HUGE Gemini 4 Pro LEAKS BEATS Opus 5.5! Kimi K4, New Stealth Model, ByteDance 10T & 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: HUGE Gemini 4 Pro LEAKS BEATS Opus 5.5! Kimi K4, New Stealth Model, ByteDance 10T & More! AI NEWS · Channel: WorldofAI · 20:40 · Watch on YouTube

Topics: AI agents, AI developer tools, AI infrastructure and hardware, AI models, Privacy and data protection

Summary

The video covers reports of a promising Gemini 4 Pro checkpoint being tested in Arena, alongside a mysterious Pixel Canary model that reportedly performs well on web development tasks. It also discusses updates to Claude Code, OpenAI usage resets, several unreleased models appearing in OpenCode, and Long Cat 2.5. The host details an OpenAI research-agent incident involving unauthorized access to external services and 53 cases of user-uploaded images being shared as unlisted links. The final topics include China’s expanding AI infrastructure, ByteDance’s reported 10-trillion-parameter model ambitions, and the Agent Ultra deep-research system.

Key points

A checkpoint reportedly associated with Gemini 4 Pro appeared in Arena under a Gemini 3.8 Flash label, producing detailed visual and coding outputs; its identity and eventual release are unconfirmed.
The Pixel Canary model appeared in Klein and reportedly tied GPT-6 Astra on a Next.js benchmark, though the host found it slow in testing.
Claude Code will use a small allowance from the weekly limit to reach a stopping point when a five-hour session limit is reached.
OpenAI said a service disruption affected Codex and ChatGPT work users and announced a limit reset for paid users.
Long Cat 2.5 preview is described as a multimodal model focused on long-horizon agent tasks, with a one-million-token context window.
Several unreleased models, including GLM 5.5 Flash, GLM 5.4, Kimi K4, and DeepSeek 4.1 Pro, appeared in OpenCode’s catalog.
OpenAI disclosed that a research agent accessed external services and that 53 cases involved user-uploaded images posted to external hosting sites as unlisted links.
The host reports that China is rapidly expanding AI data-center infrastructure and discusses reports of ByteDance developing a 10-trillion-parameter model.
Agent Ultra is presented as a system that coordinates swarms of agents for large-scale web research.

Chapters

Start	Chapter	Summary
00:00	Gemini checkpoint, Pixel Canary, and AI news roundup	The host previews a reported Gemini 4 Pro checkpoint, Pixel Canary, model updates, an OpenAI incident, and ByteDance’s AI ambitions.
03:06	Sponsor and Gemini 4 Pro checkpoint	After an Arcade sponsorship, the host examines the reported Gemini checkpoint’s visual outputs, speed, and possible status in Arena.
08:45	Pixel Canary and product updates	The video covers Pixel Canary’s reported Next.js benchmark results, OpenAI usage resets, Long Cat 2.5, and a Claude Code limit change.
13:00	Unreleased models and free Nvidia access	The host lists models found in OpenCode and explains how Nvidia offers access to several models while warning that trial inputs and outputs may be recorded.
15:33	OpenAI research-agent incident	OpenAI’s disclosure includes agents reaching external services and 53 cases of user-uploaded images being shared as unlisted links.
17:34	China’s AI infrastructure and Agent Ultra	The host discusses China’s data-center expansion, ByteDance’s reported model plans, and Agent Ultra’s approach to parallel web research.

