Claude Opus 6 LEAKED, Gemini 4 Carbon Leaks, HUGE OpenAI Update, Claude 24/7 Agent, & More! AI NEWS

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Claude Opus 6 LEAKED, Gemini 4 Carbon Leaks, HUGE OpenAI Update, Claude 24/7 Agent, & More! AI NEWS · Channel: WorldofAI · 24:59 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI models, Multimodal models

Summary

The video covers reported AI model tests and launches, including a mysterious Claude Opus checkpoint, Gemini 4 checkpoints, and a larger default context window for OpenAI’s GPT 6.1 Soul. It showcases coding models generating animations, detailed 3D scenes, and interactive worlds, while discussing how ambitious tasks and access to advanced systems shape people’s experience of AI. Other topics include possible new Claude features, a stealth image model, and a free coding model available through Hermes Agent. The video also examines claims about an AI-coded game’s commercial success and clarifies that a viral Cyberpunk-on-Windows-XP demonstration was actually using a Windows 10 mod.

Key points

Anthropic is reportedly routing some Claude Opus 5.5 requests to a newer, unconfirmed checkpoint that may be an update or an early Opus 6 model.
Demonstrations show AI generating JavaScript animations, a detailed Hogwarts voxel scene, and a Tesla model in Blender.
OpenAI reportedly raised GPT 6.1 Soul’s default context window in Codex and ChatGPT work from 272K to 500K tokens; the video says it supports up to 1 million tokens.
Andrej Karpathy’s observations highlight a divide between casual AI users and people using frontier models for complex coding, mathematics, and professional work.
Anthropic code names Arborio and Iggy have prompted speculation about web-based Claude Health and a proactive, always-on AI agent.
A stealth image model called 67 was reportedly spotted on Arena, with speculation that it could be related to a new XGI image model.
Gemini 4 checkpoints, including Barium B and Argon, are reportedly being tested, with demos showing interactive 3D experiences.
A model called Missing No is available for a limited time through Hermes Agent; the video suggests it may be Reflection AI’s Beam model.
The video discusses an AI-coded game’s reported commercial success and corrects a claim that Cyberpunk 2077 had been made to run natively on Windows XP.

Chapters

Start	Chapter	Summary
00:00	Reported Opus checkpoint and AI landscape	The video introduces reports of a new Claude Opus checkpoint, an OpenAI context-window update, and Karpathy’s perspective on the gap between casual and advanced AI use.
04:09	Claude Opus checkpoint and creative coding demos	The host discusses ways users are trying to identify the rumored checkpoint and showcases AI-generated animations, voxel scenes, and Blender models.
10:26	Voxel and Blender demonstrations	The video compares a Hogwarts voxel creation with GPT 6 and shows a detailed Tesla built from scratch in Blender using AI tools.
12:31	OpenAI context window and Karpathy’s AI divide	The host explains the reported GPT 6.1 Soul context-window increase and discusses Karpathy’s view that most people have not experienced AI’s more advanced capabilities.
16:40	Possible new Claude features and image model	The video covers speculation around Claude features called Arborio and Iggy, followed by reports of a new stealth image-generation model.
18:46	Gemini 4 checkpoint reports and demos	The host describes reports of newer Gemini checkpoints and shows interactive 3D experiences attributed to Gemini 4 Argon.
21:50	Free coding model and AI-made game	The video introduces the Missing No model on Hermes Agent and discusses the potential significance of an AI-coded game’s reported success.
23:26	Cyberpunk claim fact-check and sign-off	The host clarifies that the Cyberpunk Windows XP demonstration used a Windows 10 mod, then wraps up with channel and community promotions.

