GPT-6.1 Terra, Fable 5.5, Gemini 4 Carbon + DeepSeek US BAN & More!

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: GPT-6.1 Terra, Fable 5.5, Gemini 4 Carbon + DeepSeek US BAN & More! · Channel: Universe of AI · 16:08 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI governance and regulation, AI models

Summary

The video surveys rumored upcoming models from Anthropic, Google DeepMind, and OpenAI, including Fable 5.5, Gemini 4 Carbon, and GPT-6.1 Terra. It also covers the release of MiniMax M3.1 Flash, highlighting its reported speed, cost, and benchmark results, and discusses speculation about a new GLM model. The host reports a US directive for civilian federal agencies to remove DeepSeek products and describes Anthropic’s proposed public benefit fund and Elon Musk’s stated plan to give SpaceX shares to American children.

Key points

Anthropic is reportedly preparing a new model, possibly Fable 5.5, with stronger intent understanding and more creative reasoning.
Google DeepMind’s Gemini 4 Carbon is described as an internal checkpoint under testing, with reports that it may target coding performance comparable to Claude Opus 5.5.
OpenAI is reportedly preparing GPT-6.1 Terra, a mid-tier model expected to be priced below GPT-6.1, while the timing and capabilities of a possible Astra update remain uncertain.
MiniMax released M3.1 Flash, which it says delivers frontier-level performance on selected tasks at one-third the cost and three times the speed of M3.
The video notes M3.1 Flash’s reported benchmark results, but cautions that comparisons with other models are limited and task-dependent.
A research paper associated with z.ai’s Professor Tang Jie prompted speculation about a forthcoming GLM 5.5, possibly nicknamed GLM Omni, with integrated image and text capabilities.
The US Department of Homeland Security reportedly directed civilian federal agencies to remove DeepSeek and third-party products built on it and stop purchasing them.
Anthropic plans to seed a public benefit fund with 5% of its shares, with further contributions tied to increases in the company’s value.
Elon Musk said he plans to give one SpaceX share to every child in America next year.

Chapters

Start	Chapter	Summary
00:00	Rumored model launches	The host previews potential upcoming releases from Anthropic, Google DeepMind, and OpenAI, alongside other AI industry news.
01:34	What Anthropic’s next model may do	The video describes claims that Anthropic’s upcoming model could better infer user intent and offer less conventional reasoning.
03:13	OpenAI Terra and Gemini 4 Carbon	The host discusses reports about GPT-6.1 Terra, its expected price tier, and Google’s internal testing of Gemini 4 Carbon.
08:04	MiniMax M3.1 Flash	The video reviews the new model’s reported benchmark scores, pricing, speed, and comparisons with DeepSeek and Claude.
11:52	GLM speculation and US DeepSeek restrictions	The host discusses speculation about a new GLM model and a reported US directive affecting DeepSeek use by civilian federal agencies.
13:59	Public benefit plans	The video outlines Anthropic’s proposed public benefit fund and Elon Musk’s stated plan to give SpaceX shares to American children.

Quotes

the pressure on OpenAI is just going to get a little bit more intense — 03:13

it’s 1/3 the cost of M3 and three times the speed — 08:37

The US Department of Homeland Security is ordering civilian federal agencies to remove DeepSeek — 12:56

they’re going to start this off with 5% of their shares — 13:59

he’s going to give one of my SpaceX shares to every kid in America — 15:37

Concepts and entities

Anthropic, OpenAI, Google DeepMind, GPT-6.1 Terra, Fable 5.5, Gemini 4 Carbon, MiniMax M3.1 Flash, GLM 5.5, DeepSeek, Anthropic Public Benefit Fund, Elon Musk

Full transcript (original language)

00:00 According to some new information, it looks like Anthropic, OpenAI, and Google DeepMind are all working towards dropping a new model sometime soon. And it looks like the AI space is not slowing down anytime, either. We also have a new model from the MiniMax team, as well some interesting information coming from z.ai, and it looks like the US is working towards banning deep seek models at all. Not only that, Elon Musk and Anthropic are working towards giving back some of their money to the general public. So, let’s get into it. So, we

00:31 have some new information from the information, ironically, that Anthropic is planning on releasing a new model this month. Now, that is not a surprise to us. We have been playing around with a newer version of Fable 5.5. We’ve seen leaks, as well. It was just a matter of timing when this model would be coming out. But, it looks like they’re not the only lab that is going to be dropping something new. Gemini 4 Carbon, which was a model that was leaked internally from the Google DeepMind team and talked about over the last couple of weeks, is getting ready for deployment. Now, one thing with Google is that their