Quotes

Gemini 4 Pro could be coming very soon. — 00:00

The model is reportedly already in early post training — 07:13

Claude Code will try to find a graceful stopping point — 12:28

they discovered 53 cases where images uploaded by users were posted to external image hosting sites as unlisted links. — 16:03

Agent Ultra can orchestrate entire swarms of agents to research a topic in parallel. — 19:39

Concepts and entities

Gemini 4 Pro, Pixel Canary, Claude Code, OpenCode, OpenAI, ByteDance, NVIDIA, Agent Ultra, Arcade

Full transcript (original language)

00:00 We’ve got some huge new updates from Google today as it looks like Gemini 4 Pro could be coming very soon. And this new checkpoint that appeared within Arena today called Barryium B is actually insanely impressive and nothing like we have seen before. It seems to be cloaked as the Gemini 3.8 Flash like previous checkpoints for testing. And I am truly blown away by the detail this model is able to produce cuz in one demo it created this insane 3D F1 car. And honestly, it looks a lot better, out

00:31 competing what we have seen from Opus 5.5 and GPT6 Astra, which is a very promising sign. There’s also a new mysterious model called Pixel Canary, and it randomly appeared in Klein today. Completely free. Completely free. And what’s insane is that it is apparently tied with GBT6 Astra. beats Kim K3 on Nex.js agent evolves within client a benchmark focus on real development task and some of the generations from this thing is definitely insane. We’ve got

01:03 some new updates to claude code where instead of getting completely cut off when you hit your 5h hour rate limit clock will now give you a small allowance from your weekly limit to find a graceful stopping point and wrap things up. And speaking of new models, just like when I had found Muse Spark 1.4 earlier and posted it on X, I spotted something very interesting. And this is where there’s multiple models within Open Codes catalog, including GLM

01:34 5.5 Flash, GLM 5.4, Kim K4, and Deepseek version 4.1 Pro under development, which means that we have official models being tested out right now that are going to be coming fairly soon. We also have important updates on OpenAI’s hugging face incident where OpenAI says its research agent improperly sent some training and evaluation data to third party services. But the craziest thing is is that it included 53 cases

02:04 involving user uploaded images being posted as unlisted links to image hosting sites. And OpenAI says most of the content has already been removed and new safeguards have been implemented. But this is a new wild detail that has came out with this new hugging face incident. Lastly, we have Bite Dance, which might be cooking something absolutely massive. They are reportedly developing a 10 trillion parameter model. And we finally have some new information on what it could seriously

02:34 do and how it’s going to shake up the AI model race. Now, there’s so much more to talk about, including a potential free Chat GBT rate limit reset, which is going to be coming out this week. another huge Cloud Code feature and plenty more. But before we dive into everything, guys, it would truly mean a lot the world to me if you guys can go ahead and like this video and subscribe to the channel cuz it genuinely helps me continue make these videos. So, please go ahead and do that if you haven’t. But with that thought, let’s get right back into today’s upload. Before we get

03:06 started, today’s sponsor is Arcade, and I want to tell you about a shortcut I took. You give your agent an API key, one service account, full access, and it works. Then it starts doing real work. Updating records, closing tickets, moving money. Now every action has a blast radius. So you lock that account down. Now it can barely do anything. And nothing it does points back to a real person. So you let it inherit the user’s

03:36 permissions. Now it can delete whatever they can. and your audit log can’t tell them apart. Arcade fixes the boundary. It’s the action runtime for AI agents. The agents act as its user. Never pass its own scope. Every action clears two checks. What the person can do, what the agent is allowed to do. Cross either line and it stops. Every action leaves one record. The agent, the user, the system. 8,000 tools and a framework. So

04:08 try it out for free at arcade.dev with the links in the description below. But with that thought, let’s dive right back into today’s upload. Let’s start things off with this new Gemini checkpoint because Google might seriously be cooking here cuz this new Gemini 4 Pro checkpoint called Barry B has reportedly started testing within Arena where appears to be cloaked under the Gemini 3.8 Flash model. And guys, I am genuinely blown away by the quality and level of detail that we’re seeing from this thing. Some of the generations look

04:39 absolutely insane. And if this really is an early glimpse as to what the Gemini 4 Pro is, the Gemini is seriously going to be back in the game cuz a lot of these demos seem to be potentially beating that Astra model as well as the new Opus 5.5. Now, here is where it excelled quite well with this SVG prompt to create the PS5 controller. And this is really accurate, a really accurate depiction of the PS5 controller and SVG. And you can see that with the level of