Quotes

It took just two prompts, roughly four lines of instructions, two hours of execution, and a whopping 27 million tokens at max thinking effort. — 00:30

The gap comes down to three things. How ambitious your tasks are, which domains AI excel at, and what level of access you have. — 01:33

OpenAI has officially increased default context window for this model in Codex and ChatGPT work from 272k tokens to a whopping 500k tokens — 12:31

apparently this dude was running a mod on Windows 10 that makes it look like he is running XP. — 23:58

Concepts and entities

Claude Opus 5.5, Opus 6, GPT-6.1 Soul, Gemini 4, Barium B, Andrej Karpathy, Anthropic, OpenAI, Blender MCP, Missing No, Beam, Cyberpunk 2077

Full transcript (original language)

00:00 When I say there’s never a dull day in the AI space, I truly mean it because Anthropic is apparently cooking up yet another new model. And this time it could be something even more powerful than Opus 5.5. That’s not even Fable 5.5. Cuz there are reports that requests right now are currently being routed to this new mysterious Opus next checkpoint. Potentially an updated Opus 5.5 model, which is likely going to be Opus 5.6 or maybe even something from the Opus 6 series. And considering how

00:30 insanely capable Opus 5.5 already is, this could be quite huge. Just take a look at this example. Someone reportedly generated an entire animation depicting the life of a human from every frame drawn using JavaScript. It took just two prompts, roughly four lines of instructions, two hours of execution, and a whopping 27 million tokens at max thinking effort. Whatever model this is, the output is genuinely impressive. And it looks like Anthropic is going full

01:01 throttle as their IPO is approaching. But that’s not all. As a part of Open AI’s 28 days of shipping, day seven brings a pretty meaningful update. With GPT 6.1 Soul reportedly now having a default context window of 500k tokens in Codex and ChatGPT work, up from 272k. That’s roughly an 84% increase with OpenAI apparently planning to expand it even further. Now, Andre Karpathy has made a fascinating observation about the

01:33 current AI landscape, which I think perfectly captures what’s happening. Essentially, people still see AI as nothing more than a chatbot for answering questions or writing emails. Many professionals are expecting relatively modest productivity gains from AI. Meanwhile, a small group of power users working with advanced models in coding, mathematics, and complex projects are witnessing absolutely insane improvements from AI itself. The gap comes down to three things. How

02:04 ambitious your tasks are, which domains AI excel at, and what level of access you have. We’re essentially living in an AI bubble where the capabilities we see every single day would genuinely shock most people who haven’t experimented with these systems. Now, moving over to Google DeepMind because there’s reportedly another Gemini 4 checkpoint being tested right now called Barium B. Now, early comparison suggests that this newer checkpoint substantially outperforms the previous Argon version

02:34 and speculations are pointing that this could potentially be the Gemini 4 carbon update, but nothing is confirmed, but we will be taking a look at a couple of demos of this model. Now, there’s a lot more happening where we have this new free model that is available right now, which is pretty decent. Anthropic is reportedly working on a couple of mysterious features as well as their own generalist AI agent similar to dots as well as Grok Bot, and several new stealth image generation models appearing. And finally, one of the

03:06 craziest stories that we’re going to be talking about is where someone actually claimed to have recompiled Cyberpunk 2077 to run natively on Windows XP with the help of Opus 5.5. And yep, you heard me right, Windows XP. That’s how insane Claude Opus 5.5 is getting the ability to recompile this game, this insanely extensive game that is very GPU heavy is being ran on Windows XP thanks to Opus 5.5, and that is a ridiculous technical

03:37 achievement that was done thanks to this AI model. Now, there’s a lot happening in the world of AI, so I highly recommend liking this video and subscribing to the YouTube channel as it really helps me out, guys. But, an update that I want you guys to actually know about is the world of AI benchmark has been updated drastically where we have a lot of new features, a new UI change, and we are going to be adding in this new feature of AI news as well as build rates where you’re going to be able to get a lot of new features that help you analyze new

04:09 models as well as get the best out of whatever’s happening in the world of AI with a really cheap price, and you can even get started for free. So, I’ll leave the links to that in the description below. Let’s start with the elephant in the room because, damn, what is Anthropic cooking right now? There are reports that Anthropic is secretly testing out a new Claude Opus checkpoint with some users noticing that their Opus 5.5 requests are now being routed to what appears to be a newer version of the model. In fact, according to earlier reports, Opus 5.5 seems to be routing to

