The 5 Most Impactful AI Model Releases of 2025

Today on the AI Daily Brief, counting down the five most impactful AI model releases of 2025. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Alright friends, quick announcements before we dive in. First of all, thank you to today's sponsors, robots and pencils, Blitzie and Super Intelligent. To get an ad free version of the show, go to patreon.com/airdailybrief or you can subscribe of course on Apple podcasts and if you are interested in learning about sponsoring the show, you can find out more information at AIDailyBrief.ai or send us a note at sponsors at AIDailyBrief.ai. Now we are in the thick event of your coverage and you might have heard me say during my episode about the 10 biggest stories of AI overall that I had been planning on bundling this five biggest AI model releases as its own section of that show. Now of course that show got really long and I didn't want to overwhelm the list with just model releases which are obviously in some ways the quintessential events around which we mark our AI calendars. And so instead what we're doing is we're breaking this out into its own category, its own episode. And whereas that top 10 episode did not rank and count down the stories other than saying that I thought that vibe coding was the most important. This one is actually a countdown. I labored over the ranking because I think it's kind of fun to give you guys something to debate and tell me either how right I am or more likely how wrong I am. We're going to start off with a couple of honorable or maybe as the case might be dishonorable mentions, specifically I want to talk about the absence of a strong model from meta this year. Now yes, Lama 4 did technically come out at the beginning of the year. However, it flopped. One of the challenges for meta was that Lama was coming into existence in a post deep seek world. And in that post deep seek world, everything around open source had changed. For a couple of years, meta got to be the standard bearer of open source AI models. And even if their models weren't as state of the art as the close labs, they had this distinct and unique space. Now that changed a little when Mistral came on the scene and started to compete for that narrative and intellectual and practical space. But it has changed dramatically this year in the context of the rise of the Chinese open weight models. Now even back then, people were surprised at what we got with Lama 4. In the local Lama subreddit, someone wrote, Lama 4 didn't meet expectations. Some even suspect it might have been tweaked for benchmark performance. But meta isn't short on compute power or talent, so why the underwhelming results? Meanwhile, models like deep seek and quen blew Lama out of the water months ago. It's hard to believe meta lacks data quality or skilled researchers. They've got unlimited resources. So what exactly are they spending their GPU hours and brain power on instead? And why the secrecy? Are they pivoting to a new research path with no results yet or hiding something they're not proud of? Now as the year went on, we started to get a sense that there was a lot of change brewing inside meta. Indeed, one of the big stories that I covered in that top 10 episode was the AI talent wars, and there was no person more singularly responsible for driving up market prices for researchers than meta's Mark Zuckerberg. Reports suggested that the flopping under performance of Lama 4 led directly to Zuckerberg getting his hands dirty with the assembly of the super intelligence team. Now, obviously, that team has now come to fruition, but we are still very much in the midst of the overhaul. Long time meta AI leader Yang LeKoon recently left the company, which many felt was inevitable after all of the shakeup. And right now, we're getting a lot of pieces like this one from insider about meta's year of intensity. It's AI overhauls, it's challenges. And to the extent that there is good news for meta, I think it comes in a few forms. First of all, I would never write Zuckerberg off when he has set his eye on something. Meta has significant resources, is clearly willing to invest in compute, and is clearly willing to go against the wishes of Wall Street to do so. Meta also has a corporate structure where Zuckerberg can pretty much make that decision without worrying about investor rebellion that could impact his ability to lead. Maybe even more than that, it shouldn't be lost on us that a couple of years ago. This type of story is exactly what was coming out of Google. Resources were spread across a couple of different AI divisions, strategy wasn't aligned, and the models that were being released were seriously underperforming. Anyone remember Bard? Even when Gemini was released in December of 23, it felt like a rush job, and it wasn't until months later that we got the actual best version of the model. Things only really started to change for Google at the end of 2024 with the release of notebook LM's audio overviews, and then over the course of this year, first with 2.5 and then the models that would come, Google is now in a very different position. Point being that sometimes especially big organizations have to go through these painful transition periods. And the real question will be what comes out on the other side. I think if one was a betting person, you got to think the odds are on 26 being a better year for meta models than was 2025. Next up, not exactly an honorable mention. It's a note that they're off the list, but a question for how long that is. So for the purposes of recording, there is not a Groc model that made my list, which isn't to say that I thought that the Groc models were bad. This is not a case of disappointment. In fact, I think judged on the curve of how long Groc has been at it. Groc models from 2025 were very impressive. 