Gemini 4 Carbon Checkpoint, Qwen 4 and Kimi K3.1 Are Coming, and Anthropic Has a Problem!

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 Carbon Checkpoint, Qwen 4 and Kimi K3.1 Are Coming, and Anthropic Has a Problem! · Channel: Universe of AI · 11:54 · Watch on YouTube

Topics: Markets and economy, AI models, Local and open-source models

Summary

The video reviews early signs of a possible Gemini 4 Carbon checkpoint, which the presenter expects to be a faster, more efficient Flash model, and discusses leaked examples of its generated games and voxel scenes. It also reports that Qwen 4 and Kimi K3.1 may be nearing release, while noting uncertainty around timelines and performance. A Wall Street Journal report about Anthropic’s draft IPO filing raises questions about its reported losses, computing costs, revenue, and infrastructure commitments. Finally, the video says OpenAI acknowledged usage-limit inconsistencies affecting about 40,000 accounts and is contacting affected users with corrections and credits.

Key points

Gemini 4 Carbon is presented as an early checkpoint that may be a faster, more efficient Flash model based on the same training base as Argon.
The presenter says internal claims suggest Carbon could match Opus 5.5 on coding, but regards this as uncertain and does not expect it to outperform the model.
Early examples include a shooting-target game and voxel scenes; the presenter notes improvements in detail but describes the evidence as preliminary.
Some predictions put Gemini 4 Carbon at 400% cheaper than Astra, which the presenter takes as a sign it may be in the Flash pricing tier.
Qwen 4 is reportedly in early testing, and the video expects it to be an open-weight model.
Kimi K3.1 was reportedly spotted in a repository; the presenter says a new model could arrive within two to three weeks, while stressing that timelines can change.
The video says Anthropic is aiming for an IPO in November and discusses a reported $80 billion net loss alongside rising revenue, computing costs, and large cloud-contract commitments.
OpenAI reportedly acknowledged usage-limit inconsistencies affecting approximately 40,000 accounts and is correcting the issue and providing credits to affected users.

Chapters

Start	Chapter	Summary
00:00	Gemini 4 Carbon: early checkpoint and examples	The presenter discusses expectations for Carbon, its possible Flash-model positioning, and early generated games and voxel scenes.
06:42	Upcoming models from Chinese AI labs	The video covers reports of Qwen 4 testing and Kimi K3.1 appearing in a repository, while noting that release timing is uncertain.
08:48	Anthropic’s reported finances and possible IPO	The presenter summarizes reporting on Anthropic’s planned IPO, revenue growth, losses, computing costs, and infrastructure commitments.
10:53	OpenAI usage-limit issue	The video says OpenAI is addressing usage-limit inconsistencies affecting about 40,000 accounts and issuing credits to affected users.

Quotes

Gemini 4 Carbon, which is expected to be around 400% cheaper than GPT Astra. — 00:00

It’s okay, nothing mind-blowing so far. — 03:04

Gemini K 3.1 was actually spotted in a repo last night — 07:45

they are at an $80 billion net loss. — 09:19

approximately 40,000 accounts basically had an inconsistency on which usage limits were actually applied. — 11:24

Concepts and entities

Gemini 4 Carbon, Gemini 4 Argon, GPT Astra, Opus 5.5, Qwen4, Kimi K3.1, Anthropic, OpenAI, Codex, Wall Street Journal, DeepSeek, MiniMax

Full transcript (original language)

00:00 We have a new checkpoint from an upcoming model, Gemini 4 Carbon, which is expected to be around 400% cheaper than GPT Astra. We also have some new news from the Chinese AI labs, and it looks like Gwen 4 and Kimi K3.1 could be coming in a few weeks. We also got a new news report from the Wall Street Journal today saying that Anthropic is operating at a $80 billion loss. And lastly, OpenAI just admitted that they have an issue that affected approximately 40,000

00:31 accounts. So, let’s get into it. So, if you’ve been following my channel, I’ve talked about an upcoming model from the Google DeepMind team called Gemini 4 Carbon. Now, this model, we kind of have some idea what it’s going to be. A lot of people internally have been saying that this model is going to match the performance of Opus 5.5. Obviously, these are internal employees, so they have an incentive to say that, but we also know that this model is supposed to not only excel at coding, but also be faster at it. And most likely, based off

01:01 what I’m seeing so far and based off the leaks that I’m going to show you guys today, it seems like it’s going to be a new flash model from the Google DeepMind team. Now, what this flash model is probably going to be based off of is going to be the same training base as Argon, but it’s going to be obviously faster and more efficient than the current Gemini 4 Argon model. So, you’ll still have that intelligence level of Gemini 4 Argon. Maybe it spends less time thinking compared to the Argon model, which obviously a little bit too much at the moment. And this model is

