Claude Fable 5.5 Is BABY AGI?! NEW EAP LEAKS, GLM 5.5, GPT-6 Astra Pro Upgrade & 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 Fable 5.5 Is BABY AGI?! NEW EAP LEAKS, GLM 5.5, GPT-6 Astra Pro Upgrade & More! AI NEWS · Channel: WorldofAI · 21:21 · Watch on YouTube

Topics: AI agents, AI developer tools, AI content generation, AI models, Robotics and autonomous systems

Summary

The video covers reported leaks and demonstrations involving Claude Fable 5.5, which the host describes as a possible early step toward more capable AI systems, while emphasizing that it is not AGI. It also discusses rumored GLM 5.5 and Kimi releases, Claude performance improvements and quantum-circuit research, and OpenAI’s updates to GPT-6 Astra Pro and agent workflows. Other topics include expanded Gemini access, a free GLM 5.3 Flash offering, and a reported large industrial humanoid robot unveiled in China. Many of the model specifications, capabilities, and release timelines mentioned are presented as rumors or claims that have not been confirmed.

Key points

The host highlights reported Fable 5.5 demonstrations that generate detailed, interactive environments from prompts, while clarifying that this does not prove AGI.
Several mysterious Claude early-access model names have reportedly appeared in internal listings and a CLI binary, prompting speculation about a near-term Fable 5.5 release.
A reported Minecraft-style game demo included environmental interactions, physics-like behavior, and sound, generated as a single HTML file.
GLM 5.5 is rumored to have more than one trillion parameters and a one-million-token context window, with claims that it could outperform leading models; these claims remain unconfirmed.
Anthropic improved response startup times across Claude products, including Claude Code, and added conversation-history compression.
Claude Mythos 5 reportedly produced more efficient quantum-circuit designs in several tests and worked autonomously with an external evaluator.
OpenAI upgraded GPT-6 Astra Pro with faster responses, improved tool use, and parallel agents for Pro users.
Google AI Pro subscribers reportedly gained access to Gemini 4 Argon through AI Studio’s build mode and playground.
GLM 5.3 Flash is described as free within OpenCode, while a separate free stealth model reportedly performed poorly in coding tests.
A Chinese robotics company reportedly unveiled Neo 1, a 13-foot humanoid robot designed for heavy industrial work.

Chapters

Start	Chapter	Summary
00:00	Fable 5.5, model leaks, and the day’s AI news	The host introduces reported Fable 5.5 world-building demos, possible early-access leaks, other model rumors, and a large humanoid robot.
04:08	Claude Fable 5.5 demonstrations	The video reviews rumored early-access model names and examples of Claude generating an interactive game and recreating a 3D video from a reference.
09:23	GLM 5.5 rumors and Claude performance updates	The host discusses unconfirmed GLM 5.5 specifications and claims, then outlines faster startup times across Claude products.
11:59	Claude Slides and quantum-circuit research	The segment covers Claude’s slide-generation feature and reported Mythos 5 results designing more efficient quantum circuits.
14:02	OpenAI agent and presentation updates	The video describes improvements to context handling, parallel agents in GPT-6 Astra Pro, and slide support in ChatGPT Space.
16:36	Gemini access and free model reports	The host discusses Devin’s reported free access promotion, Gemini 4 Argon availability in AI Studio, and an underperforming stealth model.
19:13	Free GLM 5.3 Flash and a giant humanoid robot	The closing news covers GLM 5.3 Flash in OpenCode and a reported 13-foot industrial humanoid robot from China.

Quotes

I honestly believe Claude Fable 5.5 could be a step towards AGI. — 00:00

The blocks float, you have sand that actually falls with real physics — 05:41

Claude code local sessions now start responding three times faster — 10:57

multiple agents can now work on different parts of a task simultaneously. — 15:03

physical strength doesn’t necessarily mean advanced AI — 20:19

Concepts and entities

Claude Fable 5.5, Boxel Bench, Anthropic, GLM 5.5, Claude Code, Claude Mythos 5, GPT-6 Astra Pro, OpenAI, Gemini, Google AI Studio, Neura Robotics, Neo 1

Full transcript (original language)

