AI News: 6 New Models & A Nano Banana Killer

It's been another really busy week in the world of AI. We got some new image models, we got a handful of new large language models. There's a lot to cover, so let's get straight into it. I'm actually going to start in a place where I don't normally start, which is a new image model. And I wanted to start here because this is probably like the coolest, most fun thing I've played with this week, and that's Krea 3. This week they opened up their beta for it, and this is really like an image editing model. It can generate images and it kind of does it just as well as like any other image generator model. But where it really stands out is its editing. And let me show you why. I'm going to jump into Krea and I'm going to go ahead and upload an image. For instance, I'll use this image, which was, uh, one of my thumbnail tests from a week ago. And now that it's in here, if I click on the image, it'll open it in a large screen. And this is where it gets really cool. Is I can just edit things inline like this. Like I'll select this phone for instance and say, make the phone background blue. And then let's select my eyes and say, make the eyes red and bloodshot. Now I can click done and submit this and it actually works really fast. So it took, you know, 10 or so seconds and we've got 3 variations. This one probably looks the closest to what I'm going for. And again, if I just click on this, I've got my red bloodshot eyes and a blue background on the phone, and it didn't touch anything else about the image. Now let's go ahead and select the text on the phone like so, and then say, make this text say Gemini for Argon. And then I can select the little logo here and say, make this the Google logo. Submit those 2 changes. And a few seconds later, we've got 2 new variations that are both pretty much the exact same, but I could continue this on and on. Let's select this text here, change the text to new model. Let's select the logo on my hat, change this logo to a baseball logo. I don't know what I'm doing. All right, let's submit this. And you kind of get the idea. It is one of the best image editing experiences I've had so far. And to me, it does actually quite a bit cleaner of a job with these edits than something like ChatGPT Image or Nano Banana. Now, I imagine it's only a matter of time before those catch up and it feels like a similar workflow, but this is just such a simple, easy experience. Like, here's our new image. I got a baseball on my hat. You can see it looks like it's stitched onto a hat. The text changed to new model, keeping the exact same font and color scheme. Like, If you need to make quick edits on a thumbnail, this is going to be like the new tool to go do that. So I was really, really impressed with Krea 3, which is why I had it lead off this video and I wanted to show it off first. I think it's really cool. It is in a public beta right now. You can use it over at krea.ai. We actually got a total of 3 new image models this week. We got a new model out of Black Forest Labs called Flux 3 Image Klein. And this is actually an open weight image model. I believe the Krea one is actually going to be open eventually. Actually as well, but this one was released as open. You could click on download the weights and grab them on Hugging Face right now. And if we scroll down, we can actually see what kind of computers this will work on. And check this out, if you have a DGX Spark, you can do text-to-image in 7.3 seconds. You could do image editing on a DGX Spark in 16.2 seconds. But I mean, DGX Sparks are like $4,000 machines. However, RTX 4070s are much easier to find, and if you have one of these, well, it's gonna take like a second and a half longer, but you can do text-to-image, image editing, and even higher resolution text-to-image and image editing all on a consumer-grade RTX 4070. Now, when it comes to how good it generates images, you know, it's decent. I almost feel like image generation is fairly solved at this point, so I don't talk about it quite as much as I used to. But this one is open weights, and there is a way to use it for free. If I jump back over to Hugging Face here. I'm on the Flux 3 Image Klein page. Over on the right, you can see Spaces using Black Forest Labs. If I click on this top one here, Flux Image Klein, you can actually play around with this image generator for free on Spaces. Now, it's not going to be ultra fast because when you're using Hugging Face, it's sort of like shared cloud compute. So a lot of times, if there's a lot of people playing with it at the same time, it's going to be a bit slower, but you can actually use it here for For free right now, here's some of the pre-cached images that they've already