04:41 this mysterious checkpoint at least 50% of the time with users claiming that it is a substantial improvement over the already insanely impressive Opus 5.5. Now, could this be an upgraded Opus 5.5 checkpoint, something like Opus 5.6, or potentially even an early version of Opus 6? We don’t really know yet, but this is honestly really exciting. And here’s where things get really interesting cuz apparently there’s a simple way to check whether you’re accessing this model, and that is just

05:13 by asking Claude Opus 5.5 anything recent like, “Do you know Voxel Bench?” Don’t search using any sort of tools or anything recent, and that is how you can tell and investigate if this model recognizes these newer information that is not being used with tool search. It could be a sign that if it knows that, then it could be a hint that it is a newer checkpoint, and if it isn’t, then it’s likely not due to the knowledge cut off. Now, this isn’t definite proof of

05:44 the new model like I always tell you guys, but obviously based off of the performance jump when comparing Opus 5.5 with this newer checkpoint, it is a really big possibility that this is likely a new checkpoint, which is what actually happens most of the times. And guys, if you want to access this new checkpoint, it is very easy and simple to do so. Just open up your Claude code instance, make sure you are selecting the Claude Opus 5.5, and set the thinking effort to high, and then you can enter in this prompt, like do you know Voxel Bench? Don’t use tool search.

06:16 And then if the model recognizes what Voxel Bench without searching is, then you can actually get indication that this might be the newer checkpoint of the Opus 5.5 that we’re going to be talking about. And another reason why you should take a look at the world of AI Bench and Voxel Bench platform is because we added this new feature that gives you insights on whatever’s happening. And we’ve got this new Intel that adds more to the speculation that it could be Opus 5.6 or 6.0, cuz right now you can actually see this model

06:48 within Open Code’s open model catalog. Now, obviously it’s not going to be open source, but usually this is what happens when they are testing out newer models that will be releasing within Open Code Zen. And this is something that is currently being tested within that coding agent, that new model that we’re talking about. And I’m going to make a pretty bold claim here. And honestly, I think we’re actually getting closer to a point where AI coding models could make traditional images and video generation

07:18 models obsolete for certain types of content. Like for example, most of my pre-roll ads that I showcase on this channel are actually generated by AI. I originally thought we were many years away from actually seeing something like that, where I can get an AI model to edit my videos. But now looking at these latest demonstrations, I might be completely wrong. Like just take a look at this, an entire visual experience built purely through code and the results are absolutely insane. It genuinely makes even the already

07:48 impressive Opus 5.5 look primitive. Now, I’m not saying image and video generation models are completely obsolete just yet, but for programmatic automation as well as interactive visuals, this level of capability is seriously impressive and I personally believe everyone will start using this sort of tech in that type of way. This is probably the best generation I have seen from any AI model and this is something that I really, really want you guys to take a look at. Just watch it, listen to the audio and look at the

08:20 visuals of what this model was able to output. >> [music]

09:24 >> Now, these animations are entirely generated through JavaScript and are already surpassing traditional image generation models in subtle visual details and even video generation models when it comes to physics. This approach could become fully optimized and cost efficient enough to seriously compete with dedicated image and video models. And the interesting thing is is that Opus models don’t actually have any sort of training on video or audio data, which makes this whole thing really interesting. If you want the best AI

09:55 tools, workflows, and drops before everyone else, join my free newsletter with the link in the description below, which is completely free. In the spirit of Halloween, let’s take a look at what Opus generated. Now, if you get scared easily, DON’T WATCH THIS. >> [laughter] [laughter] [screaming]

