3 Major New AI Model Releases: GPT-OSS, Claude Opus 4.1, Genie 3

0:00Today on the AI Daily Brief, three big new model releases from OpenAI, Anthropic, and Google. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:18Hello, friends. Quick announcements before we dive in. First of all, thank you to today's sponsors, KPMG, Blitzy, and Superintelligent. To get an ad-free version of the show, go to patreon.com slash aidailybrief. And if you are interested in sponsoring the show, I actually have a new place to go now. For fun this past weekend, I vibe-coded up AIDailybrief.ai. If you're watching, you can see it's a cool old kind of terminal style website. And so you can now email sponsors at AIDailybrief.ai. That is sponsors with an S at AIDailybrief.ai. You can also still me, NLW, at basically anything you can imagine and it'll get to me.

0:54Check out the website, send me an email about sponsorship, and I will look forward to sending you more information. The last thing I want to note is, unsurprisingly, given how much was going on, today's episode is just one big long main episode about these three big model releases. As makes sense this week, we will be back to our normal format, or we'll just continue to focus on whatever crazy big new thing gets launched. For now, though, we've got an open source model, a world model, and a new coding champion to talk about, so let's dive in. Welcome back to the AI Daily Brief. It's been clear for a while that this was poised to be a very big week.

1:29Insid suggested that in addition to the much fabled GPT-5, we were potentially going to get a new version of Claude from Anthropic, and even some treats and something interesting from Google, to say nothing of a bunch of other smaller models from the companies just one level back from the foundation model giants. And today, indeed, we got drops from OpenAI, Anthropic, and Google. So we're going to do our best to go through all of this. This is all very fresh out of the gate with people's very first takes. And of course, we'll be spending a lot more time on this throughout the week. We kick off not with GPT-5, that's still forthcoming at this point, odds on bets are that that comes on Thursday, but instead with a release that to some is as or more significant.

2:08And that is OpenAI's two new open weight models. The two models are GPT-OSS-120B and GPT-OSS-20B. And the big blinking banner headline is that these are not just some second tier models to give developers something nice to play around with. these are actually very, very close to state-of-the-art. Sam Altman tweeted, GPT OSS is out. We made an open model that performs at the level of 04 mini and runs on a high-end laptop. WTF! And a smaller one that runs on a phone. Super proud of the team, big triumph of technology. Now, before we get into how OpenAI describes this model, it's worth putting this into context.

2:46It has been a very long time since OpenAI was actually in the business of open models. The last one that they released as an open model was all the way back at GPT-2, which by the way remains at this point the most downloaded text generation model of all time on Hugging Face, and by a fairly significant amount. And over the years as OpenAI had increasingly turned away from open source, there were still people who were asking them to reconsider. Back in December of 2023, AI entrepreneur Varun Mather wrote, Open source GPT-4 and do whatever R &D needed to ensure it can run on consumer laptops and desk tops.

3:20You got into AI to change the world. Now, in January of this year, we got the idea that OpenAI might be shifting their stance on this. In a Reddit conversation, CEO Sam Altman said that he thought that perhaps OpenAI had been on the wrong side of history, to use his phrase, when it came to open source. Now, not coincidentally, this was in the peak of deep-seek mania, when Chinese open-source models in general had not only taken a major jump ahead of Western open-source models, but were creeping up right on the heels of the closed models from the biggest foundation model companies. And so with all that in mind, developers in the AI community at large have been excitedly waiting to see what this open source model release would bring.

3:58So let's talk about how OpenAI presented these two models. GPT-OSS-20b they frame as a medium-sized open model that can run on most desktops and laptops, while its bigger sibling 120b is a large open model designed to run in data centers and on high-end desktops and laptops. One of the first things people wanted to check is what the license with this would actually be. And OpenAI led with that as one of their four main bullets, calling it a permissive license. They write, These models are supported by the Apache 2.0 license. Build freely without worrying about copyleft restrictions or patent risk whether you're experimenting, customizing, or deploying commercially.