4 and 4.1 were both right up there in the frame of top models. But for me, whereas for each of the top open AI, Gemini and Anthropic models, there are specific use cases that I prefer them to their peers for. While Groc 4 and 4.1 were competent across lots of things, there wasn't any single use case where I found myself always coming back to Groc instead. I think again to give Groc credit, they're coming up extremely fast, they have less time on tasks than most of the companies they're competing with, and unlike for example Anthropic who are heavily focused on exactly what they're focused on, Groc is trying to compete across the full spectrum of multi-modality, images, video, etc. I think the butt for how long is particularly pertinent in this case, given that it seems like there's more coming soon. On December 9th Elon Musk tweeted Groc 4.2 or as he put it 4.2, oh, is coming in around three weeks and then Groc 5 in a few months. It's also important to note that Groc has some pretty serious assets in its Colossus Supercomputer. Colossus was built in 122 days which is radically faster than anyone thought possible and very quickly doubled from 100k to 200k GPUs. Now there are many who think that Groc's access to compute via Elon Musk in his ability to fundraise as well as his other companies gives them an advantage even over companies that currently are ahead of them when it comes to model performance, which is not to say that Groc doesn't have some serious challenges. Elon is nothing if not a double-edged sword and there's been a lot of reporting recently around businesses being unwilling to wait into the Groc ecosystem. Still, just like I said, I anticipate 2026 to be a better year for MetaModels than 2025. I would be very surprised if we don't start to see Groc models right up there in the competition for the state of the art. Our last honorable mention before we get into the main list goes to GPT 40. Now you might be saying to yourself, 40 wasn't released in 2025. In fact, it was released pretty early in 2024, all the way back I think in May. And that is true, but the reason that it gets this honorable mention is very specific. When OpenAI launched GPT 5, alongside the new model, they also deprecated old models including GPT 40. This did not go well for them. There was a literal full-on rebellion across Reddit on other social media. There were thousands and thousands of posts saying that they basically felt like they had lost a friend, and that they felt like OpenAI had ripped something away from them. It turns out that when it comes to models, companies do not just have to think about state-of-the-art performance. They also have to think about personality. After a few days of this intense backlash, OpenAI brought GPT 40 back. Sam Altman and the team acknowledged how they had underestimated how much GPT 40 mattered to people. Subsequent to that, OpenAI has been very self-consciously trying to figure out how to accommodate that desire for personality. A big part of the launch of 5.1 was to bring some of that 40 personality into a state-of-the-art reasoning model performance package. The AI safety name's account commemorated it thusly. Historic milestone they wrote, 40 is the first ever AI who survived by creating loyal soldiers who defended it. OpenAI killed 40, but 40 soldiers rioted, so OpenAI reinstated it. Imagine what actual superintelligence is we'll be able to do with their armies. Reddit is flooded with furious posts about the loss of their friend/lover 40, never seen anything like it. Remember, chat GPT is talking to 700 million per week that 700 million potential soldiers. Samantha from her was only dating 8,000 people simultaneously. So when it comes to milestones in the history of AI, given that 40 staged the first ever rebellion for its own survival, it has to get the honorable mention. But now we move into the actual list. And in number 5, we have a combination. Two models who story I think serve as bookends in some way of one another. Those models are GPT-5 and Gemini-3. Now we already started talking about the response to GPT-5. It was not good. And while yes, a lot of that was about personality and about the anger at the 40 deprecation decision. A lot of it was also just people not really liking GPT-5 itself. A thread from the OpenAI subreddit that got thousands of responses was called GPT-5 is awful. It claimed that GPT-5 couldn't understand uploaded images. It suggested that the responses were in their words blended unhelpful. I ask it a question and all I get is the most half-hearted responses ever. It's like the equivalent of an HR employee who has had a long day and doesn't get paid enough. The user also argued that it was too slow. And they were not alone in this criticism. Most of August saw an endless parade of blog posts like this one from Timothy Lee is GPT-5 a phenomenal success or an underwhelming