Claude OPUS 6 EAP SPOTTED?! Muse Spark Open, GPT-6.1 Astra, MAI-Base-2, & Fable 5.5 LEAKS! 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 EAP SPOTTED?! Muse Spark Open, GPT-6.1 Astra, MAI-Base-2, & Fable 5.5 LEAKS! AI NEWS · Channel: WorldofAI · 26:02 · Watch on YouTube

Topics: AI agents, Software engineering with AI, AI models, Local and open-source models, Multimodal models

Summary

The video surveys a range of reported AI model releases, leaks, and product updates, led by Meta’s Muse Spark Open, a large multimodal open-weight model that can run locally with powerful hardware. It discusses rumored Anthropic checkpoints potentially linked to Fable 5.5 and Opus 6, alongside updates to Claude Code and Opus fast mode. Other topics include reported GPT-6 Astra and Microsoft MAI-Base-2 developments, free coding-agent models, and new releases from NVIDIA and Black Forest Labs. The video closes with Google Gemini access changes and a Tesla Roadster demonstration that the presenter says uses AI for precise control.

Key points

Meta released Muse Spark Open, described as an 816-billion-parameter multimodal model with 42 billion active parameters and a 1-million-token context window.
Muse Spark Open can reportedly run locally on high-end hardware; the video presents tests involving coding, voxel art, and game generation, with mixed results.
The presenter discusses rumored Anthropic checkpoints named Onigiri EAP and Gabagool EAP, speculating that they may relate to Fable 5.5 and Opus 6.
A newer internal GPT-6 Astra snapshot was reportedly discovered, though the video says details remain limited.
Microsoft is reportedly completing pre-training for MAI-Base-2, described as a 2.8-trillion-parameter multimodal base model; its performance claims are preliminary.
Claude Opus 5.5 fast mode is out of research preview, and Claude Code adds a command to export session history.
DeepSeek 4.1 Flash and NVIDIA Nemotron 3.5 Super VL are described as free to use in OpenCode; Nemotron is presented as a multimodal model for AI agents.
Black Forest Labs launched Flux 3 Image Client, an open-weight image model with native 4K generation and editing features.
Gemini for Argon is rolling out to paid Google AI Pro subscribers in Anti-Gravity with a separate weekly usage quota.
The video reports a Tesla Roadster demonstration involving aerial maneuvers and says AI is used for precise vehicle control.

Chapters

Start	Chapter	Summary
00:00	AI news roundup and upcoming model rumors	The presenter previews Muse Spark Open, rumored GPT-6 Astra and MAI-Base-2 developments, Anthropic checkpoint leaks, and other topics.
03:15	Sponsor message	The video promotes Oxylabs’ web API and related tools for retrieving public web data.
04:21	Anthropic’s rumored Fable and Opus checkpoints	The presenter reviews rumored Claude EAP checkpoints and shows a short film attributed to an early-access model.
08:17	Muse Spark Open: capabilities and tests	The video introduces Meta’s open-weight model, discusses its local hardware requirements, and evaluates its performance in coding and visual tasks.
14:36	GPT-6 Astra and Microsoft MAI-Base-2	The presenter covers a reported GPT-6 Astra snapshot and Microsoft’s claimed progress on its MAI-Base-2 model.
17:22	Claude Code, free models, and OpenAI updates	The video covers Opus fast mode and Claude Code changes, free OpenCode model access, and updates to OpenAI’s shipping pledge and voice mode.
20:38	NVIDIA Nemotron and Flux 3 Image Client	The presenter describes NVIDIA’s multimodal agent model and Black Forest Labs’ open-weight image-generation release.
23:19	Gemini Argon and the Tesla Roadster demonstration	The video discusses Gemini Argon access for paid subscribers before describing the Roadster demonstration and the presenter’s claims about AI control.
25:30	Closing	The presenter thanks viewers, promotes channel resources, and signs off.

