Gemini 4 Pro CRUSHES Opus 5.5! Secret Arena Checkpoint, October Release & More! AI NEWS

Author: Universe of AI · Watch on YouTube

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: Gemini 4 Pro CRUSHES Opus 5.5! Secret Arena Checkpoint, October Release & More! AI NEWS · Channel: Universe of AI · 11:22 · Watch on YouTube

Topics: Enterprise AI applications, AI evaluation and benchmarks, AI content generation, AI models

Summary

The video discusses a purported Google Gemini 4 Pro checkpoint, called “barium B,” whose outputs are presented as competitive with Opus 5.5 and, in some tests, stronger than GPT-6 Astra. The host highlights reported improvements in speed and 3D output quality, while stressing that the checkpoint, release timing, and leaked benchmarks are not officially confirmed. The video also covers unfamiliar models appearing on OpenCode and code suggesting OpenAI may be developing an ultra-fast mode and an assistant product called “O.”

Key points

A checkpoint called “barium B” is reportedly being served under the Gemini 3.8 Flash model, with some users seeing stronger outputs.
Examples of 3D and interactive generation are used to compare the checkpoint with Opus 5.5 and GPT-6 Astra.
The host says the checkpoint appears faster than other flagship models and shows improvements in design and detail, but Google has not confirmed it.
A possible October release is discussed as speculation; the video gives no confirmed launch timeline.
Leaked screenshots allegedly show a 256K output target, a 10-million-token context, browser navigation, persistent memory, robotic control, and pricing.
The host cautions that the leaked benchmarks and pricing are unverified.
Models including Jilin 5.5 Flash, Jilin 5.4, Kimi K4, and DeepSeek version 4.1 Pro reportedly appeared on OpenCode, possibly due to a glitch or upcoming releases.
OpenAI code references to “ultra fast” are interpreted as a possible new speed mode for models or agents.
The video reports speculation that OpenAI may introduce an always-on assistant product named “O” around DevDay.

Chapters

Start	Chapter	Summary
00:00	Gemini 4 Pro checkpoint and output comparisons	The host presents the alleged “barium B” checkpoint and compares its visual and 3D-generation outputs with Opus 5.5 and GPT-6 Astra.
05:42	Checkpoint details, timing, and leaked claims	The video discusses reported speed and design improvements, a possible October release, and unverified benchmark, capability, and pricing leaks.
08:47	Models appearing on OpenCode	The host lists unfamiliar model names spotted on OpenCode and considers whether their appearance was a glitch or a sign of upcoming releases.
09:49	OpenAI’s ultra-fast mode and assistant rumors	The video examines code references to an “ultra fast” mode and claims that OpenAI may be preparing an assistant product called “O.”

Quotes

this model is producing crazy outputs — 00:00

the level of outputs are matching Opus 5.5. — 03:08

these are leaked. These are not official — 07:44

We are seeing ultra fast being added after fast — 09:49

Lowercase O is the assistant’s product name. — 10:50

Concepts and entities

Gemini 4 Pro, barium B, Gemini 3.8 Flash, Opus 5.5, GPT-6 Astra, Google DeepMind, EleutherAI, OpenCode, Jilin 5.5 Flash, Jilin 5.4, Kimi K4, DeepSeek V4.1 Pro, OpenAI, DevDay, ultra fast, OpenAI O, Grok Bot, RSI, Three.js

Full transcript (original language)

00:00 It looks like we have a new checkpoint for the Google Gemini 4 Pro model called barium B and this model is producing crazy outputs that are quite competitive to Opus 5.5 [music] and GPT-6 Astra. We also found some interesting models showing up on open code today, Jilin 5.5 Flash, Jilin 5.4, Kimi K4 and DeepSeek version 4.1 [music] Pro. And it looks like OpenAI is getting ready for DevDay because they are writing ultra fast into everything. So there’s ultra fast mode coming to OpenAI models. Not only that,

00:30 there is a leak that OpenAI oh, which is going to be OpenAI’s own version of Grok bot is coming on DevDay. So we’ll talk about that as well today. So let’s get into the video. So the new Gemini 4 Pro checkpoint is called barium B and similarly like the other checkpoints, this model is being served underneath the Gemini 3.8 Flash model. So when you select this, the outputs that the model is generating is definitely stronger than Gemini 3.8 Flash. If you compare the outputs that you’re getting now, for some people this is not happening for every single account on EleutherAI, you

01:02 will be definitely seeing a difference in the output level quality because the outputs today are way more stronger compared to last week. If we look at this example that you see on screen where the Gemini 3.8 Flash model was told to create an F1 car and kind of create a simulation of it being able to change the colors, being able to, you know, mess around with it, you can clearly see that the level of output is quite competitive to Opus 5.5. This is something that you would expect from the Opus 5.5 not the Gemini 3.8 Flash model.

