This Month is HUGE! o3 & o4 mini, Llama 4, VEO 2 in Gemini & Much More!

Happy Friday folks. Welcome Back to the Mattvid Pro AI YouTube channel. You guys know that the AI space never sleeps. We've got plenty to discuss today, pretty big updates from big competitors and even new model releases, so let's dive in. First up, OpenAI has a new refresh roadmap for the coming months and for 2025 overall. Sam Altman himself the CEO, tweets Change of plans we are going to release O3 and 04 mini after all, probably in a couple of weeks, do GPT5 in a few months. This differentiates from their previous plan and here's the original tweet from that. This was in February actually, before GPT 4.5 even shipped. In this old tweet he announced that they are no longer going to even ship O3 as a standalone model, stating that GPT5 will be a system that integrates a lot of their technology, including O3 into one bundle and chooses the best model for you at the right time. But obviously things have changed. We are going to get 03s standalone and even 04 mini, which I'm really excited to check out. I'm wondering how far they've been able to push the reasoning benchmarks with these smaller mini size models. And yeah, these models are coming in just a couple of weeks, so that's a mid April release or so. That's some pretty big stuff coming from OpenAI down the pipeline and not that long of a time now. Sam states that there's a bunch of reasons for this switch up, but the most exciting one is that they're going to be able to make GPT5 much better than they originally thought. That's interesting to me. He doesn't really back that up with any specifics, but he's the CEO of the company so we'll just take his word for it. He does state though that OpenAI found it more difficult than they initially thought it was going to be to smoothly integrate all of their technologies into one GPT5 package like the tweet back in February mentioned. He also states that they want to make sure they have enough capacity to support what they expect to be unprecedented demand. And with a name like GPT5, well yes, that will be unprecedented demand no matter what. They were even shocked again at the new GPT4O native image generation release, which totally swamped up their servers. I'm definitely pretty skeptical of whether or not they're going to be able to pull it off, especially because OpenAI themselves found it more difficult than they thought it was going to be and my first impression of this idea of wrapping everything into one cohesive GPT5 that picks out and uses the best models was an idea I wasn't like super huge on. Even personally I still like to pick out my own models for independent and individual tasks, and I'm kind of in the camp where I don't think anybody, including OpenAI would be able to know the best model to use for every single use case out there. They might be able to generalize, but I don't think they'll be able to generalize it perfectly. Either way, if we scroll down here, Sam does say that they were able to improve on what they previewed for O3 in many ways, which is really exciting. So full O3 coming down the pipeline with improvements over what they previously benchmarked and in just a couple of weeks. Some of this also might be a reaction from OpenAI to Google, who really seems to be taking the lead with large language models, at least in the short term right now. Gemini 2.5 is now moving into public preview in their AI studio with higher rate limits and pricing is also available for Gemini 2.5 Pro. We've been talking about 2.5 Pro on this channel for a while now. It benchmarks really really high, it's fantastic at coding and it's very intelligent in all other regards, including creativity, reasoning and especially that high context window. A million tokens There are a lot of good things going for Gemini 2.5 Pro right now, and I'm wondering if OpenAI is sweating a little bit about that. Anyways, yes, they are moving Gemini 2.5 Pro into public preview. Sundar here says that they're setting usage records every day with a model in the Gemini app. Basically stating that this mod is attracting new users. More powerful, more intelligent AI is absolutely where people in the market are gravitating towards. So some actual clarity surrounding pricing. It's cheap at inputs over 200,000 tokens, so very large context inputs. We're looking at a $25 per million tokens with the output being $10 per million tokens. Now keep in mind, relatively speaking this is still more expensive than open source models, but this is a very intelligent state of the art model. This high up there open source models are not beating Gemini 2.5 Pro right now. Now sub 200,000 token inputs are priced a little bit higher at $2.5 per million token input and $15 per million token output. So you're getting a little bit of a discount if you submit in bulk essentially is what the Point of the pricing is here. Regardless, it's