Quotes

It’s never a dull day in the AI space — 00:00

Meta has officially released Muse Spark Open — 08:17

there is potentially a newer internal snapshot of GPT-6 Astra — 14:36

Opus 5.5 fast mode out of research preview — 17:22

Blackforest Labs has launched Flux 3 Image Client — 22:15

Concepts and entities

Muse Spark Open, Meta, Claude Opus 6, Fable 5.5, GPT-6.1 Astra, MAI-Base-2, Microsoft, Claude Code, Opus 5.5 fast mode, DeepSeek 4.1 Flash, Nemotron 3.5 SuperVL, NVIDIA, Flux 3 Image Client, Black Forest Labs, Oxylabs, Hugging Face

Full transcript (original language)

00:00 It’s never a dull day in the AI space, as Meta is back with another major model drop, claiming to have released the most capable US open weight model to date. It’s called Muse Spark Open, a massively natively multimodal model with 816 billion total parameters and 42 billion active. And it has a 1 million token context window. This is a model that is built for coding, AI agents, complex software engineering, and long horizon reasoning. But the craziest part is that you can actually run this frontier-class model locally, assuming you have some

00:31 seriously beefy hardware. Now, it seems to be based on Muse Spark 1.2, which isn’t necessarily the strongest model anymore, but regardless, it is a huge step forward for local AI, and we’re going to be testing it out throughout today’s video. But that’s just the beginning of today’s video, cuz we have a potentially newer internal snapshot of GPT-6 Astra, which has reportedly been discovered, and this could be a new model that will be dropping most likely within Open AI’s 28-day shipping pledge. And Microsoft isn’t sitting around,

01:02 either. They’re reportedly finishing pre-training for their next major model called MAI-2, with early claims suggesting that it will outperform nearly every Chinese AI model available. Now, things are getting even crazier within Anthropic. We still don’t know what’s happening with Fable 5.5, as we thought it would release this week. We still have tomorrow, but another early access checkpoint called Onigiri had reportedly surfaced alongside several other series checkpoints. But here’s where things get

01:32 interesting. There’s also a rumored early access model called Claude Gabagool EAP, potentially connecting to Opus 6. So, we have a Fable 5.5 checkpoint and a new Opus 6 checkpoint. And from the demonstration circulating, this thing is absolutely insane at generating 3D visuals, working with tools, and coding in Python alongside Blender. Now, imagine Anthropic releasing Fable 5.5 followed by Opus 6 ahead of its potential IPO. That would completely shake up the AI race and it would honestly pull them ahead over any

02:07 other lab. Speaking of Anthropic, Opus 5.5 fast mode is officially out of research preview with the waitlist removed. We’ve also got several important Claude code updates that we’re going to be talking about. And for those who love free AI models, we’ve got Deep Seek version 4.1 flash from Purdue being free for the week as well as Nvidia’s newly announced Nemotron 3.5 super VL. OpenAI is also continuing its rapid cycle of releases a part of its shipping pledge. And finally, we’re ending things off with something that involves Tesla

02:41 and SpaceX SI where Tesla has reportedly demonstrated its new Roadster, which is literally going airborne. This is something that is soaring over a row of Tesla vehicles using advanced software controls and physical thruster technology, but the exact role of AI is something that is being played throughout all of this, which is why I’m really talking about it throughout today’s video. Also, majority of you guys aren’t subscribed yet, so it would truly mean a lot to me guys if you guys can just subscribe to the channel, like this video as it also helps the algorithm out

03:15 and it also helps us make videos for you guys on a day-to-day basis. But with that thought, let’s just dive into today’s upload. Today’s video is sponsored by Oxylabs, MC Mag’s best proxy service of 2026. Their new web API gives your AI agents real-time data from almost any public website by combining search and scrape in one unified API. Search discovers URLs from the live web and scrape extracts agent-ready data even from complex hard-to-reach sites at scale.