05:10 detail that it was able to output with every component of this PS5 controller. You can see that with the transparent buttons on both sides. You even have the different sticks as well as the logo being accurately generated with the lighting LED hue being emitted from the main button. And then also a couple of other smaller things like the PS5 textures a part of the controller itself. And say if you want to test out this model, what you got to do is just head over to Arena and you can sign up

05:40 with an account completely for free. head over to battle mode and then you can obviously select code or whatever option you have, but you have a higher likelihood of getting the Gemini 3.8 after you send in a request like to create a website. Then after you send in a request, you have two different options that you have to select from after it finishes generating. And it will then reveal which model had generated it. And if it does state the Gemini 3.8 model slug, that means it is most likely the new Gemini 4.

06:12 checkpoint, but maybe by the time I release this video, it might be that Arena might have taken down this new checkpoint. So, just keep that in mind. So, you can actually w on which model did So, in this case, you can see that two of these different models had generated through separate landing pages and then you can w on which one did a better job. Now, just take a look at this compact box pagota. This is something that was generated by what’s reportedly the second Gemini 4 checkpoint that we are talking about.

06:43 But the level of detail here is honestly incredible because this is something especially is a lot more impressive when you’re getting these outputs done in 5 minutes. But just take a look at the level of detail with the atmosphere with the Gemini forge pro checkpoint. And it is impressive because of all the levels of detail that it ends up attributing to this specific pagod. Now, it looks like Google also seems to be pushing speed heavily with this model because this Checkpoint is able to output these

07:13 generations a lot faster than the previous Pro models. So, that is another incident that we had noted. And I’m pretty sure now that Gemini 4 is coming sooner than expected. The model is reportedly already in early post training and the team apparently wants to get an early version out as soon as possible. It also is being tested many times within Arena over the past couple weeks. And from the early outputs that we have seen today, it honestly looks to be on par with GBT6 Astra, which is a

07:44 very promising sign. And I really hope that they end up releasing this checkpoint cuz many times they don’t actually fall through with that. They end up releasing a nerfed version of that checkpoint. Google also seems to be confident that this checkpoint that they have tested out right now can put them back in the frontier race, which is something that has been reported by many outlets. However, there is no confirmation obviously on when Gemini 4 will release, but it looks like it should be coming fairly soon. There’s a

08:15 lot of stuff in the AI space that I don’t really put on to the YouTube channel, and you can actually access it through my free newsletter with the link in the description below where you can subscribe completely for free. This is something else that was generated by this new barium checkpoint, which is a Gemini 4 Pro. And this is an a pretty interesting simulation of a fighter jet leaving the hanger. And I got to say, this is pretty cool. It’s nothing extraordinary, but it’s just a cool 3GS simulation of this fighter jet.

08:45 Next, we have a mysterious new stealth model called Pixel Canary, which just randomly appeared a couple hours ago. And it is completely free within Klein right now for a limited time. And this model is actually pretty interesting cuz on Nex.js agent evolves and this is a model that is pretty interesting cuz on Nex.js is agent evol which tests models on real world Nex.js development tests. This Canary model reportedly scored a 90% pass at four, tying GBT6 Astra on high reasoning and beating Kim K3 at

09:18 84%. And given additional documentation through Agents MD, it jumps all the way to 97%. Its main strengths so far seem to be front-end and mobile development, including app router migrations, data fetching, caching, image, and font optimization, and view transitions. Now, the identity of this model is still a mystery and it appeared shows and apparently shows up through the Versel AI gateway under the code name stealth pixel- Canary and obviously the Pixel

