Sonnet 5.5 Next Week, GPT-7?, Qwen 4 Spotted and DeepSeek V5 Beats Astra

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: Sonnet 5.5 Next Week, GPT-7?, Qwen 4 Spotted and DeepSeek V5 Beats Astra · Channel: Universe of AI · 12:53 · Watch on YouTube

Topics: AI infrastructure and hardware, Markets and economy, AI models, Open source and software ecosystems

Summary

The video discusses reports that Anthropic may release Sonnet 5.5 next week, with leaked tests suggesting competitive coding and video-recreation abilities, though the host says its performance against other models remains unconfirmed. It also covers OpenAI’s reported focus on GPT-7 and GPT-8, and speculation about how future model bases could change the competition. The video then reviews possible upcoming Chinese models, including Qwen 4 and MiniMax releases, before discussing claims about DeepSeek V5 and China potentially allowing major AI labs to use Nvidia chips again. Many of the model details and comparisons are presented as leaks, reports, or predictions rather than confirmed releases.

Key points

The host predicts Anthropic could release Sonnet 5.5 around OpenAI’s September 29 Dev Day, citing reports of an updated checkpoint shared with partners.
A demonstration reportedly showed Sonnet 5.5 recreating an AI-generated video using a single HTML file, with the host estimating about a 70% frame match.
The host expects Sonnet 5.5 to be more efficient and cheaper than Opus, but says its actual standing relative to GPT-6 models can only be judged after release.
An OpenAI executive is reported to have said the company spends 80–90% of its time or research on GPT-7, GPT-8, and beyond.
The video describes Doug as the base behind current OpenAI models and Bell as a more capable future base, while suggesting GPT-7 may already be in fine-tuning.
Leaked Qwen 4 outputs are described as potentially competitive in some areas, particularly front-end coding, though the host says more testing is needed.
The video discusses uncertainty over whether MiniMax M3.1 Flash and the leaked “Space Bunny” are separate models.
DeepSeek V5 is reported to be an imminent launch and is rumored to have 2 trillion parameters and to train on Huawei Ascend chips.
The host says China may be considering allowing major AI companies to use Nvidia hardware again because of compute shortages, which could accelerate model development.

Chapters

Start	Chapter	Summary
00:00	Sonnet 5.5 may arrive next week	The host reviews leaked Sonnet 5.5 checkpoints, possible timing around OpenAI Dev Day, and early reports of its capabilities and efficiency.
05:23	OpenAI’s focus on GPT-7 and GPT-8	The video discusses OpenAI’s reported research priorities, its current and future model bases, and speculation about the next generation of models.
08:32	Qwen 4 and MiniMax leaks	The host considers leaked Qwen 4 generations and possible MiniMax M3.1 variants, noting that their capabilities and release plans remain uncertain.
10:43	DeepSeek V5 and Nvidia chips in China	The video covers rumors about DeepSeek V5’s size and training hardware, followed by reports that China may ease Nvidia chip restrictions for AI companies.

Quotes

next week is going to be a huge week in AI — 00:00

the model is consistently getting a frame match of, I would say, a 70%. — 03:18

There’s a newer model base, which is way more capable than obviously Doug, called Bell. — 06:25

This is going to be a 2 trillion parameter model. This is the rumor. — 11:14

this would be a complete 180 turn for China, but who knows? — 12:20

Concepts and entities

Sonnet 5.5, OpenAI Dev Day, GPT-7, GPT-8, Doug, Bell, Qwen4, MiniMax M3.1, Space Bunny, DeepSeek V5, Huawei Ascend, NVIDIA, Anthropic, OpenAI, MiniMax, DeepSeek

Full transcript (original language)

00:00 So, it looks like next week is going to be a huge week in AI because number one, we have OpenAI’s Dev Day, but it looks like Anthropic is getting ready to drop a newer model, and that is going to be a new Sonnet model, specifically Sonnet 5.5. We saw this model kind of leak a couple of days ago, and we saw that checkpoint, and that checkpoint was very competitive. The model was producing outputs that you don’t expect Sonnet model levels to be producing, and that kind of indicated that Anthropic was testing a newer version of Sonnet, which is probably going to be the Sonnet 5.5 model. Now, we have some new information

00:33 that some partners actually got access to another checkpoint of Sonnet 5.5, and that checkpoint is apparently producing better outputs than the one that was previously available last week, which usually indicates that they took the learnings from the leaked checkpoint, they took that and made the model better. And that model checkpoint, the updated one, is probably going to be coming next week. That is my prediction. There’s a couple of reasons why. Because OpenAI is having their own Dev Day on September 29th, which is this Tuesday,

01:06 and Anthropic has a tendency to spoil other people’s parties, and they’re planning on spoiling OpenAI’s party, meaning their Dev Day, by dropping a strong model. Now, Opus 5.5 is obviously a pretty strong model. It’s already better than GPT-6 Astra. It’s also better than Anthropic’s own Fable 5.1. Now, Sonnet 5.5 is going to be Anthropic’s answer to GPT-6 so and Luna because Sonnet models are obviously cheaper than Opus 5.5 model, and the Opus models already have a the newest ones already have a 20% discount on it.