00:00 This might be a bold claim, but I honestly believe Claude Fable 5.5 could be a step towards AGI. Not literally AGI, but an early step towards models that can go beyond smart auto complete and actually generate and simulate interactive worlds. Just take a look at what Fable reportedly created on Boxel Bench. From a single prompt to build out a Boxel art, and this is something that is just mind-blowing. Compared to Opus 5.5, which does look clean but

00:30 relatively flat and static, Fable’s version feels like a consistent 3D world that you can actually explore with impressive lighting, interconnected environments, as well as spatial relationships. While this doesn’t prove AGI, obviously, it’s a fascinating demonstration as to how far spatial reasoning [music] has come. And what’s even more exciting is that we’ve got new Fable 5.5 leaks suggesting a release could be extremely close. Two mysterious Claude EAP models are showing up more

01:00 and more. We have the ones that we had talked about yesterday, but we have a few others that are now inside Devin’s CLI binary. On top of that, leakers claim that Anthropic’s partners are recently receiving access to Fable 5.5, with speculations pointing towards a possible release next week. Meanwhile, GLM 5.5 is rumored to outperform both Fable 5.1 and GPT-6 Soul, which is an incredible bold claim considering how capable these models already are. And it’s reportedly a massive trillion

01:32 parameter model expected to launch very soon. Although these specifications and performance claims are still kind of unconfirmed, so you take that with a grain of salt. Anthropic also rolled out a huge Claude code performance update, significantly reducing response startup times. Claude code now compresses conversation history using gzip while Claude sessions begin spinning up. These might just be small optimizations, but for developers constantly working with coding agents like Claude code, you will definitely notice the difference as you use it

02:03 more. And speaking of Anthropic, Claude Mythos 5 is reportedly designing quantum circuits that outperform some of the best published designs by 10x or more in certain comparisons. That is an incredible achievement if validated, and it shows how AI is increasingly contributing to many different fields like scientific discoveries, even math like we had seen with GPT Bell. And if Mythos 5 is already capable of results like this, imagine what future models like Babel 5.5 could potentially

02:34 accomplish. Over at OpenAI, GPT-6 Astra Pro has received another significant upgrade as a part of its 28-day shipping pledge. We’re seeing faster performance, improved plugin and tool usage, and more importantly, the ability for multiple agents to tackle different parts of a task in parallel. This is now available to pro users, but I’m curious whether business accounts will also get access to and whether GPT-6.1 Astra Pro could be coming soon. Google is also expanding

03:06 Gemini for Argon access to AI Pro subscribers through AI Studio’s build experience and playground. This is obviously great news for anyone who has a pro subscription, which is going to be able to get more usage out of it. And there’s still plenty more to cover including two mysterious stealth models, now GLM 5.3 reportedly becoming completely free inside open code, and an absolutely massive development coming out of China’s robotic industry. Where this Chinese robotics company has actually unveiled an enormous 13-ft

03:38 tall, 2,400-lb humanoid robot capable of lifting over 1,100 lbs. This week, my dreams were truly fulfilled, my childhood dreams. I got to see a flying car from Tesla, and now we have this full-on Mecca robot that is going to be developed that uses AI to do various sorts of things. And before we even dive into anything, it would truly mean a lot to me, guys, if you guys can go ahead and give this video a like, as well as subscribing, as it truly helps me out, and it allows me

04:08 to make videos for you guys on a day-to-day basis. So, it would really mean a lot to me if you guys can go ahead and do that, as a majority of you guys aren’t actually subscribed. But, with that thought, guys, let’s go ahead and dive into today’s upload. Starting off with Claude Fable 5.5, where things are getting really interesting once again, because we’ve got even more clues suggesting that Anthropic could be preparing its release this upcoming week. First, Fable 5.5 reportedly appeared within Microsoft’s internal model listing, which already sparked speculation that an upcoming release

04:39 would come in this current week, but, as we all know, it didn’t end up dropping. But, now, two mysterious Claude early access models have been spotted inside Devinci light binary that is called Claude Orzada EAP and Claude Vanilla EAP. But, that’s not all, cuz we have two additional names that we had talked about yesterday called Varen and Onigiri, which were reportedly added around a week ago. Now, this doesn’t necessarily confirm that we’re going to see four separate models, but it does