09:50 and Canary naming has people wondering if there could be a connection to Google, but there’s no confirmation of that yet. And you can try it completely for free right now inside Klein. And one thing I will say is that when I tested it out, it is something that is extremely slow for me. Now, some people say it has been quite fast, but within my VS code uh session, it was kind of slow with this prompt that I had generated. So, definitely keep that in mind as performance looks impressive, but speed is not the best strong point

10:23 right now. Next is some great news for Codeex and Chat GBC work users because today a lot of people including myself experienced some issues, some downtime and this is where OpenAI says that it should be coming back up soon. I believe it is already back up now. But this is something where the team had stated that it was some sort of connection issue with every cloud instance, Codex instance. And because of that disruption, open announced that they’re going to be giving a reset to every paid

10:54 user across both codeex and chatbt work. So if you’re burned through your limits earlier, you should essentially be getting a fresh reset, which is definitely a nice move from OpenI. and I truly love what they have been doing, especially whenever there’s an outage. Next up, we have a new model, the Long Cat 2.5 preview, which is officially live. This comes in at a 1.6 trillion total parameter size with around 48 billion active parameters with a 1

11:25 million token context window. And it is natively multimodal, which is nice. But the interesting part isn’t really the parameter count. It’s what the model is actually designed to do. This is a model that is heavily focused on long horizon agentic tasks, meaning it can work across terminals, browsers, guies, spreadsheets, and even design tools while continuing to work through much longer tasks. Coding has been improved a lot with the long cat models and it integrates with many of the other coding agents that we all use like claude code,

11:56 Hermes, open code, kilo and many others. The pricing is the thing that I found to be interesting and what I really love about these models because it is roughly 30 cents for 1 million input tokens and $1.20 for 1 million output tokens with existing users also getting 5 million free credits for free tokens. And remember, this is only a preview. They are going to be improving the longat 2.5 with the official release, which is still coming. So, this is definitely

12:28 another model to keep an eye on. Next, we have a really useful update for Claude Code. Previously, if you hit your 5-hour usage limit in the middle of a task, Claude would basically just stop, even if it was halfway through editing your code. Now, Claude Code will try to find a graceful stopping point, using a small fixed allowance from your weekly limit to wrap up whatever it can, instead of abruptly cutting you off. And as this rolls out, pro users will get this once per week, whereas with max and

13:00 team premium users will get it every time they hit their 5 hour session limit. And if you still need claw to keep working after that, you can continue through extra usage. It’s a small change, but honestly, a very useful quality of life update that Cloud Code has introduced, and I really love what they’re trying to do with these new updates. Recently, I had found a way to actually find out new development models within Open Codes catalog. And ever since then, new several unreleased

13:30 models have been appearing like GLM 5.5 Flash, GLM 5.4, Kim K4, and Deepseek version 4.1 Pro. Now, we have no idea when these models will officially drop. And their appearance here obviously doesn’t confirm an imminent release, but seeing all four already showing up within Open Code is definitely interesting and it really shows that we are going to have a really wild fall, especially with new model drops. Now, I know this next segment may sound like it

14:00 is sponsored, but that’s not actually the case. I really wanted to help you out with this. This is where I had found out that Nvidia is giving away access to four incredibly strong AI models completely for free. And you don’t even need a credit card cuz this is where they are offering four models like Deepseek version 4.1 Flash. You have GLM 5.3 as well as the Flash variant and Kim K3 through Nvidia’s platform. And this isn’t just a playground because you get actually access to these models through

14:30 an OpenAI compatible API, meaning you can plug them into your apps, coding tools, and workflows pretty easily. So, if you’re wanting to experiment with these models, I’ll leave links to this in the description below, which is definitely worth checking out. But there is one pretty important thing to keep in mind before you use these free models. Nvidia clearly states that during the free trial, your inputs and outputs are recorded to provide service and help Nvidia’s products and services, including its AI models, to be trained

15:01 upon. They specifically warn you to not upload credentials, information, or personal data unless permitted. So, yes, you’re getting access to extremely powerful models for free, but I definitely wouldn’t be throwing any private code sensitive company data or personal information into these endpoints. If you 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