4:33Simon Willison was enthusiastic, tweeting excitedly that both had been released under a quote, proper open source Apache 2.0 license. But were these just underperforming models designed for tinkerers and hobbyists? The short answer is absolutely not. They shared comparative benchmarks across four tests, including two competitive math tests, the AIME 2024 and 2025, as well as the MMLU, GPQA Diamond, and Humanity's Last Exam, which were all reasoning and knowledge tests. On the competition math, both 20B and 120B actually beat OpenAI 03, and were just behind 04 Mini, which had been specifically designed for math.

5:10On the reasoning tests, both models were behind 03 and 04 Mini, but not by a ton. For example, the 120B model scored a 90 on the MMLU as compared to 03's 93.4, an 80.1 on the GPQA Diamond as compared to 03's 83.3, and there was a slightly bigger gap on Humanity's last exam, scoring a 19 as opposed to a 24.9. Still, that 19 was better than 04 Mini, which got a 17.7. The point is that these models are distinctly not second tier. Matt Schumer writes, Are you kidding me? OpenAI's new open-source model is 03 level. This is going to disrupt the market in a big way. In addition to the permissive license and the benchmarks, OpenAI called out three other aspects of the models.

5:52Full chain of thought, giving developers the ability to access the full chain of thought for easier debugging and higher trusted model outputs. Deep customizability, OpenAI writes, adjust the reasoning effort to low, medium, or high, plus customize the models to adapt to your use case with full parameter fine tuning. And the last one and one that caught my attention was designed for agentic tasks. Leverage powerful instruction following and tool use within the chain of thought, including web search and Python code execution. Matthew Berman honed in on this point, writing, Excellent instruction following, function calling, web search, or Python tool use, adjustable reasoning effort, and structured outputs.

6:25And it's very clear, as we'll get into in a minute, that again, far from being some hobbyist project, I think they have some real specific use cases in mind, and a big part of it is building agents. Sam Altman also wrote a longer post about OSS. He said, GPT-OSS is a big deal. It is a state-of-the-art, open-weights reasoning model with strong real-world performance comparable to O4 Mini that you can run locally on your own computer or phone with the smaller size. We believe this is the best and most usable open model in the world. We're excited to make this model the result of billions of dollars of research available to the world to get AI into the hands of the most people possible.

6:59We believe far more good than bad will come from it. For example, GPT-OSS 120B performs about as well as O3 on challenging health issues. We have worked hard to mitigate the most serious safety issues, especially around biosecurity. GPT-OSS models perform, comparably to our frontier models, on internal safety benchmarks. We believe in individual empowerment. Although we believe most people will want to use a convenient service like ChatGPT, people should be able to directly control and modify their own AI when they need to. And the privacy benefits are obvious. As part of this, we're quite hopeful that this release will enable new kinds of research and the creation of new kinds of products.

7:35We expect a meaningful uptick in the rate of innovation in our field, and for many more people to do important work than we're able to before. OpenAI's mission is to ensure AGI that benefits all of humanity. To that end, we are excited for the world to be building on an OpenAI stack created in the United States based on democratic values, available for free to all and for wide benefit. Now, there is a lot to unpack in there. If you listen to my episode about the White House's AI action plan, you will have heard me talking about the fact that one of the things that made that document so interesting is that while a lot of foreign policy in the United States right now involves withdrawing from our traditional role in the world, the AI Action Plan basically made it a prerogative of the US government to use open-source AI as a sort of soft power.

8:16That sort of theme isn't new for Sam Altman, but with this model and the narrative around its release, they're definitely leaning all the way into that. When it comes to these safety and risk issues, Kai from OpenAI added a few more details. They tweeted, Open weights can't go back in the box. Before permanently releasing GPT-OSS models into the ecosystem, we estimated marginal frontier risk by deliberately eliciting bio and cyber capabilities via malicious fine-tuning. While MFT improves performance, GBT-OSS stays below the high threshold in OpenAI's preparedness framework. These findings contributed towards the decision to release these models for the world to use.

8:50Alongside the release, they shared a paper called Estimating Worst-Case Frontier Risks of Openweight LLMs. But what about what these models are actually useful for? Why is it important if you have access to the closed models that that OpenAI has now made these near-state-of-the-art models available in this open way. The team at Every spent a little time with the models before their release, and here's what they had to say about them. They write, I can think of a few ways right off the bat where we're going to start experimenting with these models internally. First, Every has a consulting practice with hedge funds and private equity firms that are bound by significant privacy and security regulations.