05:10 add to the fact that they are testing multiple checkpoints under Claude Next, which you can access within Claude code right now. And like I had titled this video, this is a model that I see as a step towards AGI. In my case, has named it baby AGI. Obviously, it’s not actual AGI, but you get what I mean. We’re reaching a point where you can give an AI a relatively simple prompt, and instead of generating basic code, it builds an entire interactive experience with details you

05:41 never even asked for. Just like this case over here, where Claude Next, which is routing all these EAPs, reportedly ran on medium effort. Medium effort, not the highest effort. And this is where the prompt was to simply create a Minecraft game, the sandbox game, and create a clone of it, but in a Japanese style. All written in a single HTML file. And the results are absolutely insane. The blocks float, you have sand that actually falls with real physics,

06:13 you have torches that break underwater, trees fall, and there’s even beautiful background music with wind sounds near the trees. But what genuinely shocked me was that the dolphin that you saw at the start of the video suddenly started performing tricks towards the end. Now, think about that. The prompt didn’t explicitly ask for every one of these details, yet the model reportedly managed to put together an entire interactive world with environmental interactions, physics-like behaviors, and sound design. And yes, all of this is running inside a single HTML file.

06:45 Now, this doesn’t mean that the model truly understands the world like a human does, but the ability to take a broad idea and turn it into something that’s this detailed, it is coherent and interactive, and it is making this whole experience something that feels otherworldly, is just genuinely incredible. And I don’t even know even if Astra 7 could even pull something like this off. Also guys, if you guys haven’t checked out the world of AI benchmark tool, I highly recommend you do so cuz this is

07:16 where we have launched this new thing where we’re going to be releasing AI news every single day of the hour, where we’re going to be releasing leaks as well as whatever’s happening in the world of AI. And we have also added in a lot of cool things like build your own app, build race, comparing different models, exporting model comparison videos, and a lot of other things. So, I highly recommend checking it out so that it helps you get started with AI as well as getting started with which models to actually use. If you want the best AI tools, workflows, and drops before

07:47 everyone else, join my free newsletter with the link in the description below, which is completely free. And what’s even crazier is that we’re reaching a point where you can literally just sit in your chair, type out what you want, and Fable 5.5, in this case, could potentially build out some insane generations for you. Like, in this case, someone had recreated the entire Tesla Roadster launch video using pure Python code inside Blender. This is alongside 11 Labs for the audio, but the craziest part is that the reported clip was just

08:18 sent into the cloud code session as a reference, and the model was able to code out the entire recreation based off of telling it to do the so. But, just think about how much work something like this would traditionally require, cuz you need to have experience with 3D modeling, animations, lighting, camera movements, and programming. Yet, here we’re seeing all of that brought together with all these elements through code. This is exactly why I keep saying that we’re truly entering a different

08:48 era of AI assisted creation, cuz the barrier between having an idea and having it actually come to life is getting smaller and smaller day by day with these newer models. Also, just watch this.

09:23 Next up is that we already know that the Chinese AI race is heating up, but things could get really interesting because GLM 5.5 and even Kimi Chat 3.1 are both rumored to be getting closer to launching. Starting off with GLM 5.5, several sources are claiming that this model could potentially outperform the Claude fable 5 model as well as GPT-6 model, which is an absolute bold claim considering how capable these models already are. But, what’s even crazier is the rumored specification of the GLM

09:55 5.5. We’re reportedly looking at a model that is over 1 trillion total parameters with the 1 million token context window. And, there’s even speculation that GLM 5.5 flash is a variant that has already surfaced inside open code. But, from what I have seen with different catalogs, I have started to notice that there are strings pointing towards GLM 5.5 and Kimi K 3.1 releasing most likely within October if everything goes smoothly. Moving forward, Anthropic has

10:25 rolled out another major performance upgrade across Claude. And, this time it is all about making the model respond significantly faster cuz just yesterday Anthropic did release Opus 5.5 fast mode, which was made generally available. And, it allows users to run the model up to 2.5 times faster the speed for roughly twice the cost. But, now they’re improving the baseline response speeds across their products without requiring users to switch to fast mode. Claude code now compresses conversation history. And, this is using