03:47 Try it for free at oxylabs.io/worldofai. Under the hood, you get 175 million plus residential proxies across 195 countries, rotating to handle captchas plus JavaScript rendering with no browsers, servers, or proxies to manage. >> [music] >> And with Oxy Co-pilot, just paste a URL, describe what you need in plain English, and AI builds the scraper, parser, and browser instructions. Plus, it plugs

04:21 into Puppeteer, Playwright, Any end, MCP for Claude and Cursor. Go to Oxylabs’ website in the description below and use my code for $5 off in free credits. No time limit, no credit card required. Just use code World of AI for 20% off Oxylabs credits. Let’s start things off with Anthropic’s next major model, Fable 5.5, which I genuinely thought was supposed to drop this week. Now, we still have tomorrow, but at this point, it

04:51 that I really doubt that will end up dropping tomorrow. However, there are some interesting developments as Fable 5.5 could potentially be hiding under the code name Claude OnlyGiddyEAP. Now, this is another mysterious checkpoint that was found within Artificial Analysis’ front end, featuring four reasoning variants: low, medium, high, and ex high. And interestingly, this model reportedly falls back to Opus 4.8, which is something we’ve seen with other Fable and Opus tier models. But that’s not all

05:23 because several other checkpoints are being spotted, and it is something that is potentially being connected to Opus 5.6 or even Opus 6. And like I mentioned about Artificial Analysis’ front end having all these checkpoints, they actually have a history of using food-related code names for unreleased Claude models. For example, Claude Beniaz EAP was linked to Sonnet 5.5, and while Claude Viren EAP was linked to Haiku 5.5. Now two newer code names, Claude Onigiri EAP, which I believe is sushi, and Claude Gabagool EAP has

05:57 reportedly appeared with several configurations resembling Fable and Opus tier models before being removed from the site today. The speculation right now is that Onigiri could be Fable 5.5 and Gabagool could be potentially Opus 6 or an updated Opus checkpoint. It all falls under Claude Next, which is available for some people in Claude code CLI, as well as Claude web through the silent routing when Opus is selected, which is something that legit API had spotted. But what we have seen from this Opus EAP is just absolutely insane. Just

06:32 watch this short, small little film that it had created. >> Hi. >> [gasps and laughter] >> Who are you? >> Come on. >> Come on. What happens next? >> Tomorrow. >> It’s itchy. >> It’s warm. >> Will you be here? >> Where else? >> Mom, come look. >> How far? >> Far. >> I’m going there. >> Take a scarf. Tomorrow.

07:12 Got everything? >> Mom. Scarf? >> Come on. Come on. >> Don’t understand a word. It’s only me. No reason. >> 3 2 1. >> It’s only me, Mom. No reason. >> I had this one. >> Me, too. Mom, this is

07:43 >> I know. Leave it. I like it. Breathe. It’s itchy. >> It’s warm. >> When? >> Summer. >> Took me ages. Hi, who are you? Oh. What happens next? >> Tomorrow.

08:17 Hi. >> Moving forward, Meta has officially released Muse Spark Open, which they’re calling the most capable US open weight AI model to date. This is massive cuz it’s a natively multimodal model with 816 billion total parameters with 42 billion active and has a 1 million token context window. This is a model like I mentioned built for coding, AI agents, complex reasoning, and long-running software engineering tasks with support for text, images, video, and audio. But here’s the crazy part. You can actually

08:50 run this frontier class model locally by using two Apple M5 Ultra Mac Studios. Now, that obviously has some serious expensive hardware attached to it, but it’s impressive that a model at this scale can run entirely on your desktop. Meta has also showcased this model autonomously making changes to an iOS app and controlling DaVinci Resolve to edit video footage. And it’s already available through platforms like Open Router and a couple of others that you can get started with and the open weights are available through Hugging

09:21 Face. But to me, it’s kind of funny that Muse Spark Open is the first gated open model cuz to download the weights, you have to give your full name and date of birth on Hugging Face. Also, something else I saw is that Muse Spark Open runs quite surprisingly well with llama.cpp. It is working on M3 Ultra Mac Studio with about 512 gigabytes of unified memory. Obviously, that is still pretty beefy in terms of the hardware, but the fact that you can run this type of model on something like that is nice to see. If you want the best AI tools,

09:53 workflows, and drops before everyone else, join my free newsletter with the link in the description below, which is completely free. Now, benchmark-wise, it is doing quite well across a lot of benchmarks like Agentic, knowledge work, as well as even coding. But, obviously, it’s not beating all of the Chinese frontier models like Qwen 3.8, even Mistral. But, in many other cases, it does quite well, which is a really good sign. And remember, models don’t always revolve around coding. It does quite well in other domains like knowledge