generated: this still water record cover, this Lido Lido poster, gold leaf bakery sign, Dutch still life, brutalist church. I mean, again, I don't feel like there's anything particularly special about this image generator other than the fact that it is open source and that you can run it on consumer-grade computers right now. We also got another open weight model, this time out of Alibaba in their Quinn Image 2.1 Turbo model. This model is also designed to be really fast. They already released QuinImage 2.1. This Turbo model is an accelerated checkpoint, so you probably lose a little bit of quality in exchange for a little more speed. And like the other one, it is available to test over on Hugging Face. Head over to Hugging Face. Down in the bottom right, there's this Hugging Apps QuinImage 2.1 Turbo. You can select that and you can see here's an example of like a chemistry poster that it generated. This is a transparent anime bride. I don't like the bride's transparent. I think the background is transparent. A travel planner user interface. Again, it's hard to impress me with image models because I do feel like these image models have been fairly solved for a while now. I get more impressed by image editing than image generation, which this one can also do image editing. But for me, Krea 3 is the big winner of the week when it comes to image. This next update is from Genspark, and they just launched a new AI design tool called Genspark Design. I'm going to see if I can turn my Future Tools site into a fully mapped out mobile app. So here I can drop in some screenshots of Future Tools for design inspiration, and then I'm going to ask it to keep the overall style similar but rebuild the experience so it feels mobile first. And you can see it didn't just create one generic mockup, it designed the full app with a home feed and individual tool pages and search and filters and a news section. And everything it generates is layered and editable. So I can click into any screen, rewrite the text, change the colors, adjust the spacing, or completely restructure the layout without starting over. What's great about designing all of this in GenSpark instead of in something like Figma in the same workspace, I can also create the images, launch video, website, and other assets I'd need for the app and collaborate on it with my whole team. And when the design is ready, I can export its components and design tokens or bring it into GenSpark Code to start turning it into a working product. If you want to try it for yourself, you can get free credits when you sign up for GenSpark Design using the link in the description box. And thanks so much to GenSpark for supporting my channel and sponsoring this portion of today's video. We also did get a ton of new large language models this week. Probably the biggest model and the most impressive one was this Gemini 4 Argon model. I'm not going to go too deep into this model on today's video because I did already make an entire breakdown here in this video titled Google Just Caught Up: Gemini 4 Argon. I will I will, however, give you like a little bit of a TL;DR. Didn't watch. This model was actually announced a few weeks ago, back on September 30th. So all of the benchmarks and everything we know about it was already sort of put out into the public. But for the most part, it's caught up in almost all the benchmarks with GPT-6 Astra, Claude Opus 5.5. Like, it is kind of state of the art in most benchmarks. As far as availability, you can use it via the API and you can use it inside of Gemini if you're on an AI Ultra or Pro plan. So one of their paid plans, the API cost $2 per million input tokens, $10 per million output tokens. But that is for a limited time. That's through the end of November. Then that price pretty much doubles. Taking a look at artificial analysis, which is sort of this combination of a whole bunch of benchmarks that are all sort of mashed together to try to find like the objectively smartest model, actually puts Gemini 4 down here with a score of 53. A full 5 points below Opus 5.5. However, as far as cost per task goes, Opus 5.5 is way up here at the 3rd most expensive cost per task model, where Gemini 4 is way down here. So a lot cheaper per task of a model to use with a very sort of similar level of intelligence. I of course ran it on BUCI Bench. This was the BUCI that it generated, and it did it for about $0.34, used 34,000 tokens, and took about 4 minutes to generate. It did, however, land On our leaderboard where an LLM judges these images at number 2. I also ran it through my normal set of tests and had Gemini 4 generate its version of Mega Bonk, and this is what we got out of it. It's not bad. Not the best game generation we've ever