10:57 gzip before sending it to the model. While Claude sessions will begin spinning up as soon as you start typing. And, the improvements are pretty impressive cuz Claude code local sessions now start responding three times faster, dropping from about 4 seconds to just 1.33 seconds. Claude sessions are about 2.8 times faster, going from nearly 17 seconds to 6 seconds. Meanwhile, regular Claude chat and co-work are about 1.5 x faster. And,

11:28 even the Claude API has seen a 1.5 times improvement in time to first token. Now, these are baseline improvements primarily reducing the time spent waiting before generations begin, rather than making the model itself generating tokens three times faster. But overall, it’s definitely a great update. It is something that is going to make using Claude, as well as all of its suite of tools, a lot easier cuz it’s just going to be more efficient overall. Also, this is something that I heard

11:59 about for the first time, but it was something that was already launched long time ago. This is where Claude Slide is a feature in your Claude artifacts, and it’s honestly pretty impressive cuz you can use Claude to create an entire slide deck, including custom visuals, layouts, and even things like iPhone mockups, without manually searching for assets online. It’s another great example of how AI is going to be used in these sort of cases where you can practically visualize whatever you’re trying to create.

12:29 And speaking of Anthropic, seriously, how do they keep doing this? I know their IPO is soon, and I know their valuation dipped a little bit, so maybe they’re squabbling a little bit to release all these little updates to hype it up even more again. I don’t know what they’re doing, but they have launched another fascinating research, and this shows how Claude Mythos 5 can design quantum circuits that are significantly more efficient than some of the best previously published designs. And the results are absolutely insane cuz Mythos

12:59 5 reportedly outperformed published designs in seven out of eight tests, which made one circuit 39 times more efficient, and reduced another circuit’s gate count by just 17x. In one experiment, the model worked autonomously for 53 hours, starting from a single paragraph of plain English instructions. What’s fascinating is how it achieved this, cuz Claude would write a circuit, test it using an external evaluator, receive feedback, and continuously improve its design without

13:30 human intervention between attempts. And this actually matters because more efficient quantum circuits could reduce the computing resources needed for applications like chemistry, material sciences, as well as optimization. And think about that, we’re reaching a point where AI isn’t just helping us write code, but potentially discovering better solutions to complex scientific problems. And if Mythos 5 can already achieve results like this, imagine what Fable 5.5 is going to be capable of doing. If you like this video and would love to

14:02 support the channel, you can consider donating to my channel through the Super 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 over to OpenAI, as a part of their 28-day shipping pledge, we’re now on day 12, and the company has introduced some pretty useful improvements to Codex and ChatGPT work. Essentially, OpenAI has optimized how these tools handle context

14:33 and requests, which allows users to get more effective usage out of their existing limits. They have achieved this by reusing context across multiple sub-agents, which keeps context cached for longer and improves request routing to increase cache hits. This means less unnecessary processing, as well as more efficient agent workflows, especially when you’re handling complex tasks involving multiple agents. And while this isn’t a flashy new model release, it’s definitely those updates that make Claude code

15:03 Codex, sorry, a bit easier to use. And that’s not all cuz OpenAI has upgraded the GPT-6 Astra Pearl, which is receiving this pretty significant upgrade inside ChatGPT with faster response times, as well as improved support for plugins and tools. But the biggest addition is parallel agents, meaning multiple agents can now work on different parts of a task simultaneously. You can even track each agent’s progress and open completed tasks to review their work, similar to

15:34 what we had just saw within Claude code. This is going to make it a massive improvement for complex researching as well as coding tasks and multi-step workflows, especially when you have several tasks running at once. It’s an upgrade that is available for pro users directly in the chat tab, and I’m curiously wondering whether OpenAI plans to bring this to other subscription tiers soon. And could we actually see a GPT-6.1 Astra Pro soon? Another update from OpenAI is ChatGPT Space, which the team had actually launched on Dev Day.