01:31 going to basically trade away some of that thinking for faster responses. And the outputs I’m going to show you guys today kind of show that this model has improved in 3D scenes and voxel worlds. Hey, before we get into the video, I just want to take a moment to tell you guys that AI is not slowing down anytime soon. And if you guys want to stay on top of everything that is happening in AI, be on top of all the new model drops, then make sure you guys are subscribed to the channel. Your support means a lot, and now let’s get back into

02:02 the video. Now, this is something that we love testing with these models, especially flash models, because these are obviously one-shot prompt generation. And what these generations test is the model’s ability to think on the go super fast and generate something that’s actually comprehensive and works. So, obviously these outputs are very early. We don’t know how they’re actually going to be in deployment, but if we just take a look at this, we have a I guess a shooting target game that somebody was able to generate with this. Now, obviously we don’t have the actual

02:34 demo, unfortunately, so I’m showing you guys early leaks. And as I said, this is very early. If yesterday was the first time we heard about them actually speeding up the deployment of this model, and today we have some idea of what this model looks like, and I’ll be tracking this story just in case if there are more leaks I can show you guys. But what I’m showing you guys right now are very early generations of this model. So, what we can see from this model, at least at the moment, design-wise, this is nothing crazy. Obviously, this depends on the prompt, and if we just look at it, it looks like

03:04 the person was able to create a simple shooting target practice game what we are seeing at the moment. And it’s okay, nothing mind-blowing so far. But then if we look at some other examples, especially on the Voxel world side, we’ve seen this test evolve a lot in 2026. In 2025, we were generating simple gardens with this. Now, we’re adding elements like this, where we have the garden, the trees and everything are much more detailed. It’s less blocky, obviously still going to be blocky cuz it’s a Voxel world, but less blocky

03:35 compared to before. And we are seeing the models actually add design elements on their own that surprisingly make sense to add. Sometimes we would have some elements that didn’t work. For example, if you guys remember in earlier videos, I would show like a pond with fishes in there, but the fishes did not look like they were actual fishes. And sometimes these new models still make those mistakes, but what we can see from this generation is that it looks like the model is adding more details. However, I feel like this model is not going to like outperform Gemini for

04:08 Oregon. I feel like the carbon model is going to be a more efficient version of Gemini for Oregon. So, the earlier test that people have been saying internally that this model supposed to match Opus 5.5 on coding. Maybe it gets on par with it. I don’t see it actually outperforming Opus 5.5. Obviously, this is just my assumptions based off simple test that I’m seeing. Now, here’s the famous PS5 SVG controller test. We’ve actually seen better demos of this. We’ve seen at least the Opus 5.5 model basically

04:39 create something that pretty much looks like a image of the controller. Now, we’re seeing more generations of this from the new checkpoint, which is probably the carbon model. And as I said, more and more tests give me more confidence to say that this is probably going to be another flash model from the Google DeepMind team. They love flashing us. They’ve been flashing us throughout this whole year with like 3.8, 7, whatever flash models that they dropped. I think they dropped about four to five flash models this year, and they’re probably going to add to that list. And flash models are great. They are fast,

05:10 they’re efficient, and they’re also most of the times cost-effective. So, some people are actually predicting that this model is going to be way cheaper compared to before, and I’ll show you guys some prediction. But here are some more generations. Here’s a voxel castle, which I thought was pretty well designed by the model. And you know, it looks pretty neat. Like this is something that, you know, we kind of struggled with the models literally, especially the Google models, couple of months back. Now, this is becoming normal for them. So, the new models, as I mentioned, some people are guessing that the Gemini

05:42 carbon model is going to be 400% cheaper than Astra. Which, if we do the math would mean that this is most likely in the flash tier when it comes to the model pricing. It’s not going to be a frontier pro model like the Gemini 4 Argon model. It’s going to be a flash model most likely and that is why I can confidently say that right now at least we are leaning towards a new flash model rather than a new frontier pro model. Then we also have Fable 5.5 which, you

06:12 know, we all have been waiting and I feel like in topic has that on the go. I don’t know when they’re going to drop it. I hope they drop it soon because we’ve seen the outputs from this model and they are very, very different from anything that we have seen so far. So I agree with this person saying that this will top all of the leaderboards. I wouldn’t be surprised. And then we also have a new model from OpenAI which I talked about yesterday a lot which was a Terra model and that is going to be their mid-size model and the pricing at the moment is going to be $5 per 1