still cheap in comparison to competitors and that's the most important bit here. GPT4 Omni not OpenAI's smartest model. In fact, standard model for chat GPT is priced at the higher side, relatively speaking compared to Gemini 2.5 Pro with the same price for input and a little bit cheaper output. Sonnet 3.7 though is more expensive than that. And as I just said in my last video, a lot of people, a lot of taste testers are saying that Sonnet 3.7 does not stack up to Gemini 2.5 Pro, especially when it comes to coding. And yet it's more expensive. $3 per million input and 15 bucks per million output. Now GPT01 more expensive OpenAI model here is far, far more expensive while benching worse than Gemini 2.5 Pro $15 per million input, which is orders of magnitude greater than any of these prices from Google and even more outrageous with the output $60 per million tokens output. So Google is really playing the inexpensive, cheap API game here and they people who just want to mess around with a model to mess around with it for completely free. I haven't had to pay a dime to test out any of Google's latest models. Definitely something to think about here. Google might be able to win in pricing, at least in the short term. We'll see. You know, 03 and 04 Mini might come out swinging, really taking down Gemini 2.5 Pro with ease. And even if it's more expensive, people might still prefer that higher intelligent model. But if the benchmarks are about the same as Gemini 2.5 Pro here, well, it'll be no brainer to go with the Google model. In other Google related news, VO2 is also rolling out to Gemini. It seems Cole points out here that he hopes this is not the horrible turbo model that we got inside of YouTube shorts, and I agree that model was not nearly as good as the full big bad VO2. As it turns out, I actually do have VO2 inside of Gemini Advanced, so why don't we try it out? But first I have a quick message from our sponsor. I'm sure I don't have to explain this to you guys, but I am a huge fan of AI image generation and Recraft's image generation is at the top of the list. You might remember I did a full video talking about them all the way back in November and here's why they still stand out even today. First off, Recraft's V3 model was and still is ranked all the way towards the top of the hugging face leaderboards. This is a difficult user preference based benchmark. Make hyper realistic scenes where even more complex nuanced visuals. Recraft will have your back with lifelike textures and great anatomy overall. But it gets better because Recraft can also do fantastic text. That's right, this sucker works great for banners, posters or even more detailed layouts. Wrap this all up together and you get an AI powered professional design toolbox with powerful image generation and editing capabilities. Now here's the best part. If you want to try Recraft out for yourself, you can get a whole entire month free. If you use my exclusive promo code, you get $12 off of any Recraft plan. And for the basic plan, yeah, it's 12 bucks a month, so that's a free month. So check out that link down below and take your inner creativity for a test drive with Recraft. And if you end up making anything cool, feel free to share it to my Discord server too huge thank you to Recraft for sponsoring today's video and subsequently supporting the channel. Now back to your regularly scheduled content. Welcome back folks. Thanks for sticking through sponsors with me. They make pursuing this channel full time an absolute reality for me. Oh, big deal. Input image is not available with VO2 and Gemini yet. That's to be a deal breaker for many. Still though, even natively the images look pretty great. Let's start easily. Slow motion footage of a fox jumping around in the snow. VO2 is surprisingly quick to generate for how high quality its outputs are. All right. And the video has generated it looks like pretty decent quality. You can see that YouTube training data shining right through. We've got this GoPro footage, you know, slowly moving backwards from the fox and yeah, it's slow motion. Great anatomy. It looks real. I mean that's impressive. I don't think this is a turbo model, at least not from this one example. Let's continue. Next prompt. Physics simulation. Jelly is raining from the sky on 3D animated characters holding umbrellas in a city made from ice cream. Little bit more complex. Obviously we're trying to see some cool Jelly esque physics and it has to do 3D animated characters as well. Alright, here is our physics simulation. Definitely some 3D characters walking around with their umbrellas. We've got Jelly C raining from the sky. Overall, doesn't really seem to be doing too much in relation to the physics simulation. I guess you can kind of see the cubes bouncing in the background. Not a bad Little generation. Though the characters are pretty consistent and lively, I'd say. Definitely