01:33 So this is probably a signal that Gemini 4 Pro, this checkpoint is being tested. And if people have been enjoying this checkpoint and it’s been performing well on EleutherAI, then we probably are going to see this checkpoint being used for the Gemini 4 Pro model. That is usually how Google has, you know, determined what checkpoint to use. But one thing is, they have delayed this model quite significantly. This was originally supposed to be the Gemini 3.5 Pro model. That didn’t work out for them. They dropped a bunch of flash variants. Now they are working on the

02:05 Gemini 4 Pro model. And now I think they have produced a checkpoint that is going to compete with Opus 5.5 and GPT-6 Astra. In some cases, the outputs that you’re getting from this model are better than even GPT-6 Astra. Here’s another output which is a lighthouse. This is something that you do not expect the Gemini 3.8 flash model to produce. You can see from the lighthouse we have, I don’t know, it looks like a rocket ship being produced from there and like it’s being launched. This level of

02:36 detail is something that most of the models now kind of struggle with. And if this output is just created using 3.js, 3D quality textures and effects, all of those elements have definitely improved in the model. So it looks like Google might be finally ready to drop something serious because we are all probably tired of their flash variant models. They’re fun to use, they’re effective, they’re cheap, but we need a pro model from them, especially from a lab like Google DeepMind. Here’s another output from Gemini 4 Pro. And if you compare this to Opus 5.5, this is why I was

03:08 saying this model is producing output similar to Opus 5.5. And Opus 5.5 has only been out for about a week now. And that is probably the best model when it comes to generating outputs, 3D quality, things like that. Opus 5.5 is a current state of the art model. Now Gemini 4 Pro, the level of outputs are matching Opus 5.5. But one thing to remember, usually the Gemini team, whenever they drop a model, their pricing is more competitive than Opus 5.5, meaning it’s always on the cheaper side. So let’s see

03:38 if this time around, if the Google DeepMind team decide to go with this checkpoint and finally drop this model, they might be putting some pressure on not only Anthropic but even OpenAI because their newest model is going to be most likely, in my opinion, cheaper than Anthropic’s and OpenAI’s and going to be creating outputs that are on par with Opus 5.5. So, that will be a big win for the Google DeepMind team because they have been very quiet in 2026. We don’t really have a strong pro model from them yet, and Gemini 4 Pro might be

04:09 the answer to that question. Now, if you guys wanted to see what this model looks like compared to GPT-6 Astra, yeah, it does not look that great. Like, GPT-6 Astra, it generated the horse, but the horse is, you know, his legs are backwards, I guess, and I moving in the same spot. I don’t know. I don’t know how the physics of that works, but you can see at the top the Gemini 4 Pro model in arena, obviously, 3. underneath 3.8 flash is producing a horse that actually looks okay. Obviously, the physics of generating a

04:40 horse for these models is quite difficult, and I I would love to see what Opus 5.5 created on this test, but you can see that the Gemini 4 Pro model is definitely producing a way stronger output compared to GPT-6 Astra at the bottom. So, yeah, OpenAI, I know their dev day’s coming up on September 29th. They might have some pressure from the Google DeepMind team. Is Google DeepMind trying to pull an Anthropic move by dropping something powerful right before OpenAI’s dev day? That could be the case, who knows, or this model is still

05:11 early because this is just a checkpoint. We have no confirmation from the Google DeepMind team about any timelines, and I just hope that this time around they don’t delay it further because finally they have something that is producing similar outputs to Opus 5.5 and beating GPT-6 Astra. So, this would be a perfect time for them to drop something new. This is another output where we’re comparing Gemini 4 Pro to Opus 5.5. In my opinion, in this specific test, the level of detail that the Gemini 4 Pro model has versus Opus 5.5,

05:42 I think it’s a bit better. Even if we look at like just the colors and things like that versus Opus 5.5. The Gemini 4 Pro model is producing sharper outputs compared to Opus 5.5 in this specific test. So, in my opinion, I think Gemini 4 Pro takes the lead on this specific test. Now, what do we know about the new checkpoint? There are a couple of leaks that I’m going to show you guys. Leaked benchmarks, some leaked pricings, but what we do know is that this checkpoint is almost two times faster than any of the other flagship models, which is a

06:12 good thing. So, I hopefully I don’t think this is going to be a flash model cuz if this is a flash model once again, that would be crazy to see if they drop a capable flash model and if it’s going to be faster and more efficient than the other frontier labs. It has very good design taste. We all saw that. And this person is saying it has the best taste I have seen in any model. It now creates things in much more detail in 3D and pays attention to small details. And we saw that with some of the outputs where the colors were sharper, the level of detail was stronger compared to the other labs.