06:42 million tokens. So at least Carbon and Terra are expected to be on the cheaper side of newer models to come. Fable 5.5 it could be cheaper than Fable 5.1 based off what Anthropic has been doing recently but obviously I think it’s still going to be on the more expensive side. So these are the upcoming models. We have some early checkpoints of the Carbon model and I’ll show you guys more leaks as they come through. Now it looks like the Chinese AI labs are also not slowing down. We already know that we have a new DeepSeek model, a new MiniMax

07:13 model but it looks like Qwen 4 is in early testing and the Qwen lab has been on a roll in 2026. All of their open weight models were very successful. They competed with the frontier labs. They overtook sometimes even the mid-tier models from the frontier labs like OpenAI and obviously Google DeepMind earlier this year and we could expect them to do something similar with Qwen 4 because this model is expected to be once again a powerhouse of a model and it’s going to be open weight. So for all of us who enjoy open weight models, open

07:45 source models, this is going to be another enjoyable few periods for us because new models are coming. And even Gemini K 3.1 was actually spotted in a repo last night, which usually indicates that this model is almost ready for deployment. Maybe they’re doing some last-minute testing. Obviously, we don’t know what the timeline is going to be because this usually sometimes happens. We’ve actually seen it come up in the repos and then has been delayed, but we I can say with some sort of confidence that within the next 2 to 3 weeks, we

08:17 should have a new Gemini model. Now, Zeta AI, there’s some information on it that we saw yesterday’s video where Gemini 5.5 was kind of hinted at through the new paper that they dropped, but you know, there isn’t more evidence other than that. So, we don’t know when Zeta AI is going to drop a new Gemini model, but at least Gwen 4 and Gemini K 3.1 seems sooner than Gemini 5.5. Now, something interesting happened today. The Wall Street Journal actually posted an article today on Anthropic’s

08:48 draft IPO filing. So, it looks like Anthropic is aiming for an IPO sometime in November, and this is probably going to be one of the biggest IPOs this year and probably one of the biggest IPOs for the AI space. This will kind of determine how big of an AI bubble we are in. The term AI bubble has been used a lot recently, and Anthropic’s IPO is going to be a good test to determine how big of a bubble we are in. At least at the moment, what we are seeing doesn’t seem super positive. Anthropic is

09:19 operating at really high computing costs, and they are at an $80 billion net loss. Now, if you are an optimist, when you look at this news, you’re not surprised. You understand that this is kind of like the early stages of this technology. These are the foundational years, and most of these companies are probably operating at a loss with high debt and you know, you are hoping that their capex eventually leads them to make a return in the long term. And but if you are AI, I guess not an optimist, a pessimist, you are probably going to look at this news and expect that we are

09:51 all in the AI bubble. I’ll let you guys make your own decisions, but I’m just going to show you guys couple of highlights from that article. Number one, the first half of revenue this year jumped from 1.1 billion to 16.3 billion. So these companies are generating revenue. They have a lot of revenue coming in, but their cost and the amount of debt they have is the key concern for many companies. Their computing infrastructure spending actually rose from $2.3 billion to roughly $9 billion. And the revenue

10:22 obviously is scaling much faster than the compute cost, but their commitments, meaning debt, are enormous. 450 billion in cloud contracts, mostly non-cancelable as this person mentions, plus other infrastructure deals. Another 150 billion is under negotiation. So a lot of their deals are obviously, you know, laid out, but there’s no cash flow coming in. And obviously their interest payment on their debt is going to obviously catch up as well. But obviously this is early news. We don’t have official balance sheets or income

10:53 statement from these companies yet. But as Anthropic is preparing for the IPO, especially in sometime November, we’re going to get more news like this. So keep that in mind because this will determine as I said if we are in {quote} an AI bubble or not. Now some housekeeping news for you guys. If you guys have been using your OpenAI models and you noticed that your usage was actually draining faster than you expected, that is not something that you were hallucinating about. That is exactly what happened. OpenAI or Tibo on

11:24 from Codex basically admitted that approximately 40,000 accounts basically had an inconsistency on which usage limits were actually applied. And you know, because of that, some people were experiencing their usages empty faster. So, what they’re doing right now is that they’re reaching out individually to all of those 40,000 users and is correcting the issue. And if you have been affected by this, so check your email, you’ll also get credits to your account. So, good thing that OpenAI is fixing this, so keep that in mind if you see some new

11:54 credits or you miss an email, might be the time to check your email just to make sure that you are not one of the users that had this issue. 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 universeofai.z as well. Until then, I’ll see you guys in the next video.