10:23 work, which is something that could be useful for a lot of other users. Now, on the world of AI benchmark, after running it across multiple tests, this is where the model is slightly behind the Qwen 27B. This is the Qwen 3.8 27B by just a couple of points. And guys, we’ve launched a new feature, AI news and leaks, where you can get up-to-date leaks on whatever’s happening in the world of AI, as well as a new build race feature that lets you evaluate multiple models across different prompts. The prompts that you see across my videos,

10:56 and you can essentially run these prompts across all of these different models to get a better idea of how it is going to be able to proficiently generate whatever you’re looking to have it generate with the prompts you send in. Not just that, you get access to the leaderboards, you get access to what reasoning levels, what sort of models you should be using, and a lot of other information that helps you out in the AI space. But now, let’s actually take a look at what this model is capable of doing across a couple of the different domains. So far, what I’ve realized with this model is that its front-end taste

11:28 is pretty good. Obviously, not the best, but for a local model, it does pretty decent overall with what it’s able to output. With the scroll triggers, the the design taste that it adds in with its visual effects, and overall, it just does pretty decent with its generation quality. Now, here is a test for Muse Spark Open. It actually came on top in a coding comparison test against three other open-weight models. All four models were tested with building a Vampire Survivors Cells pixel game using HTML canvas and with no external

12:00 libraries or external images. Muse Spark Open delivered the most playable game for about 59 cents in around 20 minutes. Obviously, if you were to locally host it, it would be free. But, this is where it features enemy orders as well as damage numbers and a level-up system with zero crashes. Meanwhile, Inkling was the fastest at just 11 minutes and had roughly the best visuals, but its weapons didn’t actually deal any damage, meaning the functionality of the thing wasn’t actually the best. Each rough four preview took about 52 minutes,

12:32 which is way too long. Whereas, with Step five preview, it crashed during development and needed an additional fix. So, while Muse wasn’t the fastest or cheapest, it produced the best overall playable game in this particular test, which is definitely impressive for an open-weight model. Here is the Voxel Pagoda test. This is a useful way to evaluate how well an AI model is able to handle voxel art with 3D spatial reasoning, coding, and complex visual instructions. Essentially, you’re testing whether the model can translate

13:02 a description into a functional 3D scene while maintaining the accurate geometry, the proportions, and the structural details. It helps reveal whether a model can build multiple interconnected objects together and understand the depth positioning as well as generating working code without any errors. You have seen this test multiple times, and it gives you a clear picture as to what the model is capable of doing. The bigger models tend to have those more creative generated voxel arts, but you can see with this model it did

13:33 successfully generate a five-tier pagoda with the gate on its first attempt, which is definitely impressive, but the result was cropped with the lack of detail. This is a limitation in its visual composition or instruction following, which is probably one of the big issues with the Muse models. And it did take 37 minutes with over 15k tokens to generate running at seven tokens per second across three DGX sparks using a two-bit quantize version. So, this test isn’t just about making pretty pagodas all the time. It’s not about just making

14:06 3D environments. This is a real test that evaluates how effective a model is in terms of reasoning and outputting functional code. Now, in this test by Atomic Chat, this is a really cool one cuz Muse Spark open was tasked to build a voxel city and blow it up with TNT going up head-to-head with the Muse Thunder, which is the model that released this week claiming to be a part of the big three, which it’s not. But, Muse Thunder produced a bigger explosion, which you can see over here, and Muse Spark open did complete the

14:36 task faster at roughly 10 times lower the cost, which is obviously impressive, but not at the same sort of quality. Small little update for OpenAI, but there is potentially a newer internal snapshot of GPT-6 Astra, which has reportedly been discovered. Now, the details are still quite limited, but this is suggesting that OpenAI is preparing another major model release. They had initially scrapped the GPT-6.1 Astra when the new Opus had dropped cuz they knew that this model wasn’t as competent as whatever Anthropic had