9:23I imagine this is going to be immediately interesting for these businesses because they can now run the models themselves in the security of their own data centers. Previously, they may have wanted to use open-weight models like Kimi K2 or Kuen, but for firms with strict regulatory or compliance requirements, there's no better way to make your head of IT break out in hives than to try to install Chinese-developed AI in your secure private cloud. OpenAI's OSS series models boast the same open-weight flexibility and security benefits without the geopolitical drama. Every also points out that even beyond big strategic risk, there are a lot of companies that just don't want to send their information to anyone's cloud.

9:57They write, Every has a suite of Mac apps that could use these models to offer AI features with a better security and privacy profile. Sparkle, our AI file organizer, sends your file names to OpenAI's cloud, a no-go for security-conscious users. Similarly, Monologue, our AI dictation app currently in beta, sends your dictation transcript to the cloud for processing. We've experimented with on-device processing for both of these apps, but speed and reliability haven't been good enough to release it. And I think that the point here is that in addition to just the principle of the thing, and giving companies and users the choice of how they want to toggle their security and privacy settings, there are actually very specific use cases in industries where this might matter significantly.

10:36Basically, the more regulated an industry is, and the more complex it is, the more likely to want custom solutions they're going to be. We see this all the time at Super Intelligent when it comes to how they're thinking about agents. I don't think it's an accident that OpenAI is really honing in on how good at agentic capabilities like tool calling these models are, because I think that they anticipate that a lot of the usage is going to be in the form of privacy and security conscious enterprises using these models to build their own custom agentics. Harrison Chase and the team of LangChan already started writing about how to use GPT OSS for this type of agentic use case.

11:09Chase wrote, deep agents require good tool calling capabilities, something that OpenAI's new open source model is pretty good at. And for companies that are doing a lot of that usage, there is another part of this story, which is also equally compelling, which is the cost profile. Matt Schumer again writes that the pricing when delivered through Grok is 91 % cheaper than O3's pricing. He continues, because of this, you can likely use this model in much more comprehensive chains and agents to eke out more performance than O3 dollar for dollar. Example, running five OpenAI OSS model generations in parallel and then selecting the best is still way cheaper than just one O3 query.

11:44Grok's head of partnerships, Jacob Lowenstein, wrote, in turn accidentally priced OpenAI GPT OSS at 1 100th of Anthropic. Bad intern, but what's done is done. Now go build. Cerebris writes, OpenAI GPT OSS 120B is live on Cerebris. 3 ,000 tokens per second, the fastest OpenAI model on record. The point being that with these models, you have something that is not only available with a privacy and security profile that was not previously possible, but at near state of the art and for a fraction of the price and at significant speed. I'm sure we'll be spending a lot more time on these open source models and how people are using them in the weeks to come, but this is just one of three announcements that I wanted to speak to in today's show.

12:25Stephen Heidel, who works on the API OpenAI, wrote, also congrats to my friends at Anthropic on the release of Opus 4.1 today. Today's episode is brought to you by KPMG. In today's fiercely competitive market, unlocking AI's potential could help give you a competitive edge, foster growth, and drive new value. But here's the key. You don't need an AI strategy. You need to embed AI into your overall business strategy to truly power it up. KPMG can show you how to integrate AI and AI agents into your business strategy in a way that truly works and is built on trusted AI principles and platforms. Check out real stories from KPMG to hear how AI is driving success with its clients at www.kpmg.us slash AI.

13:08Again, that's www.kpmg.us slash AI. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context. Blitzy 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 Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy 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.

13:45Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Blitzy is providing a limited-time, 30-day free proof-of-concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org. Visit blitzy.com and press book demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native. That's B-L-I-T-Z-Y dot com. If you are a regular listener, you will have heard about Superintelligent's agent readiness audits at this point.