01:40 So, if they drop a cheaper model that is producing stronger outputs than the sole model lineup, or even if it’s a little bit more expensive, but it’s producing way more capable outputs than the sole lineup, that is going to put some pressure on OpenAI’s newest model. So, what What know about Sonnet 5.5? Well, we do know that it’s been stealth tested. That’s pretty much it. But, we do know that some of the outputs that the model is producing is quite competitive. Before we get into the video, I just want to take a moment to thank you guys for all the support of the last couple of months. Every single

02:11 view, every comment, and every share is helping this channel grow, and I don’t take any of that for granted. And if you’re new here and haven’t subscribed yet, hit that subscribe button so you don’t miss what’s next in AI. It’s free, and it genuinely helps grow this channel a lot. Thanks for watching today’s video. Now, let’s get back into it. And we also know that recently this model’s and the model slug was kind of leaked as well. If you take a look at this picture, we can see Anthropic’s model lineup, and Sonic 5.5 is being referenced now, which usually means that the model is going to come soon. Also,

02:44 if we take a look at this output that we saw, we see on the left there’s an AI-generated video of, you know, “It All Started with the Middle Ego,” the famous AI video that gained a lot of traction over the last couple of weeks. Now, Sonic 5.5 was told to recreate that video in just pure code, one HTML file, and not use any assets. And if you put these side by side, they look like they’re identical. Obviously, there are some differences in some elements, and obviously this is code versus, you know, using a image generation or video generation on the left. But, if we look at this, guys, I

03:18 think the match rate of the video and the generation is pretty much strong. At the bottom, you can also see the frame match, and the model is consistently getting a frame match of, I would say, a 70%. But, anyways, what this is clearly saying is that this model is going to be competitive, too. Opus 5.5 probably, you know, is the strongest model right now, but they might be taking the number two spot with Sonic 5.5. This person is saying that this is not going to be Astra level, but it’s obviously better than Soul and closer to Astra than it

03:50 should be, which I agree with because, you know, Astra is still going to be, I would say, in most cases a better model. In some tasks and some generations, people might be, you know, enjoying Sonnet 5.5’s ability to design things, things like that, but overall, I think Astra is still going to be a model that is going to be stronger in production. But, when it comes to Soul, I don’t think the GPT-6 Soul currently stands a chance against Sonnet 5.5 based off the outputs that we are seeing, but I could be wrong because we’ll only find

04:20 out on the actual release date. And we can see that the sentiment around Anthropic’s recent model drops has changed everyone’s opinion on where they are in the AI lead at the moment. This person says that I’m not sure how Anthropic managed to pull this off. And when he looks into it, not only is Opus 5.5 an incredibly good model, but the rates are also incredibly good, which is one of the plain points that people had with using Anthropic’s model, but clearly they have made some changes that made Opus 5.5 more usable for a lot of

04:51 people. And he also mentions that Opus now appears to be an outstandingly efficient model, which is something that you probably wouldn’t expect people to say about Anthropic’s model lineup. That’s something that they’re not really known for. They have always been known for creating the, you know, some of the best models out there, the smartest models, but efficiency was something that they always lacked in. But, with Opus, it looks like they have answered that question. And this person ends off, “It makes you wonder how much more efficient Sonnet 5.5 will be.” I think this is going to be the more efficient

05:23 model out of Opus 5.5. It’s always been like that. And I do expect Sonnet 5.5 to be model that most people will be using anyways on a daily basis once they update the Sonnet lineup. So, yeah, the Sonnet 5.5 coming very soon. Is it going to live up to the hype? I think so, but we’ll find out pretty soon. Now, at the same time while all this is happening, we have some internal news from OpenAI. There’s a lot happening with OpenAI at the moment. Obviously, they have their dev day. They have this whole agent leak and like, you know, the whole hugging

05:53 face incident plus other incidents like that. Even hacking the Australian government files and things like that, which is interesting. But, one of the heads at OpenAI says that right now the company is spending 80 to 90% of their time or research on GPT-7 and GPT-8 and beyond. So, OpenAI obviously is confirming that they are working towards creating a stronger model. We know that because current model base right now is code named Doug. The models that we have currently, the Astra model and probably even the Soul model, they’re all based

06:25 off of this model base called Doug. There’s a newer model base, which is way more capable than obviously Doug, called Bell. And that is probably going to be GPT-7 and GPT-8. And the way they have been doing it, their playbook is basically build a frontier model, then use it as a teacher, then distill that model into tiny cheap models. And that’s how we probably have models like Luna. Because obviously, you have a strong model like Astra, and then you can distill that model, make it more efficient, make it a little bit, you

06:55 know, less intensive than the larger base, and serve it to people at a cheaper cost. That’s what they have done with Luna. That’s what they have to done with the Soul model. Because we see that with Astra, it’s Soul and Luna, their main selling point is that they are more efficient versions of Astra and cheaper versions of Astra. And that’s probably what we’re going to be expecting from them. And they are working towards creating obviously a new base for GPT-8 and beyond. But, GPT-7 is something I feel like they already have ready, and