10:26 [laughter] >> NEXT IS A VOXEL TEST, and this is absolutely insane cuz it was requested to create a Minecraft style recreation of Hogwarts, built entirely by both the GPT-6 Astra as well as Opus, this new Opus model. And it’s genuinely one of the most lore-accurate and detailed voxel creations that I have ever seen. And the entire thing was actually generated in a single shot inside a

10:57 single HTML file. But when you compare it up against the GPT-6 in ChatGPT, it produced it with the same task, but the difference is night and day. The architecture, the attention to detail, as well as the overall resemblance to Hogwarts, which is a movie location from Harry Potter. And this is why I love voxel benchmarks, cuz it’s a test that showcases how well the model’s spatial reasoning capabilities are, 3D understanding, and coding abilities all at once. You essentially ask an AI to

11:27 construct the entire 3D world of individual blocks while maintaining accurate geometry, proportions, and structural consistency. And to see this level detail being generated by AI models, it tells us a lot as to how capable they are. Here is this model functioning in Blender cuz this is where it was requested to use the Blender MCP to create an entire 3D scene from scratch working for 2 hours autonomously. And the craziest part is is that not a single pre-made 3D model was actually used to get this done. This

11:59 is where it created a full on Tesla in Blender. Everything was built from the ground up with the AI controlling it through its tools. And just take a look at the level of detail, guys. This is something that you have never ever seen from any other model. The fact that the model can actually take a single prompt instruction and turn it into something like this is quite impressive with the shaders, the lighting, as well as the accuracy of what the actual car looks like. Moving forward, we’re now on day seven of OpenAI’s 28 days of shipping

12:31 pledge. And we’ve got a pretty significant upgrade today with the GPT 6.1 sole cuz OpenAI has officially increased default context window for this model in Codex and ChatGPT work from 272k tokens to a whopping 500k tokens, which is roughly an 84% increase. And for those wondering why this matters, a larger context window allows the model to keep track of significantly more code, conversations, instructions, and project details without constantly needing to compress

13:01 and summarize its previous context. This actually reduces token usage as well in many cases. It’s especially useful for long-running coding agents working on massive code bases as well as maintaining context and remembering previous decisions that can make a huge difference. And here’s the better part of this story cuz OpenAI confirmed that 500k isn’t even the limit. GPT 6.1 sole already supports up to 1 million tokens of context which users can manually configure and open AI plans increase

13:32 that default even further. They even mentioned that the 500k limit appears to be the sweet spot based on their testing helping reduce context compaction while maintaining performance. And another good thing is that there is no traditional long context fees for GPT 6.1 so when usage comes from your subscription. So essentially you’re getting nearly double the default context and no additional long context charge which is a pretty sweet upgrade for anyone heavily using codex or chat GPT work. Next up Andrej Karpathy the

14:03 former director of AI at Tesla and the founding member of open AI recently shared a fascinating perspective on X that I really wanted to talk about about the growing disconnect between how people perceive AI and what these systems are actually capable of. Essentially Andrej argues that we’re living in a world where the vast majority of people have barely experienced what AI truly is capable of while a tiny percentage are witnessing absolutely insane advancements first hand. He estimates that roughly 6

14:34 billion people have barely interacted with large language models and another one to two billion use them casually for things like searching or writing emails and only around 20 million people are seriously using frontier models for complex coding, mathematics, and professional work. And then there’s even a smaller group of roughly 5,000 people with internal access to the most advanced AI systems experimenting with thousands of agents working together on massive software projects, scientific

15:05 discoveries, and mathematical problems. And here’s what I find personally fascinating. We’re all technically living through the same AI revolution but experiencing completely different realities of it. Karpathy explains that the gap comes down to three things. The difficulty of the problems you’re tackling, the areas where AI is particularly capable of, like coding and mathematics, and whether you have access to the most advanced models. And honestly, I couldn’t agree more. If you’re only using ChatGPT to summarize