06:43 Now, if you compare this to the previous checkpoint, it did make a good improvement, especially in speed and overall design. Now, one thing that we do know or we don’t really know for sure, but what has been leaked is that Google DeepMind has solved RSI, which is recursive self-improvement. And if that is the case, their models are working on creating stronger versions of themselves. So, if the checkpoint last time was okay, then the model might have improved that checkpoint and created this new version, which is clearly stronger compared to what we saw couple of weeks back. And if this continues the

07:14 speed of improvement that they’re saying, they might launch it in the first or middle of October according to this person’s predictions. I think that would be fair to see. Obviously, we are entering the early stages of October. And as I said, OpenAI’s DevDay is on September 29th. So, either they wait for the DevDay to finish and drop something after to kind of spoil OpenAI’s DevDay. I don’t think we’ll expect this to come out before September 29th. So, yes, I agree with this. The October timeline does make more sense. Now, I’m going to

07:44 show you guys some leaked benchmarks. And I remember, these are leaked. These are not official, so I would take everything that you’re about to see with a grain of salt. So, what we do know from the alleged internal screenshot is that this model is supposed to have a output target of 256 K, 10 million token context, autonomous browser navigation, persistent memory, robotic control capabilities, which I think is interesting. And the claimed pricing is $2.25 for 1 million input, $11.25 for 1

08:16 million output, which is around, I would say, OpenAI and Anthropic’s model, a little bit cheaper than them. And if this is all true, if you look at the benchmarks on the left, you can see that this model is pretty much beating majority of the models that are out there. We can see that even on like deep software engineering, it has a 88.7% versus, if you look at Fable 5.1, 69.1, GPT-6 Astra, 86.9. They don’t have Opus 5.5 on here because that model is stronger than, obviously, GPT-6 Astra and Fable 5.1. So, I’d love to see that

08:47 side by side. But these are leaked benchmarks, so I’m not going to waste so much time going through them. I’d rather wait for the official benchmarks, but just wanted to give you guys some idea that if these screenshots are real, you would expect this model to, obviously, outperform GPT-6 Astra and Fable 5.1. Now, at the same time while all this is happening, something weird happened on OpenCode where interesting models showed up. And these are models that don’t exist, which was Jilin 5.5 Flash, Jilin 5.4, Kimmy K4, DeepSeek Version 4.1 Pro.

09:19 Now, this could be just a glitch on OpenCode, or this could be them getting ready to drop a new model, meaning the newest models from the open-source labs. And that could be the case, but I just wanted to show you guys just in case you hear more news about this. Obviously, this is the first time I’m hearing about Jilin 5.5 Flash and Gemini 5.4. I do know that Gemini K4 was coming. We thought it was going to be Gemini K3.1, so I’m not sure why it’s Gemini K4. Obviously, this is just a speculation, but let’s see if there’s more meat to

09:49 this story. Now, as OpenAI is getting ready for Dev Day, we are seeing something interesting happen in their code. We are seeing ultra fast being added after fast, which probably means they’re working towards creating an ultra fast mode on a tier that you can set on the agent. Now, which is a good thing, which means that you can work with OpenAI’s models and their agents a little bit faster compared to fast ultra fast. Obviously, we’ll find out very soon cuz Dev Day is only around the corner, but you can see like across their code bases we are seeing ultra fast show up in many areas.

10:20 We can see it over here, over here as well. So, it looks like they’re working towards dropping an ultra fast mode. Not only that, there’s something interesting happening. So, Tibor first mentioned, “Oh, yes, we’re back in action and we’ll reset usage limits for all paid users.” Blah blah blah blah. This was just a reset limit, and then he keeps on using O, and then he also said, “Oh, no.” Which is sometimes signal that something named O is coming. So, what are people speculating? Well, not even speculating is coming up in their code as well. O, you’re always on assistant, whatever.

10:50 People are saying that this is going to be OpenAI’s response to Grok bot. They’re dropping something called OpenAI O, and this is going to be the assistant’s product name. So, we can just take a look at this screenshot. You can see O, you’re always on assistant. Lowercase O is the assistant’s product name. You can see that it’s being called that. And then we also see O, benefit shared by all pro tiers. O, and this is a different language. I think this is German, and yeah, I don’t know what these languages are. This I think is Hindi. This I do not know.

11:22 But anyways, we know that a product from OpenAI is coming that is going to be called O. It’s going to be an assistant product. We don’t know how powerful this assistant is going to be, but keep that in mind because DevDay is right around the corner. But, that’s it for today’s video. Make sure you guys are subscribed to the channel. Follow our new newsletter as well at universeofai.beehive.com. As well subscribe to the main channel World of AI and support us on X by following the as well. Until then, I’ll see you guys in the next video.