seen, not the prettiest we've ever seen, but you know, it's decent. Can't really complain. Uh, this is one shot. I've got my camera angles that I can move around. I could jump, I could sprint. It's just none of the characters or graphics look that pretty. We've definitely gotten better looking graphics out of some of our other models. Now, one thing that was interesting was I was on this like pre-brief call with the Google team like a couple of days before this launched and they were sharing some details. And one of the details they shared was that it's really good at creating 3D print models. So I actually gave it a prompt and said, generate a 3D bust of Gary Busey. And well, wouldn't you know, I got my Gary Busey bust right here. It did it pretty good. Now, if you look at the profile, it's not amazing, but straight on, that's a pretty good Busey bust. And again, that was me just asking Gemini 4 Argon to generate an STL, and I gave it an image of the front of Gary Busey's face and said, generate a 3D bust that I can print on my printer of this person. And yeah, I mean, it's pretty Busey. All right, let's move on. If you want a much deeper dive into Gemini 4, check out that video I mentioned. You'll see all the prompts that I use and a lot more tests that I did with it. But again, don't want to go too deep in this video. Since, well, there's a whole other video about it already. We did get another big model release this week that's impressed a lot of people. I can't say I've been super impressed with it, if I'm being honest. It's just been kind of okay. But Theo here, who's typically pretty critical about most LLMs, claims that it's S-tier. He says, my current model rankings: S-tier, Opus 5.5 and Muse Wonder. A-tier, GPT-6.1 Soul. Gap, everything else falls below that. So, I mean, he put it as an S-tier model. But here's what you need to know about it. On DeepSuite, the benchmark most people look at to see how good these models are at coding. Muse Wonderful scored a 79%, which would make it the best top-tier model on DeepSuite. Here's a bunch more benchmarks. Superintelligence Index, Harvey Lab for legal work, financial analyst research. You can see it scored in the state of the art for all of those, and it scored on par with the other state-of-the-art models for all of these benchmarks here. And well, as we can see, just seems like a really, really good model. Like Meta has caught up according to the benchmarks with the state-of-the-art models. As far as pricing goes, $4 per million input tokens, $20 per million output tokens. This puts it at, I believe, roughly the same pricing as Opus. Yeah, so Claude Opus 5.5, it's about the same price as that. Looking back at artificial analysis here, we can see that Muse Wonder scores up here. It actually scores above Gemini 4 Argon, above Gemini, above GPT-6 Astra, above Claude Fable 5.1, which like a couple months ago was a model that was blowing everybody's minds and freaking people out. Well, this one scores above that, and it's from Meta, which I don't think people saw coming from Meta. It's tied with Claude Sonnet 5.5 and just 2 points below Opus 5.5. On a cost per task basis, it's kind of up there. It's the 5th most expensive model, but not quite as expensive as the other Anthropic models. I ran this one through BUCyBench as well, and well, this is the BUCy That it generated. It cost about 34 cents, used 17,000 tokens, and came in at just under 4 minutes to generate. And I don't think it fell anywhere in the top 10 on my leaderboard. No, it fell in at number 14 according to the LLM as a judge. I also gave this one the Mega Bonk test to see how it would generate. And well, this is what we got with that. Let's do Sir Bonk-a-Lot. And yeah, we got a snowman with a stick. And it's actually really hard because it only hits him like if the bat actually makes contact. But This is one single prompt and we got camera control so I can move around the camera. I can jump, I can sprint a little bit. It, it did what it was supposed to. It actually coded up a working game and it got all the game mechanics right. It's just graphically, it's not on par with some of the other models we've seen. I definitely think Argon did a better job so far. I think Opus 5.5 did the best job, followed by GPT-6 Astra. I should really make a website that shows all the benchmarks of these various Megabonk benches. So aesthetically, not the best looking of the Megabonks, but from a standpoint of actually coding up all the functionality, it seemed to do it all correctly. But we got even more models this week. In fact, we got 4 different OpenWeight models this week. We got another model out of Muse called Muse Spark Open. So previously Muse Spark was Meta's best model, but