07:25 they are probably just fine-tuning it, doing some more training on it before they release it. Because I don’t think Anthropic is going to be number one for a long. The way the space works is that somebody’s number one for a couple of seconds, and then the next lab is ready to drop something, and then, you know, it just goes on and on. So, GPT-7 is going to be a model I think is going to outshine current model lineups. Obviously, OpenAI’s own model lineup, but also in Dropic’s model lineup, because I think after Opus 5.5 and, you know, Sonic 5.5, they have some internal model name called model two. I don’t

07:59 know when they will drop that because they did talk about pacing the frontier. So, we don’t know when both of these labs are going to drop the next generation of models, because those generations of models are going to be, obviously, much stronger compared to anything that we have seen. And as this person says, they are trying to squeeze as much as possible out of the same pre-training base, which I mentioned Doug. When we start to see a newer pre-training base bell, that is when we’ll see the model that OpenAI is producing to match Opus 5.5 or also overtake them, because I think, as I

08:32 mentioned, Anthropic is already working on a newer base. OpenAI is still working on Doug, so that is why we see some sort of discrepancy between the two model lineups that we are seeing at the moment. But, it this space moves fast. This is going to change very soon. Now, at the same time all this is happening, we have some information from China, because it looks like new open-source or an open-weight models are coming from the space. Number one is going to be Qwen 4. It looks like Qwen 4, if we look at the outputs that we see on screen,

09:03 these are some leak generations, and these generations, when it comes to Voxel, they kind of look way competitive now, even this test. Obviously, Qwen is going to be a model that might not take number one spot, but as this person says, it might get close to Opus 5.5 and Astra. I don’t know if it will overtake it. Based off the results so far, the Voxel ones give me high confidence, but if I look at this one, the pelican riding the bike one, I don’t have some high confidence, so I need to see more tests before I make my decision on Qwen 4. But, the Qwen model lineup has always surprised everyone, because they are

09:38 obviously open source models. And the people who are using open source models, most of the time they’re not concerned with having the best model. They want a model that is competitive. They’re wanting a model that they can deploy locally. So, they’re going to be happy with Klein 4. And people are saying currently that the front end inside, when it comes to front end coding, this model is pretty close to Opus 5.5 and Astra range. So, in some areas this model will be competitive. I guess front end coding is one area that currently this model is expected to match these other models. We also know that another lab from China that is working towards dropping

10:11 something new is the MiniMax lab. And they had this leaked model, Space Bunny. Well, it was called leaked, but I think the model itself just told everyone that it was going to be the MiniMax M3.1. And both of these models that you see on screen, MiniMax M3.1 flash and Space Bunny, some people are projecting that these might be two different models. Originally, we know that MiniMax team was going to drop something new. But when you compare the outputs that you see with Space Bunny and then MiniMax M3.1, they’re different. So, we don’t know exactly if the MiniMax team is going to drop two new models, the flash

10:43 variant that we see on screen, and then like the Space Bunny model, which is going to be maybe they’re just MiniMax M3.1. But I just wanted to show you guys this just in case the lab decides to drop something new upcoming week, which I think is going to be the case most likely. We also have some news and I think I made a dedicated video on this as well, where DeepSeek version 5 is supposedly beating Astra right now. This model is something that the DeepSeek team is working on, and they’re saying that this is going to be an imminent launch. The company’s founder calls it the biggest bet that the company has

11:14 made, meaning the DeepSeek team has made. So, they’re going all out on this model as well. This is going to be a 2 trillion parameter model. This is the rumor. Originally, it was 3 trillion. I guess it’s 2 trillion now. And >> [snorts] >> they’re saying that this is going to be the first DeepSeek model to train fully on Huawei Ascend chips instead of Nvidia. And it needs roughly four times more Ascend accelerators than Nvidia equivalent to hit the same training skill. Now, one thing to note is that obviously these are leaks, but one thing I want to show you guys is that I got some new information that China might be working on lifting the ban on Nvidia

11:48 chips at the moment. Currently, if you guys are not aware, China has a ban on those chips, so that’s why all the labs have been using obviously Huawei Ascend chips, but it looks like the compute constraints are forcing them to actually remove that bit. Because there is some information that China is now considering letting its biggest AI companies back onto Nvidia hardware because they’re running short of compute. And they’re asking all these labs, Alibaba, ByteDance, how many new RTX Pro 5500s they want and what they’ll use them for. So, yeah, it looks like things are changing in China as well

12:20 because obviously China is working hard to compete with the frontier labs. So, and that is why they put an emphasis on working on using their own chips, Huawei Ascend, but I think this chip shortage is getting to them and they need support. And Nvidia’s chips are probably going to be the ones that they’re going to use. And obviously this is early news, but this would be a complete 180 turn for China, but who knows? I think this could lead to a couple of things. If the Chinese labs, the Chinese biggest labs get access to Nvidia chips, it’s going to speed up development in China and make them more competitive. So, it

12:53 might work out for China, they might have to obviously, you know, use the Nvidia chips, but in the long run it might be good for them especially as they train these stronger next generation of models. 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 as 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.