a city made of ice cream as well. For the most part. I'm pretty impressed. Let's now do something even more complicated. The camera moves towards the left hand side of a red Miata car on the highway and then stops facing the rear of the car. So the camera is gonna move and then stop and then the car is going to speed off, leaving smoke and a fire trail in a cinematic fashion. Wow. Alright, so the video is generated and yeah, there's the fire and the smoke leaving the trail. You know what? Not too bad. Not a bad video at all. Indeed. I'm pretty impressed by this. Nice and consistent. That training data is really, really solid. It understands prompts well and it can replicate them very coherently. This model clearly just has the best data backing it. I don't think anyone can doubt that YouTube is going to be the best source possible for training AI video models because everything, and I mean everything is on there. All right, one more prompt. The camera follows a cute robot riding a rocket ship through space. The rocket lands on the moon and then the robot gets out of the rocket. Cute 3D animation style. All right, so we've got our robot on our rocket. He is landing on the moon and then kind of hops out and runs around. This generation's kind of a little bit of a flop. I mean, it's not terrible. I like the opening, but this part where he kind of starts running literally on nothing to land on the moon is a little bit strange. And then the rocket kind of takes off again. It gives off that weird AI generated video tab. The second half though, where the robot is running, that's actually kind of crazy good. Where he hops around and yeah, he looks like a lively little dude. Yeah, so this one's a little bit of a flop, but you can see VO2 is a strong video generator. This definitely looks to be a better model than what we got inside of the YouTube Shorts app. So overall I'm pretty impressed. I'd love to check this out in a little bit more depth. In other AI video related news, LTX Studio has some brand new updates. LTX V Distilled, which is their proprietary video model, is now getting some optimizations for more speed and clarity. That's great to see because speed and clarity is absolutely what LTX Studio needs. In my opinion, that app works at its best when it works pretty decently and can get you your generations quick. They also redesigned their video upscaling model. Fewer artifacts, sharper details more consistent frames. If any of you guys are big LTX Studio users, let me know if this makes a big difference for you. In other news, llama4 is also probably dropping this month. Information comes from Chubby here, a phenomenal account to follow if you want to stay up to date in the AI world. It does look like Meta has been struggling to get llama 4 out the door though. Apparently they've been facing some performance issues. It's not benchmarking as high as they'd hoped to see in reasoning, math or even human like conversations. And to be fair, the competition is super fierce. Not only is the open source competition really insane from folks like Deep Seek, but even in the closed source we've got Google coming out with Gemini 2.5 Pro OpenAI later this month is going to be dropping O3 and 04 mini. Either way, they do plan to have their launch also come out this month. Apparently Llama 4 is also switching over to mixture of experts. And of course this comes from Deep seeks surprisingly strong open source performance earlier. Honestly, I'm sure that llama 4 is probably gonna be pretty darn good. It'll be multimodal, it'll probably have reasoning, I'm sure there will be different variants to run locally and if they know what's good for them, they will make it open source in the similar ways that they've done in the past. I'm excited for Llama 4. I think it's gonna be a pretty kickass model, but it is a little bit worrying that they've delayed it so long and that they switched techniques after Deep Seek came out absolutely swinging. I believe in them. Next up, Ideogram V3 has an official placement now on the Artificial Analysis Image arena, so it's clearly a bump up from the previous ideogram 2.0. Right now the arena score is around 1095 puts it in fourth place. Competition is fierce. We've got Recraft in second place right now with an Elo score of 1105 in and Recraft sponsored today's video of course. So you guys know we got the top pack leaders sponsoring us these days. Anyways, Reeve also is a touch behind Recraft here and still above ideogram 3.0. But the number one king seems to be GPT4O that auto regressive native image generation. The fact that it's one model that's also an LLM is giving it a serious leg up against these traditional diffusion models. Very important to point out there are some other models though that are released that are aren't exactly on here just yet. Midjourney V7 is actually finally here. Yep, as you saw my last video, they were