15:09 shipped. But, now it looks like they have post-trained the model further, and it looks like it is ready to be released, which means that they are really likely to release this model a part of the 28-day shipping pledge of OpenAI. So, this is something that we will likely see within the next couple of weeks. Next up, we have some massive news from Microsoft cuz they are reportedly completing pre-training for their next frontier model, the MAI base 2, and this thing looks like an absolute beast. We’re talking about a 2.8

15:40 trillion total parameter model with 80 billion active, trained on a massive cluster of 16K Nvidia GB200 GPUs over just 38 days. It’s also multimodal, capable of understanding both text and images. But, this is where things get really interesting cuz Microsoft is claiming 4 to 12 times more efficient pre-training compared to DeepSeek version 4 Pro base, which is pretty significant and it’s an achievement with those numbers, and if it holds up, that is just insane. And according to their early benchmark comparisons, MAI base 2

16:16 is already outperforming nearly every Chinese base model tested, suggesting that Microsoft could be becoming a serious competitor in the frontier AI race. Now, something to also keep in mind is that this is still a pre-trained base model, meaning it hasn’t gone through the full post-training process that makes these models better at following instructions, reasoning, and handling real-world tasks. So, while the early results look extremely promising, we still have to wait for the final model performance against other giants like GPT, Bard, and even now the new Gemini. And what’s particularly

16:50 interesting is that Microsoft built this model entirely in-house, which shows that they’re becoming increasingly capable of developing their own AI system rather than relying on others. If you’d like this video and would love to 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 back to

17:22 Anthropic, we’ve got some important updates to Claude Code and Opus 5.5 fast mode. Starting off with the fast mode cuz it is something that has officially taken Opus 5.5 fast mode out of research preview, which removes the waitlist and it makes it more widely accessible through the Claude platform. This is essentially making Claude Opus 5.5 up to 2.5 times faster for those who do not know, which is pretty significant and it is especially an improvement for developers working on complex coding tasks and AI agents. However, the extra

17:54 speed comes at a premium cost with pricing at $8 per million input tokens and $40 per million output tokens. It is now available through Microsoft Foundry on Azure as well, with Google Cloud support coming soon. And for Claude Code users, fast mode remains unchanged and continues to use extra usage credits. Then, another cool Claude Code feature is the {slash} export file name command, which lets you export your entire session history into a plain text file. This is actually pretty handy because you can feed that session into another

18:28 AI agent to continue where you had left off or analyze your previous workflow to create custom skills and improve how you use Claude Code. It’s a simple feature, but it’s honestly quite useful. Next up, Open Code has now made DeepSeek version 4.1 flash completely free to access within their coding agent, as well as this new Nimatron 3.5 super VL model, which is also free within Open Code. But, it’s kind of crazy that they’ve made DeepSeek version 4.1 flash free for

18:58 you to actually access right now, cuz it is something that also provides a 200K token context window for free a part of this model tier. Moving over to OpenAI, it is It’s of wild that these guys are now on day 11 a part of this shipping pledge and mostly we have seen a lot of new features coming out of this shipping pledge. But honestly, I don’t know how they keep on giving resets out like it’s candy cuz on day 10 they ended up giving a reset to everyone cuz they ended up missing that day for adding a new feature. But now

19:32 we’re on day 11 for their shipping pledge where they’ve rolled out improvements to significantly reduce those frustrating selected model is at capacity errors. They’ve also added more compute and made changes so that even if you encounter that error, your session should automatically resume instead of getting stuck. And interestingly, a developer has also released an open-source project called chat on steroids which reportedly turns chat GPT into a codex-like coding environment while using your chat GPT usage limits

20:05 instead of your codex limits. The developer claims that this could provide practically unlimited GPT 6 usage, although that hasn’t been fully confirmed yet. But a pretty cool project developed. OpenAI has reportedly also rolled out with a major upgrade to voice mode across chat GPT, codex, and dots. The update introduces a new model called GPT live 1.1 which is bringing support for additional voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese. And here’s another important improvement. That annoying checking

20:38 message which would sometimes get stuck during voice interactions and vibe coding sessions should finally be resolved. Overall, this makes voice-powered coding and AI agent workflows much smoother, especially for developers who prefer speaking to their agents rather than constantly typing. And the cool part is is that there’s going to be API support which is going to be coming fairly soon. Moving forward with NVIDIA’s new model that we had mentioned, that is free right now within open code, it is something that is called Nemotron 3.5 SuperVL, and this is a powerful new open