14:23But I wanted to tell you today about the full suite of agent readiness products that go beyond just the initial readiness report. Over the last six months, Superintelligent has built out an entire agent planning suite. We help you move from discovery to planning to implementation. After you've completed your agent readiness audits, we help you double-click on your most important use cases with what we call our use case planning reports. These reports are going to help you understand what sort of technical preparation you need to do to be ready for a use case, what challenges you might face in implementation, and whether you should be thinking about building, buying, partnering, or some combination.

14:57After that, you can even get a spec document in what we call our technical blueprint that gives either your developers or the developers of the partner you work with what they need to build exactly the agent that you're looking for. If you want to learn more about Superintelligence Agent Planning Suite, we've built a custom GPT to answer your questions. Just go to bit.ly slash super agent. That's bit.ly slash super agent, all one word. And if you have any questions, the agent can even help you book an appointment with our team. Just before the OpenAI announcement, Anthropic tweeted, Today we're releasing Claude Opus 4.1, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning.

15:37And the big story here is software engineering. On the SweetBench verified test, Opus 4.1 scored a 74.5 % as compared to Opus 4's 72.5 % and Sonnet 3.7's 62.3%. The quote that everyone has been grabbing from their announcement post is this one. Windsurf reports Opus 4.1 delivers a one-standard deviation improvement over Opus 4 on their junior developer benchmark, showing roughly the same performance leap as the jump from Sonnet 3.7 to Sonnet 4. Menlo's Didi-Dos writes, Anthropic just dropped Opus 4.1, the best coding model in the world. State-of-the-art 74.5 % on the sweet bench, surpassing both Gemini 2.5 Pro and OpenAI 03.

16:16This is the product right now to use. Chubby points out, however, that maybe the thing to really focus on is the comment that follows the announcement post, We plan to release substantially larger improvements to our models in the coming weeks. That's echoed, by the way, in their announcement blog post. Now for some, it feels like that was the main point. New Form AI's Alec Velikhanov said, Is this it? Opus 4.1 feels like a rushed release to get ahead of GPT-5. Look at how it struggles with making a UI that Horizon, assume GPT-5, nailed in a single shot. Now this has been out for less than two hours at the time of recording, so we haven't really gotten a chance to see a ton of people playing around with it yet.

16:52But there's certainly no doubt that the timing is not a coincidence, and that part of this is the competitive pressure to always be at the state of the art. We haven't gotten GPT-5 yet, but we have heard tons of rumors that it is leagues better in coding than any previous OpenAI model, and people have had tons of success with the assumed tester models like the codenamed Horizon, which could theoretically threaten Anthropik's crown at the top of the coding heap. Much respected Simon Willison says, My favorite thing about this model is the version number. Treating this as a 0.1 version increment looks like it's an accurate depiction of the model's capabilities.

17:25Basically, he's saying that yes, it is a modest improvement, but an improvement nonetheless and one that is appropriately pitched. Now, the other new model that I'm really excited about that was announced today is Google's Genie 3. This they're calling the most advanced world simulator ever, and there's really some amazing stuff going on here. So the basic idea of Genie 3 is that rather than prompting a video, with a few words, you're prompting an actual dynamic world that you can interact with. They write, given a text prompt, Genie 3 can generate dynamic worlds that you can navigate in real time at 24 frames per second, retaining consistency for a few minutes at a resolution of 720p.

18:05Let's actually just watch part of the video to get a sense of how they explain things. Each one of these is an interactive environment generated by Genie 3, a new frontier for world models. With Genie 3, you can use natural language to generate a variety of worlds and explore them interactively, all with a single text prompt.

18:31Let's see what it's like to spend some time in a world. Genie 3 has real-time interactivity, meaning that the environment reacts to your movements and actions. You're not walking through a pre-built simulation. Everything you see here is being generated live as you explore it. And Genie 3 has world memory. That's why environments like this one stay consistent. World memory even carries over into your actions. For example, when I'm painting on this wall, my actions persist. I can look away and generate other parts of the world. But when I look back, the actions I took are still there. And Genie 3 enables promptable events, so you can add new events into your world on the fly.