01:04 timelines are always, you know, a little bit delayed. So, we don’t know how fast this model will actually come out. But, we do know that RSI has been achieved inside Google DeepMind, and that is allowing them to create stronger models quite rapidly compared to before. At the same time, OpenAI is gearing towards dropping a new model, but it’s supposed to be the new Terra model, which is a mid-level model in their whole AI model catalog. Astra being the top model, then Sol, and then Terra, and then Luna. So, it looks like they’re updating a newer

01:34 version of Terra. So, what exactly are we expecting from this new model? Well, one thing that we do know about the Anthropic model is that this model is supposed to be really good at understanding what you’re trying to say and your goals, your intentions, and your goals, even if you don’t give the model much information. Meaning, it’s getting better at predicting what the the token is going to be compared to before. Next thing is that the Entropic models also better at thinking outside of the box rather than regurgitating common sense advice or solutions, which is interesting because obviously most of

02:08 these models are trained on public data and it’s supposed to basically gather that information and provide you that answer as fast as it can. But then now the models are shifting away from just doing that. They’re also getting better at thinking outside of the box meaning we’re not going to be expecting the same type of answers that you would get. To give you guys an example, the article gives a specific example. Let’s say if you were to ask the model specific information about a stock or just ask like, “Okay, how do I invest in this stock?” Rather than just giving you the generic information like price to

02:40 earnings ratio, things like that, or looking into like, you know, the historical performance, the model might actually come up with its own creative algorithm to predict future stock performance. This is going to be a new layer for these models. If you ask any information to these models, most of the time it’s going to give you public information or give you generic information. It’s not going to try to obviously predict the stock performance, which is something most of these models don’t do when you ask it about it. So yeah, this is going to be a different space completely for these new models and this is all for Entropic. This is the information that we have for the

03:13 upcoming Entropic model, which tells you that the pressure on OpenAI is just going to get a little bit more intense because if Entropic is working towards dropping something like this and OpenAI is dropping Terra, which is a mid-tier model cuz they already have 6.1 soul, then I don’t know how they’re going to be able to compete with them in the short term at least. 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

03:46 you guys are subscribed to the channel. Your support means a lot and now let’s get back into the video. And then, we also know that Google is obviously developing Gemini 4 carbon, which is going to be upgrade to the argon model. Some people are saying that this model, Gemini 4 carbon, is supposed to be an upgrade to argon, but we also have some people saying that this is a more efficient version of argon. We don’t know exactly how this model is going to, you know, be deployed, but we do know some internal leaks, which I’ll show you guys in a bit. And as I said, OpenAI is just preparing capable terra model for

04:21 the release in the coming weeks. But one thing that we do know is that this model is going to be priced below GPT 6.1. So, so it’s very possible that they drop actually GPT 6.1 terra, and this model is outperforming soul on some, you know, benchmarks, things like that, and it’s cheaper, which would be a good win for OpenAI in the short term. But at the moment, if Anthropic is dropping or gearing towards dropping Fable 5.5 or a next level Opus 6 or something like that, which it seems like it, then OpenAI still doesn’t have an answer to

04:53 that level of model, you know what I mean? So, yeah, it’s kind of all over the place. But at the same time, we do see some news like this where Tibo, who’s at Codex, actually posted this today saying that I bought a new domain, and I’m very excited for it. Very Now, this does not indicate that he’s talking about obviously a newer model from OpenAI, but usually when he posts something like this, we do know that there’s probably an update coming from OpenAI. We do know that he is working towards dropping new updates for the next 28 days for Codex features, and so far it has been kind of whatever. It

05:26 hasn’t been anything crazy, but I’ll be monitoring this just in case something new comes out and updating you guys as well. Now, as people were mentioning 6.1 terra, we actually are seeing it in the Microsoft PR, which gives us more confidence that this model is coming compared to something else. And you know, this is usually indication that it’s going to be tested quite soon. We might have some leaks from 6.1 Terra and that will give us some idea of what this model will be like. And as this person says, this model is probably not going

05:56 to compete with Opus 5.5, but does this mean that we could be getting GPT 6.1 Astra soon? Because if they’re updating the whole 6.1 lineup, probably. And maybe 6.1 Astra is able to match Opus 5.5, but at least at the moment a point one update, I don’t know if that will be able to match a newer version of Anthropic’s model like Fable 5.5 or Opus 6. Now, some updates on the Gemini 4 Carbon model. So, there were some leaks. Business Insider actually leaked this model a few weeks back and now it’s coming up again. And what we do know is