now that this new Wonder model came out and that's their new state of the art, you know, internal state of the art, they released their Muse Spark in an open version. And according to their website, it is the most capable US open weight model to date. A lot of specification in there, right? It's the most capable US open weight model because obviously there's a lot of other open weight models out of China that are possibly more capable, but this is the best one One thing that's interesting about this model specifically is that they actually worked with Apple to optimize it to run on Apple Mac Studios. So this will actually run on an Apple M5 Ultra Mac Studio, which is pretty impressive considering that it's an 816 billion parameter model with a 1 million token context window. Usually models that size, you're not even gonna be able to run on any sort of, well, consumer computer. I don't know if I'd really call an M5 Ultra a consumer computer, seeing as you're looking about $2,000. about 10 grand to build a system that'll run it. But hey, it'll run on a Mac Studio. When we look at the benchmarks here, you can see it's not comparing it with like the most state-of-the-art models. It's comparing it to the other open weight models like Nimotron, Mr. All Large 4, Quinn 3.8, GLM 5.3, Kimi K3. And on most of these benchmarks, it is the top DeepSWE, which is the coding benchmark we pay attention to the most. It's not quite as good as Kimi K3. I ran this one through BuseyBench. There was no cost to run it. It generated in 1 minute, 45 seconds and used 9,280 tokens. But you know, it's not winning any beauty contests. I haven't installed this one offline. I unfortunately don't have an M5 Ultra. This one ran through OpenRouter. I haven't done any tests to try to see if it'll, you know, make 3D print files or build a good Mega Bonk game because, well, it's an open weight model that's not gonna be super great at that stuff yet. Microsoft also announced a new open weight model. They say, we recently completed pre-training on our new state-of-the-art model, MAI Base 2. It's got 80 billion active parameters and supposedly outperforms every open weight model they tested. This model hasn't completed post-training yet, so it's not actually available and not totally benchmarked after post-training. So we don't have a lot of information about it, and it's not one we have access to. So I haven't been able to do any testing with it or anything Yet. Nvidia also released a new open weight model. This one's called Nimotron 3.5 Super VL. This one's a 120 billion parameter model. They actually shared where it lands on artificial analysis, and we can see it falls right here with a score of 26. So not quite up there with like DeepSeek and Quin 3.8 or GLM 5.3. So, you know, it's a decent model, I guess, but it's not totally up within the state of the art or even really state-of-the-art among open weight models yet, but it appears to be designed to be really, really fast. So output tokens per second, we can see it's all the way up here where some of those other models like DeepSeek, GLM 5.3, Quin 3.8, etc., are way down here in speed. They're smarter but not nearly as fast. Like, that seems to be its biggest selling point. Of course, I BUCI benched this one. It had no cost to Took two and a half minutes and used 14,400 tokens. And yeah, that's its interpretation of Busey. And finally, the last open weight model we got this week: Ultramarine One Flash from Paper Instruments. They claim this one is Fable class on knowledge work, but at 60 times cheaper than Fable. So we can see the benchmarks that they sort of cherry picked to show off here: Job Bench, Pi Knowledge Work, Frontier Finance Automation Bench, Diligence Bench, A. Apex agents. You can see it does well on those, but the models it's comparing it to are models like GLM Flash, Gemini 3.8, Claude, Fable, and 5.6 SOL actually come in underneath on these, but neither of those are these companies' current state of the art either. The real selling point on this one is the cost. You're looking at 15 cents per million token input, 50 cents per million token output, so that puts it fairly close in cost to GPT-6 Luna Max, but slightly smarter. This one wasn't available at the time of this recording inside of OpenRouter, so I haven't been able to do any sort of benchmarking on it myself yet, but I imagine it'll probably, you know, look in this general range like a Nimotron or Spark Open sort of level. Okay, so last week I talked about OpenAI's 28 Days of Shipping. Thibault put this post out on October 4th. Over the next 28 days, each day we'll either ship one thing is a clear improvement and relevant for most Codex/work users, or ship a full reset. In last week's video, I