19:18Something like another person, or transportation, or even something totally unexpected. You can use Genie to explore real-world physics and movement, and all kinds of unique environments. So the real world physics are a really important part of this. We talk about AGI like something that is a straight line path from where we are to there. And certainly as more models become more capable, that feels accurate. But the reality is there are different beliefs about the different paths that it's going to take to get us there. And if you've ever listened to our coverage of world models before, you'll know that this is a different path that some people are convinced is a better path towards AGI than the sort of text-based LLMs that have become common at the core of our experience with generative AI.

20:06Google writes that world models are AI systems that can use their understanding of the world to simulate aspects of it, enabling agents to predict both how an environment will evolve and how their actions will affect it. World models they continue are a key stepping stone on the path to AGI, since they make it possible to train AI agents in an unlimited curriculum of rich simulation environments. So part of what Genie has that's different is that it can model physical properties of the world, it can retain environmental consistency like we saw with that painting example, and it can be prompted to actually change the generated world as well.

20:39Now, environmental consistency was one of the things that people latched on as a really big update here. Andrew Curran writes, they solved environmental consistency with Genie 3 and this was an emergent capability. You can see the trees remain the same even after being out of line of sight. Visual memory extends back one minute now. Google is on a steady path to a real world simulator. Chubby writes, The breakthrough, minutes of consistency at 720p using only text descriptions. No pre-built 3D models required. They continue, The technical breakthrough lies in autoregressive generation. Each new frame must take into account the entire trajectory up to that point.

21:14At 24 frames per second, this means 24 complex calculations per second that must access minutes of context. Google DeepMind researcher Jack Parker Holder writes, Genie 3 feels like a watershed moment for world models. We can now generate multi-minute, real-time interactive simulations of any imaginable world. This could be the key missing piece for embodied AGI. Theoretically Media writes, Google's Genie 3 shows the beginning of the endgame for VO3, full world simulation. Just look at this and tell me you don't see the start of the holodeck. And that word holodeck comes up a lot. There are a few different things that I'm seeing people excited about when it comes to the announcement.

21:48One of them is robotics, or to use the term that we just heard, embodied AI. Robert Scoble writes, Having the ability to create a real-looking world that a user can move around in and interact with will speed up robotics. What Google just showed off this morning is a simulator. It simulates the real world, perfect for training robots of the future in. Robots need to learn to navigate the real world. Stairs, for instance. Now developers will be able to develop thin stairs, thick stairs, spiral stairs, muddy stairs, icy stairs, wet stairs, stairs with kids, stairs with dogs, etc. And run virtual robots up and down those stairs billions of times to train new robot models.

22:21Roan Chong also points out the implications for training. He writes, We're entering the era of infinite AI training environments, where world models like Genie 3 enable AI agents and robots to learn from their own experiences in simulated real-world environments. This is the convergence of world simulation, AI training, and creative expression all happening in real time. Still, as you might imagine, a lot of the implications that people are jumping to are those for games. Bilawal Seedhu writes, Genie 3 just achieved what AAA game engines do, but without any 3D models. Wild how this model figured out complex effects like exposure shifts, volumetric god rays, and phenomena we need to code explicitly in 3D engines.

22:57If you squint, he writes, you can see where this goes. VR and AR, half robots, robotics, jungle gyms for robots, sims made from real-world data to train robots to then operate in the real world, real-time virtual production, motion-tracked iPhone and you, quote-unquote, directing actor agents. Jim Fan from NVIDIA writes, This is Game Engine 2.0. Someday, all of the complexity of Unreal Engine 5 will be absorbed by a data-driven blob of attention weights. Those weights take as input game controller commands and directly animate a space-time chunk of pixels. Andrew Curran joked, Given the current rate of world model advancement, the end result of Bethesda taking 15 years to release the next Elder Scrolls is that fans may end up making it first.

23:35To which Elon Musk responded, For sure. Common Sense Machine CEO Tejas Kulkarni had early access, and wrote, I spent the whole day playing with the system and when it works, it's truly mind-blowing. It's the first neural game engine and world model I've tried that generalizes so well and has long-term world consistency. Where it shines, he writes things like that it's truly general purpose and has a quick startup time. Works exceptionally well for gaming environments, but also generalizes to other industrial and real-world scenarios. It learns physics. He wrote, although there are systematic failures even for rigid body physics, it was clear to me that it can learn game engine and non-rigid physics without an underlying engine.