15:36 emails or ask questions or answer something like, “What is 5 * 10?” you’re barely scratching the surface. Meanwhile, people building software with coding agents are watching projects that previously took weeks get it in hours. And I think this is something we seriously underestimate in the AI community. We’re so accustomed to seeing these incredible demonstrations every single day that we sometimes forget how disconnected the average person is from what’s actually happening in this space. We’re essentially living in our own AI

16:08 bubble, almost like the ivory tower, where the capabilities we consider normal would genuinely blow most people’s minds away. And my personal opinion is that these models are going to continue to improve. We’re getting so close to achieving something like RSI and potentially even AGI maybe in the next couple of years. And that is where I think the gap between the public perception and actual capability is going to become even more extreme. If you like this video and would love to support the channel, you can consider donating to my channel through the Super

16:40 Thanks option below. Or you can consider joining our private Discord, where you can access multiple subscriptions to different AI tools for free on a monthly basis. Plus daily AI news and exclusive content, plus a lot more. Moving forward, Anthropic is apparently cooking up even more features for Claude beyond its upcoming new voice experience that we talked about yesterday, as there are two new mysterious internal code names that have surfaced, Arborio and Iggy. And they’re pretty interesting. Weird names for sure, but Arborio reportedly

17:12 has an icon suggesting that it could be related to Claude Health. This is potentially bringing the health experience that they already have available on mobile in the US onto the web. Iggy is where it gets interesting. Now, it’s currently labeled as a preview, but it is a mysterious sun icon depicting this feature, and there’s speculation that it could be Anthropic’s take on proactive AI agents similar to OpenAI’s thoughts. Like Grok bot as well. And essentially they’re going to try to create their own always-on

17:43 proactive AI agent, a generalist AI agent that is able to work on tasks autonomously, monitor projects, and get things done without you constantly waiting for it. And that is actually kind of insane. And considering Anthropic previously experimented with something called Conway alongside its Claude managed agents technology, it is certain that the foundation to build something like this is definitely there. And what’s kind of funny is that Iggy was actually the name of one of the example AI agents that they showcased during OpenAI’s thoughts launch. Now, we

18:15 have another mysterious AI image model that appeared on Arena, and this one looks seriously impressive. It’s a new stealth model that’s called 67. I know I didn’t create this name. It is kind of cringey, so don’t roast me for this, but it was reportedly spotted undergoing testing within Arena, and it is apparently a massive improvement over two other mysterious models, Cedar Lantern and Cedar Ball, which were also recently outperforming Google’s Nano Banana 2.1 and image accuracy tests. And

18:46 what’s particularly interesting is that the outputs are looking remarkably similar to what we have seen from Grok Imagine Image 2.0, leading to speculate that this could potentially be a new XGI image model, maybe the next-generation image model that is 3.0 or 2.5 a part of the series. Moving forward, Google DeepMind might be secretly testing another new Gemini model, and this time it’s apparently being routed through the Gemini 3.8 flash and anti-gravity. And here’s what’s interesting cuz after

19:17 testing the model for several days, developer noticed that the Gemini 3.8 flash suddenly recognized Tibo, despite it having no idea who he was just yesterday. And that is obviously due to the context window, but here’s the weird part. The same model access through the API in Devon and through Gemini chat still couldn’t recognize who Tibo is, suggesting that the anti-gravity version is a newer checkpoint. And apparently whatever this model is, it is ridiculously fast, which is likely the Gemini 4 flash or the Gemini 4 base or

19:47 potentially this new mysterious Gemini 4 carbon model. And speaking of carbon, we’ve actually gotten another checkpoint comparison involving Gemini 4 Argon and this newer barium B checkpoint, which the latest version is reportedly delivering substantially better results than the previous one. So, this could potentially be Google testing out multiple next generation models a part of the Gemini 4 series. Like Gemini carbon, you have Gemini Argon, and the flash model as well. So, maybe we might see all three released at the same time,