shared everything they shipped last week, so let's see what we got this week. Quick spoiler, wasn't as great of a week. So last Friday— I record these on Thursdays, so this one came out after I recorded it— they released Composer Predictions. If you're wondering what that is, basically when you prompt something and it gives you a response, it sort of guesses what it thinks your next prompt is going gonna be. So it sort of pre-types the next prompt for you. So that was the day 5 treat. They also released the ability to create your Dot straight from your phone. Before, you had to create it on the desktop app, and then you can go and use it from your phone once it was created. But now they gave you the ability to actually create it from your phone. They also gave you the ability to delegate work in Codex to your Dot. So if you're doing something in Codex And you have your dot. They can sort of communicate back and forth now. So three, you know, kind of minor updates on last Friday, which was day five. Then day six, no new updates. So they gave everybody a reset. Day seven, they increased default context length for GPT 6.1 Soul to 500,000 when used through Codex and Work, which is you know a little less than double. 272 was the previous. So you got a larger context length. On 6.1, day 8, no updates, so we got a reset. Day 9, they rolled out optimizations to how they store state on disk for ChatGPT and Codex desktop app that should reduce the amount of stored disk by 60%. This was kind of like a who cares update. People are calling them out saying this isn't really as much of a feature as it is you fixing a bug that was already in the system. Day 10, no new updates, so they gave a reset. And then day 11, which happens to be the day I'm recording this, Thursday, we've landed a series of changes and more compute to very significantly decrease the occurrence of disruptive selective model isn't at capacity, please try different models. So their big update for day 11 was OpenAI got more compute so it wouldn't screw up as much, which once again, people are complaining this isn't improvement, this is a bug fix, and basic service reliability shouldn't be a headline of a 28-day shipping event. But that was their headline for Uh, day 11, which is the day I'm recording this. So there should be 17 more days of updates or resets, and I'll update you in the next Friday's videos with what we got the following week. But again, all fairly minor stuff, nothing that anybody's getting super excited and shouting from the rooftops about. All right, so let's move on. There's a lot more that I want to share, but a lot of this is sort of smaller, quicker stuff, so we're gonna jump into a rapid-fire. So recently, decision models like Jev have been all the rage, and even OpenAI put out their own decision model last week. These aren't LLMs. These are models that they read text or look at things and then make a quick decision about those things as opposed to outputting a ton of text. Well, we got 2 new decision models this week. We got one out of Microsoft called Microsoft Decision One. Pretty much does the exact same thing as Jev, but supposedly with a bit higher accuracy. And we can see, just like Jev, when given a fixed set of answer options, Microsoft Decision One provides a calibrated probability score for each option. The model supports yes/no, multiple choice, and rating options, as well as a rubric-based grading of AI responses and agent actions. So it does the same thing that Jev does and the same thing that the Decisions model out of OpenAI does. But we also got a second decision model this week from a company called Bespoke Labs. They released Nimble V4 and Nimble V4 Max, which are pushing the frontier of decision models. Apparently, they also outperform Jev. Here's some benchmarks they shared. They also have multiple choice, yes, no, and rubric scores. Has fewer errors. And yeah, I mean, again, another thing that kind of does the same thing as Jev, Microsoft's decision model, and OpenAI's decision model. They all do the same thing. They're all just slight differences in accuracy. SpaceX. SpaceX AI or SpaceX SI, whatever they're calling themselves. They released a new voice model, Grok Voice 3.0, and it's pretty decent. I actually built a little like mini agent earlier that worked pretty well. I'm gonna link up everything I talked about in the description below the video, but if you click on try it for free, they do have a little sandbox that you can play around with it in. And I created this like future tools support, and it actually just asked me questions about what I wanted support to do, and it wrote the custom instructions and the welcome message and all of that kind of stuff. You pick which voice Have