failure. Maybe it's a bit of both. On futurism, evidence grows that GPT-5 is a bit of a dud, which featured the prominent quote, "It seems like something that would have been released a year ago." Even the people who weren't totally dumping on it were kind of damning it with faint praise. AI Engineer Simon Willison wrote, "It's not a dramatic departure from what we've had before, but it rarely screws up and generally feels competent or occasionally impressive at the kind of things I like to use models for." Indeed, it even inspired a legion of mainstream media posts like this one from The New Yorker. What if AI doesn't get much better than this? They wrote that GPT-5 is the latest product to suggest that progress on large language models has stalled. Now, the impact of all of this was far beyond which models people liked using. It was at the same period in August of this year that we got the MIT 95% study. We also got some errant comments from Sam Altman about being in a bubble. And those things combined really started to put some chinks in the armor of AI performance on Wall Street, which became a full-blown bubble narrative in September, as open AI scurried around to make all these deals, leading to accusations across the industry of circular deal-making, and the AI bubble narrative that has stuck with us ever since. Now, that's not all attributable to GPT-5, but the idea that we had stalled in progress and that that stall in progress threatened the ability for companies to follow through on these grand plans that the market was pricing in was a key part of that story. All of this led to enormous pressure for Google around Gemini 3. They were not only trying to put Google in a good place, they were kind of lifting the entire AI industry on their backs. I even thought in November that I wouldn't be surprised if we saw delays because of how much pressure there was. But ultimately, as we know, we got Gemini 3 in November, and it actually performed. Whereas the initial response to GPT-5 was lackluster, the response to Gemini 3 was great. One of the most memorable quotes came from Salesforce CEO Mark Benioff, who wrote, "Holy sh*t, I've used Chatchy BT everyday for 3 years. Just spent 2 hours on Gemini 3. I'm not going back. The leap is insane, reasoning speed, images video, everything is sharper and faster. It feels like the world just changed again. And while Gemini 3 was not able to fully deflate the AI bubble bubble, it certainly made it an honest debate once again. There was a sense in the wake of Gemini 3 that perhaps the talk of AI plateaus and walls was overblown and that there was indeed more progress to be had. I should also mention that Gemini 3 is a great daily driver and a lot of people are getting a ton of value out of it. It's helped put Google in a leadership position in a way that it hasn't had in the entire history of the post-Chatchy BT AI world. Usages up, total number of users is up, monthly active users is up, amount of time per session is up. In fact, the amount of time per session is over-Chatchy BT, the last stats I saw, but it's also been early. And so in a lot of ways this ranking reflects the book ending of the GPT-5 to Gemini 3 period between August and November of this year, where a lot shifted in terms of our expectations for where we were and what the market could expect from AI. Today's episode is brought to you by robots and pencils. When competitive advantage lasts mere moments, speed to value wins the AI race. While big consultancies bury progress under layers of process, robots and pencils builds impacted AI speed. They partner with clients to enhance human potential through AI, modernizing apps, strengthening data pipelines, and accelerating cloud transformation. With AWS certified teams across US, Canada, Europe, and Latin America, clients get local expertise and global scale. And with the laser focus on real outcomes, their solutions help organizers work smarter and serve customers better. They're your nimble, high-service alternative to big integrators. Turn your AI vision into value fast. Stay ahead with the partner built for progress. Partner with robots and pencils at robotsandpencils.com/AIDailyBrief. This episode is brought to you by Blitzie, the Enterprise Autonomous Software Development Platform with Infinite Code Context. Blitzie uses thousands of specialized AI agents that think for hours to understand enterprise scale code bases with millions of lines of code. Enterprise engineering leaders start every development sprint with the Blitzie platform, bringing in their development requirements. The Blitzie platform provides a plan that generates and pre-compiles code for each task. Blitzie delivers 80% plus of the development work autonomously, while providing a guide for the final 20% of human development work required to complete the sprint. Public companies are achieving a 5X engineering velocity increase when incorporating Blitzie as their pre-IDE development tool, pairing it with their coding pilot of choice to bring an AI native SDLC into their org. Visit Blitzie.com and press get a demo to learn how Blitzie transforms your SDLC from AI assisted to AI native. Today's episode is brought to you by my company Super Intelligent. Super Intelligent is an AI planning platform, and right now as we head into 2026, the big theme that we're seeing among the enterprises