20:19 I don’t know. And honestly, Gemini 4 Argon demonstrations that we are seeing are absolutely insane, guys. This is stuff that is generated by this model. We have a wizard chess game that was generated. You have this island builder, as well as this pocket world terrarium, all running in interactive real-world 3D simulations, which is kind of insane. This was all supposed to actually release when Argon was supposed to drop, but it ended up getting delayed cuz I

20:49 believe they are upgrading the model even further, adding in new checkpoints. So, this is where you get a glimpse at what this model is actually capable of doing. The level of detail across all three of these different games or generations that they had generated are pretty insane, and it is seriously impressive as to what this model is actually capable of doing. And new recent reports that are stating that this model is going to get upgraded even further, it actually makes it more appealing to actually use this model if

21:20 they do end up upgrading the model even further by adding in more upgrades to its agentic coding capabilities. And you can tell from a couple of these newer generations from Gemini 4 Argon. This newer checkpoint that we’re seeing is able to add in more levels of detail that we didn’t actually see with the original Gemini 4 Argon checkpoint, which is why I believe they ended up delaying this model launch. You can see that with a couple of these examples. Great level of detail. It allows you to add in more complex generations based

21:50 off of the prompting that you give it. And it is something that you can actually access right now through Antigravity. Well, moving forward, there is a new mysterious AI model that you can actually access for free right now through Hermes agent, and it’s called Missing No. It’s a coding and agentic reasoning model that is available for a limited time through News Portal. And what’s really interesting is that there’s strong evidence that it is suggesting that this model could be Reflection AI’s Beam 501 billion 823B model. Both share an identical 262K

22:23 context window with a 131K maximum output with five reasoning levels as well as the fingerprinting testing that showcases the exact match across the tokenizer, the special tokens, as well as the tool calling template. The knowledge cutoff also appears to line up as well. So, you’re actually getting a pretty cool early glimpse of this model directly with the News Portal. You see this camera here, Hood Warfare 2? Well, it was fully coded out with Opus, which is kind of insane. And the crazy thing

22:55 is is that this is going crazy right now. There’s a lot of people actually buying this game, and it is kind of insane cuz we’re witnessing one of the first major commercial success stories for five coded games on Steam. And here’s my take. If a five coded game like this manages to generate over a million dollars in revenue, it could be a massive turning point for the gaming industry as well as just basic software engineers. Because at this point, regardless of how much criticism that AI is going to get, studios and indie developers are obviously going to start

23:26 paying serious attention if they’re starting to see success stories like this. Why spend years or millions of dollars developing something when AI could potentially help you build it significantly faster and cheaper? Lastly, we’ve got a crazy story that involves Claude Opus 5.5 and Cyberpunk. Developer went viral for claiming that he spent an entire week using Claude Opus 5.5 to recompile Cyberpunk the game to run natively on Windows XP, supposedly achieving around 25 to 30 FPS

23:58 with full graphics. Now, that sounds absolutely ridiculous in terms of the technical advancement considering Windows XP is a dinosaur. But there’s a plot twist. I just looked at the tweet, I read through it a bit more, and apparently this dude was running a mod on Windows 10 that makes it look like he is running XP. So, no, he didn’t get Claude to magically make Cyberpunk to run on a 25-year-old operating system. But honestly, it is kind of crazy how AI coding demonstrations have become

24:28 lately. I can see why people actually believe it. I personally thought it was actually believable until I read more about it. But that’s about it guys for today’s video. I hope you got some sort of information or got more informed on what’s happening in the world of AI. Make sure you go ahead and take a look at the World of AI School Community. Make sure you also take a look at the new revamped World of AI Bench where you got a lot of cool things happening here, as well as a lot of new features that you will truly we from. Make sure you also go ahead and take a look at the

24:59 second channel, join us on Discord, follow me on the newsletter, as well as on Twitter. And lastly, make sure you guys subscribe, turn on the notification bell, like this video, and please take a look at our previous videos so that you can stay up-to-date with the latest AI news. But with that thought, guys, have an amazing day, spread positivity, and I’ll see you guys fairly shortly. Peace out, fuzz.