you heard the new Grok voice? Let me tell you a secret. Have you heard the new Grok voice?

Have you heard the new Grok voice?

You get the idea. There's pretty decent sounding voices, but here's what it sounded like when I created my little agent. I can actually—

Hi, thank you. It sounds like you might have gotten cut off. What were you about to say?

Yeah, you were talking when I was trying to talk. Stop talking when I'm trying to talk.

You're right. I'm sorry. Go ahead.

I make YouTube videos. What are some good tools over on Future Tools for people who make YouTube videos?

Let me check the site for that. Here are a few that stood out. Tube Spanner is an all-in-one YouTube toolkit. It helps with scripts, video ideas, thumbnails, titles, chapters, and promotion. InVideo lets you turn a text prompt into a full-length video with scripts.

Okay, wait, wait, wait. I've got another question. What is the latest news in the AI world?

Sure. Today's top stories on Future Tools are OpenAI is bringing GPT Rosalind to academic research labs. Meta released Muse Spark Open, an 816 billion parameter open weight model.

You get the idea. I basically created a little customer support bot where somebody can call into Future Tools and it only has the context of the Future Tools website. So it pulls news from the site, it pulls tools from the site, and if you chat with this assistant, it will use that as its source of information. And it's pretty fast. I think that's what they were kind of trying to show off is that it feels quick. Like I ask a question and it responds almost immediately. It has less of a pause where it's like waiting after I ask to like think it through before responding. Since we're talking about audio, might as well mention this as well. ElevenLabs introduced instant soundtracks. You can generate a soundtrack and synchronize sound effects for any video you upload in Eleven Creative. So you give it a video, it watches the video, and then figures out how to soundtrack that whole video. Facebook made an update to their algorithm this week where you can use AI to shape what you see on Facebook. So now apparently you can directly input your interests to fine-tune what you see, and the ranking system will adapt in real time. I don't think I've logged into Facebook in like 3 years, so I haven't even tested this myself. But if you are somebody that's using Facebook, well, now you can fine-tune your algorithm a little bit better, I guess. Google rolled out a new feature to help you see if a text message is actually a scam. So when a text comes in, you can sort of long-click on it, that'll activate Circle to Search, and then you draw a circle around the message, and then an AI overview will let you know if there's any red flags around it. I don't have an Android phone handy at the moment, so I can't test this out. And even if I did, I would have to have like a scam text sitting there for me to test it with. But good to know that Google's got your back on that stuff. A new AI agent orchestration tool called Kylon was launched this week. I'm sharing this one because it's one I actually got a little bit of a demo of earlier this week and found it kind of impressive. It's a tool where you can have multiple agents and multiple people all in a similar chat messaging system, all working together and all communicating together. And yeah, I just thought it was a really cool tool and thought I'd shout it out for you to go check out. It's called Kylon. We're getting some AI shakeups in the gaming world. I don't know if you've been seeing lately, but there's been a lot of videos about people sort of mashing up games where they're taking Minecraft and Grand Theft Auto and mashing them together, and their Minecraft characters running around in GTA. And people are taking like Counter-Strike and putting the guys on skateboards from like Tony Hawk or something, and they're like skating around in Counter-Strike and things like that. And they're using AI and reverse engineering tools to be able to do that. And well, Microsoft saw that and they created Xbox Labs, which is an experimental sort of division of Xbox where they're going to start playing with this stuff, like the ability to merge games and use AI to generate games. And it's like this whole new side of Xbox where they're going to explore like AI mashups and things. At least that's my understanding of it. And apparently Epic Games is also embracing AI. Tim Sweeney here says, we have a huge pivot underway here at Epic, moving developers around to retool for maximum AI acceleration. They're not trying to do the same stuff with fewer people, but to enable everyone here to build more and better stuff. Now, I don't know if most people in the gaming community are super excited about these The gaming community is one that's been very, very resistant of AI. So as you can probably imagine, it's not going over super well if you read the comments here. But if you ask me, I do think it's kind of inevitable. I think a lot of coding is going to be done by AI in the future, whether it's with games or not games. I think more people get frustrated when assets are made with AI, like the graphics and the sound effects and the voiceover. and things like that. I think people are gonna be more frustrated by that stuff than they are that some of the code was written with AI. I think people are gonna start to get over that pretty quickly if you ask me, but I don't know. Apple had an event this week where they unveiled a few new gadgets. Like there's the new Apple TV 4K, which is of course going to have Siri baked into it. And finally, Apple launched this new Home View thing, which to me literally looks like a bigger version of an Apple Watch that sits on a counter. This is like, I guess, their version of an Amazon Alexa Like little screen dock thing. It pretty much does everything like an iPad or an Apple Watch will do, but on a bigger screen. And of course it has Siri and is going to be designed for you to use AI with like your smart home and stuff. I haven't played with one, so I don't totally know what I'm talking about, but they rolled this thing out. And that's what I got for you today. Again, there was a lot of models. I had a lot of fun playing with Krea this week. I had a lot of fun playing around with Kylon this week. I'm trying to get more hands-on and test more and more of these tools and play with them more, as opposed to just like reporting the news. As with every week, it's my job to drink from the fire hose all week. I'll be overwhelmed. I'll read all the news and keep up with all the tools and test stuff and spend my own money to try to figure out if stuff is worth it for you and just keep my finger on the pulse. So if you want to just watch one video a week where you get the full breakdown of the week of everything you need to know and be completely looped in on AI, well, you should probably subscribe to this channel and like this video, 'cause this is that one video every week. I'm gonna do this every single Friday and completely loop you in. Now, I do record these videos on Thursdays, so if any major news breaks on Friday or late Thursday night, that'll usually end up in next week's video, or, you know, I'll make like a standalone video about it. But yeah, that's what I got for you today. I really, really appreciate you tuning in, hanging out with me, nerding out with me. Lots of stuff this week, and just like every week, I don't think this is slowing down anytime soon. So again, consider subscribing, and I'll make sure you're always in the loop with just one video a week. I do make more than one video a week, but these news videos where I break everything down are just one video a week. So again, you know what to do. Thanks again. Really, really appreciate you. Hopefully, I'll see you in the next one. Bye bye.