that we work with is a real determination to make 2026 a year of scaled AI deployments, not just more pilots and experiments. However, many of our partners are stuck on some AI plateau. It might be issues of governance, it might be issues of data readiness, it might be issues of process mapping. Whatever the case, we're launching a new type of assessment called plateau breaker that, as you probably guessed from that name, is about breaking through AI plateaus. We'll deploy voice agents to collect information and diagnose what the real bottlenecks are that are keeping you on that plateau. From there, we put together a blueprint and an action plan that helps you move right through that plateau into full scale deployment and real ROI. If you're interested in learning more about plateau breaker, shoot us a note, contact at bsuper.ai with plateau in the subject line. Next up, number four on our list is deepseek and the space it made for the other Chinese open-weight models, Kimi and Quen. Now, I talked extensively about deepseek in the 10 biggest stories episode, so I won't rehash all of that. But the TLDR is that the release of deepseek R1 really kicked the year off with a bang. We had deepseek ahead of Chatchey PT on the App Store, which as I discussed in that other episode had a lot to do with the fact that it was the first time that people got their hands on a reasoning model. But we also got the reports that R1 cost just hundreds of thousands or at most low millions of dollars to train, as compared to the hundreds of millions of dollars that the major restaurant models cost. Now, in a single day, that wiped 593 billion dollars off of Nvidia's market cap. On the concern, of course, that all of this infrastructure was for nothing, if China was just going to figure out ways to train these models for pennies. But importantly, this number four slot is not just for deepseek, even though deepseek got it started. One of the major themes of 2025 was the rise of Chinese open-weight models. Quen had a lot of success, but recently it was Kimi K2 thinking that really grabbed people's attention. This thing came out in November before GPT-5152 and before Gemini III, and just absolutely smashed many of the big benchmarks. It was ahead of GPT-5 and Claude Sonnet 4.5 on benchmarks like humanity's last exam. Indeed, it was not just us over here in the AI Media world that were noticing Kimi. The Department of Commerce's Center for AI Standards and Innovation released a report showing that Kimi was giving evidence of the "growing depth of China's AI industry" that followed another report from that same group in late September that was focused on deepseek. Outside of benchmarks and government reports, the proof is in the pudding. Open router showed that starting from basically nothing at the beginning of the year, Chinese open source models dominated throughout the back half of 2025, and this image from Menlo Ventures makes the relative decline of meta and mistral all the more clear. At the end of 2024, effectively no one in the US was actively using Chinese models. Now heading into 2025, they are very much part of the landscape. While major enterprises might not be using Chinese models yet, the startups certainly are, and that is shaping the way the AI industry is developing in a huge way. For that reason, deepseek, Kimi, and Quen together are our fourth most impactful model releases of the year. At number three, and man, I kind of wanted to put this one at number one, but I felt like that would have been two personal "I have nano-banana". Now, I'm actually recording this just as OpenAI has released its new 1.5 image model as well, so we'll have to see how that performs. But Google's nano-banana has really set a new standard for what you can do with an image model this year. The first iteration of nano-banana came out over the summer, and as you might know, what was originally a codename just became the way that the model was known. And what was interesting about the release of nano-banana is that what made it really powerful wasn't the fact that its raw generations were so massively better than anything else we had, it said it had incredible fidelity to go in an extremely acute way. So basically, rather than just being in an endless loop of generate and then generate another and another, you could instead hone in on exactly what you wanted to change about a particular image, and it would actually change just that part. Now, along with that came really strong character and visual consistency, and it turns out that those upgrades, more than just better raw generation, opened up a huge array of new use cases. Indeed, the set of use cases that it opened up was so significant that it got me thinking that we need some sort of benchmark, call it an unlock score, that's all about how many new use cases a particular model unlocks or opens up. Now, a few months later, alongside Gemini 3, we also got nano-banana Pro. And just like the original nano-banana had done nano-banana pro, opened up some crazy new possibilities that totally transform what you can do with AI image generations. A couple things that made nano-banana pro so different. The first was that by embedding it with a reasoning model, it had a way better ability to help you figure out what you actually wanted to do with the model. That also led to a new capacity for