21:11 model designed specifically for AI agents that need to reason, code, and understand visual information. The model features 120 billion total parameters with just 12 billion active and can process text, images, documents, charts, and videos all within a single model. But, this is where I really like this model cuz it is something that can complete 10K coding tasks up to four times faster in comparison to multiple other domains while maintaining strong accuracy on Sweep Bench Pro. It also is designed to handle massive amounts of

21:44 content, maintaining fast output speeds with even over speeds even with 100K input tokens, which is especially useful for long-running AI agents working across large code bases and documents. And what I find particularly interesting is that it combines visual understanding and reasoning into one model, meaning agents don’t have to constantly switch between separate vision and text models. Overall, it’s a really nice update from Nvidia, and I truly love what they have been doing with the Nemotron models.

22:15 Next up, Blackforest Labs has launched Flux 3 Image Client, a new open-weight image generation model that is designed to be fast, customizable, and capable of running locally. It’s a model that features just 5.6 billion parameters and supports a native 4K image generation capability, which is pretty impressive considering that it can run on a single consumer GPU. But, what’s particularly interesting is its pixel-perfect editing and bounding box support, which is going to give you more precise control over image

22:46 generation and editing. It is also fully customizable, allowing developers to fine-tune and post-train the model for their own use cases. And for those interested in running it locally, Blackforest Labs is offering optimized versions in FP8 NVFP4, as well as int8, allowing a low-latency variant for faster generation. The distilled 4K model is also coming to their API, making another exciting release for the Open Weight image generation community. Moving on to Google, Gemini for Argon,

23:19 like you guys had all asked for, is now finally rolling out to paid Google AI Pro subscribers within Anti-Gravity. Previously, access was more limited, but Google now says that they are gradually scaling up capacity to meet the incredible demand for this model. But, the interesting part is that Argon now gets its own separate weekly usage quota for Pro subscribers, meaning it won’t eat up your existing Gemini 3.8 flash allowance. So, you can use Argon for more demanding coding and reasoning tasks, while flash is available for

23:50 everyday workflows. But, to be fair, I don’t even know who actually uses Anti-Gravity right now, cuz Codex and Claude Code are just way better than Anti-Gravity. Lastly, Tesla and SpaceX SSI had actually hosted their next-generation Roadster unveiling today, and this is where this car fully goes airborne. I am not joking. During the live demonstration, the Roadster launched itself over a row of Tesla vehicles, performing aerial maneuvers, and it landed using a system of compressed air

24:24 thrusters developed alongside SpaceX SSI. It was even demonstrated driving upside down using a powerful suction system. But, here’s where things get interesting from an AI perspective. Tesla is already heavily invested in AI powered driving and robotics, and a stunt like this requires incredible precise real-time control to manage the vehicle’s movement, balance, and landing. Now, this is where Tesla has been using AI to control these maneuvers, cuz that is where it was able to perfect all of these systems down as

24:58 well as nailing how it’s able to land in that precise manner, which is kind of cool to see that AI is being used in things like this. Obviously, this is not totally practical, but it’s just insane what this dude is capable of doing in terms of what his team and the company has been able to achieve. But that’s about it, guys, for today’s video on whatever’s happening in the world of AI. I hope you enjoyed today’s video and got some sort of value. It would truly mean a lot to me, guys, if you guys go ahead and like this video and subscribe as well as take a look at the world of AI Insiders School, which

25:30 we’re going to be launching officially very soon. And this is where you get access to a lot of insane perks, which I highly, highly, highly recommend checking out. Make sure you also take a look at the world of AI Benchmark and Vibe Coding Platform as well as the second channel. Make sure you go ahead and take a look at the newsletter, join us on the Discord, follow me on Twitter, and lastly, make sure you guys subscribe, turn on notification bell, like this video, and please take a look at our previous videos that you can stay up-to-date with whatever’s happening in the world of AI. But with that thought, guys, have an amazing day. Spread positivity, and I’ll see you guys fairly

26:02 shortly. Peace out, fuzz.