gearing up for a launch and and now it's public and it's alpha testing. Nick St. Pierre does a great comparison here by the way. Nick is a long time Midjourney fan, so you bet he knows the best prompts to give this thing the goods. Alright, for the first one here we've got V6 on the left hand side and V7 on the right. Prompt was a young Indian woman with dark hair in an open ponytail and a black jacket standing on a university campus looking directly at the camera. The image has a 1990s style movie still aesthetic with a close up portrait on a sunny day. So the first thing we're gonna notice, a clear coherency bump. I mean it's got the ponytail. University campus I think fits a little bit better and this is much more of a close up style portrait on a sunny day compared to this one. But it doesn't have the 1990s movie still aesthetic which is disappointing to see. Again, V6 on the left versus V7 on the right. Majestic barn owl perched on an ancient moss covered tree. Soft, light, dense foliage, magical and ethereal atmosphere. Photorealistic style with attention to detail of the feathers and textures. V6 again definitely nails the prompt pretty well, but it's a less coherent image. The owl honestly kind of looks a little wonky. He doesn't really look like he's actually there on the branch. V7 knocks it out of the park I think. Much more coherent, still maintains all of those great parts of the prompt. Magical, ethereal and photorealistic. Third example, again, more coherency bumps. A person's hand points towards the window of an airplane which is seen from inside with its wing visible in profile.v7 at the top here definitely has the wing plain door. I mean that looks like a real image. I wouldn't be able to tell otherwise. The six down here is lackluster in comparison. I mean that looks like a straight up open door. Like I could fall right out of the plane and the wing is nowhere to be seen. I won't spoil the entire comparison thread for you from Nick here, so this is linked down below. He's a great account to follow for stuff like this, so if you do use X Twitter, I recommend him. Krakas here has also given us an honest review. Honest opinion from his testing with Midjourney V7 so far. Aesthetics 5 out of 5. That's no surprise for Midjourney who seems to mainly focus on the aesthetic side of things. Prompt adherence only a two out of five. But I imagine that this score is so low here. Not compared to V6 but compared to the other image generators that are out today like recraft, native GPT4, ideogram 3. I mean those all have significant bumps over any mid journey model, even this brand new one. In terms of prompt adherence, finer details gets a little bit of a bump three out of five for sure. And from the examples that I've seen floating around, I'd say that it's definitely improved over V6 in this as well. But. But again, compared to some of these other models that really have nailed the punch down in terms of getting the prompt adherence down while maintaining aesthetic quality and finer details, I just don't know if it can compete. And finally, text Midjourney is not optimized for text. I mean they basically said it themselves. V7 is not designed to really do text all that well, so so we're still stuck with text that is lagging behind the competition. I mean Recraft does great text, ideogram 3.0 does great text, native GPT4 image generation does phenomenal TE most top charting models these days are able to do at least a paragraph of text if not branch out in even other more creative directions with it. TLDR is that midjourney is the king of aesthetics, but that's about it. Things might change in the future, but the competition is very tough. By the way, in this thread, which is also awesome, he does provide a plethora of examples, but mostly midjourney V7 versus other models out there, not versus the old V6. Again, I'm not going to spoil the entire thread for you guys, but but the TL doctor here is that it has great aesthetics and everything else is worse than some other image generator in some way. Again, think about the finer details of things. Majority oftentimes provides artsy examples but doesn't listen to the whole prompt. No rugged landscape like described here and of course given to us by a different image generator. So where's this leave us today? Well, there are a lot of things to anticipate in this month regarding LLMs. OpenAI is going to be doing some big drops in LLAMA 4 should also be coming out, which is huge. Not to mention all the other stuff that's constantly going on in the AI space. Thanks so much for watching guys. I hope this video was helpful. It's actually been kind of a crazy week if you want the bulk of what happened this week in AI, check out my video from yesterday. Yeah, it's another AI news roundup. There was just that much to talk about this week. See you guys in the next video and thank you so much for watching. Goodbye.