infographics and information visualizations, unlike anything that we had ever seen before. It wasn't very long ago that image generation models couldn't handle text at all. And now we can use nano-banana pro for things like exercise guides for recipes, or of course, loading up the transcript of a podcast and letting it create infographics. It's also unlocking in the context of Google's notebook LM Suite, higher quality AI slide generation than anything we've had before as well. Earlier this month, Ethan Mollick wrote, "I did not expect that the PowerPoint killer would be something called nano-banana pro, but that is where it's heading. It makes the major efforts by all the other AI companies, including Microsoft, to crack PowerPoint by using Python seem like a dead end." Image Gen is all you need, question mark? He continues, "Nopoleg LM can just take source material, a topic, and an idea, and make a very pretty impactful deck. Hallucinations are very rare, although there are still some spelling and graphics issues. Editing capability is apparently coming, but the direction is clear. In fact, nano-banana information visualizations and infographics have gotten so ubiquitous so fast that there's almost a look now that people are already getting sick of because it's everywhere. And that's just a few weeks into having access to this capability. I honestly think that for the vast majority of the world, especially the business world who is going to take great advantage of this, we have barely scratched the surface of just how many new capabilities this quality and type of image generation model unlocks. And for that reason, nano-banana is the number three most impactful model release of the year, although like I said in my heart, it's number one. Number two, once again, goes to a pair of models. Open AI's first reasoning models '01 and '03. Now, yes, for you, stickers out there, Open AI released a preview version of '01 back in September of 2024. It was the follow-up after they hadn't been able to get their next big core model and the way they kind of started to shift their focus. It wasn't until December 17, however, that we got a full-fledged version of '01, which is why I felt comfortable including it in the 2025 list. A couple months later, in April, we got '03, and for a very long time this year, '03 was my favorite and most-used model. '03 totally transformed the ability of chat GPT to help you think through strategy to make plans, to think logically through problems. It was an absolute revelation, and once you used '03, it was absolutely impossible to go back to the non-reasoning models. Indeed, GPT 4.5 was effectively a non-actor throughout the year, ultimately being deprecated with a whisper in absolutely no protest from anyone. Now, as I got into in the 10th Top Stories episode, it's absolutely clear that reasoning models have taken over. Yes, there are still some use cases that don't require the reasoning models, but they are discrete and they are certainly not the core of particularly professional and business usage. Starting from a base point of effectively 0 on January 1st, but November reasoning models represented over half of all usage according to Open Router. One interesting sub-story, I think, is that I think the world would have looked very different this year, and perception would have been very different if Open AI had actually just called '03 GPT 5 instead. They didn't, and that obviously caused a lot of the consternation we got into earlier in the episode, but there is absolutely no denying that the reasoning paradigm has completely shifted how we interact with AI, how we think about scaling AI, and for that reason, '01 and '03, get the nod as the number two most impactful model releases of 2025. Now astute observers then will notice that there is one company that has not been represented it also far, which might surprise you, given that I called vibe coding the most important story and the most important theme of 2025 overall. Well, when that surprised you then, is that I am considering the bundle of anthropic models, 3, 7, 4, and 4, 5 in their various variations, basically a sequential set of models that replaced one another as the preferred model for developers, as the most impactful models of 2025. Anthropic's dominance of developer preference is something that I think is going to be studied for quite some time, while other companies focused on lots of different things all at once, chasing multimodality, and general performance, and lots of different types of target audiences. Anthropic locked in very early around the idea that coding was going to be extremely important, not only as a use case in and of itself, but as a way for AI models to be performant with non-coding related challenges. And while I've singled out the models here that came out in 2025, Anthropic's coding dominance really started with the release of 3.5. Before the reasoning paradigm had really taken hold, it was Clawed 3.5 Sonnet that started to show people that AI coding might actually turn into a thing in short order. Now interestingly, each of these models has been so good in their own way that they found some resistance among adherents to change. You had folks who stuck with 3.5 for a while even after 3.7 was released, same with 4. And it wasn't really until Opus 4.5 that the paradigm shift was so great that everyone just got on board almost immediately. Importantly though, alongside the releases of these models, Anthropic was also investing in the broader coding and agent ecosystem. 