16:06 But, a pretty cool actual thing that they ended up releasing is slide support, which allows users to work on presentations directly within the experience, and this is a feature coming to pro, business, and enterprise users with OpenAI gathering feedback with this initial launch, but I think it is solid cuz this way you get to create these beautiful presentations directly inside their experience. This is nothing too crazy, but you can add your dot as a widget inside your iOS home screen. This is something that will

16:36 let you jump into conversations or start a call just from a tap on your icon, which is going to start up your dot. Now, this is a little random, but this is something that I found on X that I really wanted to emphasize cuz this is where Jazzy had emphasized, why would you pay for expensive AI APIs when you can access the models approaching Kimi or GLM level capabilities for practically free. This is where Devin has reportedly extended free sway to access until the end of the month, and

17:07 it is going to allow you to get access to the model within this promotion period. Now, I’m not paid by them, nor is this guy telling me to actually talk about it, but you get access to a lot of these great models for discounted prices, especially if you are a power user, cuz you get access to Gemini 4 Argon, which is about 75% off, and you can access these models through either the CLI or the desktop app. And speaking of Gemini 4 Argon, Google has rolled out with another nice update for Google AI Pro

17:39 subscribers, cuz you can now access Gemini 4 Argon directly inside Google AI Studio’s build mode and playground. And it’s going to allow you to experiment with the model and build entire applications using the build mode that Google previously had. And it’s kind of pretty decent. It is also going to get you some good usage out of your pro tier. So, this is something that you should 100% make the most out of if you have the pro subscription. Now, a new mysterious stealth model called double L alpha has appeared

18:10 within open code. And initially, people were definitely excited about this model because it was available for free, and it reportedly had a reasoning style similar to Claude. Now, some even speculated that it could be an upcoming GLM model, but after the couple of tests that were ran with this model, it is looking like it is an under-developed model. One developer tested out on two coding tasks, and the results were honestly hilarious, cuz the model spent around 40 minutes thinking and another hour working only to produce an HTML file that displayed nothing. And when it

18:42 was asked to fix it, the model insisted its code was correct and blamed the developer for opening the wrong file. Even after being asked to double-check, it kept insisting that it was right. That’s nearly 2 hours wasted on a model that refused to acknowledge its own mistake, which is why a couple of people are saying that you should just avoid using this model at the current moment. Regardless, it’s free, so if you want to use it, you can do so, but there are better options like GLM 5.3 flash, which is completely free within open code.

19:13 It’s kind of crazy. There’s so many free models that have been deployed, maybe hinting at more models coming sooner or later, or maybe it’s being used to train, I don’t know too much, but the GLM 5.3 flash is now fully free for you to actually use within open code. And lastly, we’ve got something absolutely insane coming out of China, where robotics is actually becoming quite interesting again. This is where a company called Neura Robotics has reportedly unveiled Neo 1, a massive

19:46 13-ft tall humanoid mecha robot that is designed for heavy industrial work. This thing looks like it came out of Transformers. And this is why at the start of the video I said, “My dreams are coming true, my childhood dreams.” where I had once wished for flying cars, which we did see actually with the Tesla Roadster, and now we have full-on robots coming to life. The robots that you would see within Transformers. And this is a robot that is reportedly weighing at around 2,400 lb and can lift over

20:19 1,100 lb, which is more than half a ton, which is nuts. And what’s fascinating is that we’re seeing robotics move beyond just smaller robots. And this is something that is going to be used within massive mechanisms. And it is designed for industrial applications. Now, physical strength doesn’t necessarily mean advanced AI, but imagine combining hardware like this with increasingly capable vision, reasoning, and an autonomous control that is powered by AI. But that’s about it, guys, for today’s

20:50 video. I hope you got some sort of value. I’ll leave all the links that I use in today’s video in the description below. Make sure you go ahead and take a look at the World of AI School community, which is going to get you a lot of benefits and essentially help you get started in the world of AI. Make sure you also take a look at the World of AI Bench. Take a look at the second channel. Make sure you go ahead and join the newsletter, Discord, follow me 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 whatever is happening in the world of AI. With that

21:21 thought guys, have amazing day, spread positivity, and I’ll see you guys really shortly. Peace out fams.