here itself. So, I literally just open the app store on the Mac. look at what was coming up in the

11:16 productivity recommendations and saw the apps Obsidian and Bear. So Bear is a particular kind of note-taking app. Very nice, very welldesigned, very nicely put together. Obsidian, for those of you who

11:28 don't know, is basically sort of like a thinking linking kind of app where you can have a whole bunch of markdown files in a vault. You can link from one idea to another idea. It's very popular and

11:40 very useful for things like second brains, for mapping out different projects and for dealing with thinking that's just not purely linear and stuff like that in here. So when we look at

11:50 the app that it made here, we can see it's clearly got a lot of the features from both of these. Right now this is not doing any reverse engineering or anything like that. And I'm not actually

12:03 sure how much Google is going to let you use things like the REA, MCP, etc. to do reverse engineer anything. This is basically just using the model's knowledge and its ability to look up

12:15 things on the web to work out what to build for this particular mashup. And I deliberately gave it a super simple prompt. The prompt was in a new folder, research the apps Bear and Obsidian, and

12:28 build me an app that has the best features of both of those. I got to say, while I went off and did some meetings and stuff like that, it went off and did its building. Again, when it was done, I

12:38 had a fully native Mac app here. It had a menu bar component. It was all done in Swift. It was all done in a way that I would expect a Swift developer to actually develop this kind of app. So,

12:49 playing with it, I'm sure there are some things missing. But I should stress that this wasn't sort of followups here. And the thing that I think is really big here is what's this going to do to

13:00 software development. Now, I certainly don't think that every end user is just going to start, you know, developing their own apps for everything. And I think people are going to run into

13:09 issues and stuff like that. But I do think if you're a developer and you're working at even a small medium business and you're paying for some kind of SAS that has a high cost per seat, I think

13:19 people are just going to make those things. I think the whole thing of small software houses has always traditionally been that their value is that they could build a really nice polished app and

13:29 that they understood the problem really well. The day and age where a model can basically research videos about that app, screenshots of that app, descriptions of that app, a lot of

13:39 people are going to just make their own. And I think a lot of people are just going to open source those versions and make them free for other users to be able to download and use. Now, I don't

13:49 think that that means totally that software business is going to disappear because I think there are definitely going to be people that will still pay for software. I do think with tools like

13:58 this and with the reverse engineer anything MCP, you are going to see people put a very dumb front-end interface and put all the intelligence or the complicated things in the cloud

14:09 basically to try and prevent people from cloning your software. And I'd love to hear from people in the comments what you think about this and how are you actually seeing this change in your own

14:18 life here. I do think we're now at a point where a lot of the arguments that were against things like vibe coding in the past are just falling apart. You know, as these models get better at

14:29 being able to prepare things for production, as these models get better at being able to make these things secure, it really has to make us question how we're actually building

14:38 software if we're intending to actually license or sell that software. All right, so just in closing, I got to say that Carbon is a very strong model, especially on the things like I

14:48 highlighted about building software and stuff like that. It may not be number one on artificial analysis anymore. Clearly, Fable 5.2 2 has taken that mantle and also is the cutting edge for

14:59 pretty much everything that carbon can do here. But for carbon being roughly a third of the price, I do think that's probably going to matter a lot more for a lot of people. And it is going to be

15:08 interesting to see, do we see the days of these all you can eat buffet plans for tokens slowly get reduced over the next year to the point where the best of the best models actually can become so

15:22 cost prohibitive that people end up looking for other alternatives here. And in a world like that, carbon definitely becomes one of those alternatives. Anyway, so keep an eye out for the flash

15:32 model. I think we're going to see that in the near future. I put links to the launch post, etc. in the description. Let me know in the comments if you build something with this, how are you

15:41 actually finding it? Where are you actually finding weaknesses with this model? The more that I've used Argon, that one, I've definitely found weaknesses around tool calling and where

15:50 it just tends to just try and be too intelligent rather than actually go and use search, etc. As always, if you found the video useful, please click like and subscribe and I will talk to you in the

16:00 next video. Bye for now.