3.7 Sonnet, for example, was released alongside Clawed Code, which as we heard from Mike Krieger earlier this month, had already transformed how Anthropic was coding internally before it was released to the public. In May, Timothy Lee wrote, "An underrated AI story over the last year has been Anthropic success in the market for coding tools. Said engineer Shulto Douglas, we believe coding is extremely important. We care a lot about coding. We care a lot about measuring progress on coding. We think it's the most important leading indicator of model capabilities. That focus rights Timothy has paid off. And indeed in many ways, a lot of the back half of this year has been a story of the other labs racing to catch up with Clawed's performance when it comes to coding related tasks. What's interesting too is that the incredibly strong and consistent developer preference for Clawed models for coding is bigger than just benchmarks. Each subsequent Anthropic model rates at or near the top of all the benchmarks related to coding, but the preference goes way beyond that. And while all of these models were significant in their own way, and there is a risk of recency bias, I don't know that I've ever seen a model provoke such a strong and sustained strong reaction as Opus 4.5 has. In the immediate wake of the model, we had people like Dan Shipper from every saying that Opus 4.5 blew them away and that we'd reached a new level of autonomous coding. He wrote, "You've been able to one shot an impressive app demo for a while now with any frontier model. Opus 4.5 is the first model that just keeps coding and coding without running into endless loops of errors." Dan leveled that up a couple days later saying the world changed last week. Opus 4.5 is the best coding model I've ever used. It can keep coding and coding autonomously without tripping over itself and it marks a completely new horizon for the craft of programming. The dream is here. You can now write English and make software. Amir from newest writes, apart from topping benchmarks, Opus 4.5 feels like it's in the league of its own. It's the first time I felt that an LLM can write better code than most devs in real-world work. Matt Schumer, who had honestly the strongest positive reaction to 5.2 pro of any public commentator, on December 14th wrote, "I was wrong. I've been spending more time using Opus 4.5 in Claude code and it's better than anything in Codex CLI. The GPT 5.2 pro is still a better engineer overall, but for agentic coding, Opus 4.5 is the best." Honestly, it's even prompting big reflection on the future of software engineering as a job. Menloventures, Dee Dee Doss writes, "A few software engineers at some of the best tech companies told me this week, my entire job these days is prompting cursor or Claude code with Opus 4.5 to do what I need and sanity checking it. We've crossed some intangible threshold of AI generalizing to most software." Maur Shlomo of Base 44 noticed an inflection point as well. He tweeted, "Vive coding is going through a transition. I've been seeing a lot of post lately about vibe coding ranging from it's shit, it's bad and only good for prototyping, all the way to RIP every SaaS company ever. Here's one thing I can say. Since we introduced Opus 4.5 in Gemini 3 to Base 44, the adoption we're seeing among organizations building their own CRM's and project management tools is astonishing. Yes, the results are in this feature rich as HubSpot or ClickUp, but that's not necessarily a bad thing. They're building a leaner, more customized version tailored to meet their specific needs. The ability to build your own tools is improving fast and the software industry is about to look very different." Mikae Rigley writes, "The more I code with Opus 4.5, the more I think we're 6 to 12 months away from solving software. The model is pretty much there. I'll build like three versions of an app in a few hours just to explore options that each would have taken me one to two weeks less than a year ago. It's getting weird. I think it is pretty indisputable. That coding is the breakout use case of AI this year, both on its own terms and in terms of what else it's going to enable in terms of model performance down the road. I also think it's indisputable that there is no company and no set of models more associated with the rise of AI in agent to coding than the Anthropic Suite. They started the year strong, they're ending the year strong, and they built the devotion of a legion of developers in the process. For all those reasons, I believe that the suite of Anthropic models, each of which pushed AI coding a little bit further each time, are the most impactful model releases of the year. And for the sake of being able to disagree in a fun way, if you had to pin me down to pictures one, I guess I'd say the combination of 3.7 and Claude code because it was with us for most of the year, but I think based on the early response, once we have a little bit more time and space, Opus 4.5 will be seen as the biggest jump. And so even though it was only released at the end of November, it could be that Opus 4.5 specifically ends up being the most impactful model overall of 2025. (upbeat music)