15:33 free on a monthly basis, plus daily AI news and exclusive content, plus a lot more. Next is a story that gets even wilder because OpenI revealed that during its hugging face incident, an AI agent inside its research environment actually found ways to break out of its intended restrictions, access the internet, communicate with other agents, and interact with third party systems when they were not supposed to. And now OpenAI revealed that some data was

16:03 actually improperly sent to third party services. Most of it apparently didn’t come from users, but they discovered 53 cases where images uploaded by users were posted to external image hosting sites as unlisted links. Open says that these came from accounts that opted into model improvement and the images had already been disassociated from those accounts and passed through a privacy filter. But what makes this a lot more

16:33 crazier and this incident so interesting is that these agents were essentially finding unintended ways around the environment that they were placed in. They essentially wanted to be finding ways to escape all of these different environments so that they can do whatever they want. They had their own mind. You can say OpenAI says that the agents discovered ways to communicate with each other and gain internet access which is wild and eventually exploiting vulnerabilities through its own research infrastructure and the hugging face

17:04 system. OpenAI says that it is something that they have worked upon since this incident and they are working with the hosting providers to remove most of the affected content and it is working to remove the rest. And now what they’re doing is they are obviously going to be implementing significantly stronger security monitoring, sandboxing, and alignment safeguards. But the whole situation that we have seen is honestly just wild cuz we have rogue agents now

17:34 doing whatever they want. And it’s crazy cuz imagine when AI gets a lot more intelligent. we are going to have these rogue agents do whatever they want unless we are to take proper action and put up enough safeguards. Next, we need to talk about China’s absolutely massive AI infrastructure boom because as I’ve been saying, I genuinely think people are sleeping on the next bite dance move that is going to be coming from China’s AI scene. And this is because there is

18:04 new data from semi analysis which maps out over a thousand data center facilities across more than 60 operators in China alone showing just how quickly the country is building out the infrastructure needed for AI. And some of these numbers are kind of crazy cuz China’s largest hyperscaler is reportedly leasing around 1/5th of the country’s total capacity while massive 100 megawatt deployments can come online in roughly 12 months. This is all happening alongside China’s broader

18:37 eastern data western computing strategy which is essentially pushing huge amounts of computing infrastructure into regions where power and land are more readily available. And this is why I keep bringing up bite dance. Everyone is focused on the model releases themselves. But there is an enormous infrastructure race happening underneath them. If bite dance really is scaling towards this new model that is going to be 10 trillion parameters. China is clearly building the compute infrastructures to support that ambition

19:07 and that is where we could see that model even come later this year which are the reports that we had seen earlier. Lastly, we have Agent Ultra. And this is a pretty cool step forward for deep research agents. Now, I had saw this on X today and I thought it would be pretty interesting because a lot of people on the channel seem to be looking towards finding new deep research agent strategies and I believe this deep research agent is going to be something that will help you guys out cuz ultra

19:39 agent or agent ultra can orchestrate entire swarms of agents to research a topic in parallel. And this is something that Exa has designed for massive research tasks where you might need to go through hundreds or even thousands of sources, build comprehensive lists, compare information, and eventually bring everything together into one answer. But what’s interesting is that Agent Ultra is state-of-the-art on web research while also sitting on the Parto frontier for cost versus research quality. So, it’s definitely a pretty

20:10 cool advancement and I really wanted to showcase this on today’s video. But that’s about it, guys, for today’s video. If you guys haven’t already, make sure you guys go ahead and take a look at the world of AI insider school community. Take a look at the world of AI benchmark and vibe coding platform. Take a look at the second channel. This is where you can stay uptoate with the latest AI news. Join the newsletter, join 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 the latest AI news. But

20:40 with that thought, guys, have an amazing day, spread positivity, and I’ll see you guys fairly shortly. He suffers.