06:28 that Carbon is supposed to be an internal Google checkpoint that is being tested on its own Jetski coding platform. So, this model is internally being tested at the moment and they’re saying that this is not going to replace Argon anytime soon. The Gemini 4 Argon model that we do know was built from a earlier checkpoint called Barium B, is still going to be the model that most likely is going to be publicly available, but I feel like Gemini 4 Carbon, I don’t know what the numbering could be. It’s expected to match coding

06:59 capabilities as similar to Cloud Opus 5.5 and you know, some of the employees at Google DeepMind are obviously claiming that this outperforms Anthropic’s models, but that’s just you know, some employees and we don’t have any confirmation about what the models actually going to look like. But as I mentioned, RSI, recursive self-improvement, is already on the go at Google and that is allowing them to drop new models more faster compared to before. So, if Carbon does come out, we should see a match Opus 5.5. So, it

07:31 looks like right now Anthropic is obviously dominating the space at the moment. Google DeepMind has caught up and soon they might be you know, competing for tier number two, but OpenAI at the moment is focused on other things, especially Codex updates for some reason. But let’s see, things change very fast, so it’s very likely that OpenAI could be dropping something new, and then, you know, it outperforms Anthropic, and the whole sentiment changes. Today, we also got a new model that is officially dropped. So, this is an actual model that’s live now, which is Minimax M3.1 Flash. And not

08:04 surprisingly, this was the model that all of us were probably, you know, already testing or trying out, which was the Space Bunny model. But what they did mention is that Space Bunny was an earlier version of M3.1 Flash, so the model we got today was obviously improved based on probably their testing with Space Bunny. And what they say is that M3.1 Flash is achieving frontier-level intelligence across key tasks, such as code work, and they are referencing the Automation Bench score, which is getting 62.2%. And even on coding deep software engineering, the model is achieving a

08:37 74%. Then, it also has good visual reasoning, achieving 91%. But what’s the real story here is that it’s 1/3 the cost of M3 and three times the speed, which is obviously a great thing for many people. And especially if you’re using this model, or if you’re running the open weights on this, this is going to help you a lot. And the pricing they actually mentioned, the API pricing, the flat rate is 15 cents and 60 cents per 1 million input/output tokens, and the cash reads are sitting at 0.015 per 1 million tokens. The model, which

09:11 are actually going to come on October 23rd, so you might have to wait for that. Now, I’m not going to spend too much time on the benchmarks over here, but you can clearly see that one area that this model excels in is obviously on the Automation Bench, it achieves the highest score of 62.2. But obviously, it’s not being compared to Opus 5.5 or GPT, you know, 6.1 Soul or Astra, any of those models, but it is being compared to the newest Flash model that we got from DeepSeek. So, this model does beat it on, for example, the automation bench

09:42 cuz that’s the only benchmark that we see. But, we don’t really see any of the other benchmarks being applied to the DeepSeek model, so it’s hard to compare. So, I would use this M3.1 flash model against DeepSeek’s flash model to really understand which model you would like to use in your workflow. The benchmarks don’t really tell you much. Also, it’s pretty, you know, all of these you’re not really going to compare it as much because this is a flash model. You’re not going to use it against Opus 5. And even on the benchmarks, the model still outperforms M3.1 flash. But, it’s just good to see that we have access to a

10:13 flash model like this, which I said once again is 1/3 of the cost of the old M3 model and three times the speed. Now, somebody ran a test of this model against DeepSeek version 4.1 flash, the latest one that we got 2 days ago, I would say. Yeah, 2 days ago. And MiniMax took about 40 minutes and cost $0. So, it’s probably in the cents, it’s probably not going to be $0 completely. DeepSeek took 50 minutes and cost 40 cents. And obviously, you can’t really make out which model is better. It’s all going to be on preference based off of

10:46 this simple benchmark. I would say like this is all up to the person at the end because both of these have their strengths and weaknesses. So, I would say both of these models are capable. I wouldn’t really try to like, you know, pick on one of these outputs and say like this one is better or not cuz at the moment I can see that it’s pretty neutral. Like, they’re both are producing high-quality results compared to before, and this is good for the AI space in general. Now, somebody also compared this model to Claude Opus 5.5, and you can see the results on the screen. So, M3.1 flash,

11:20 which was about 88 times cheaper, cost about a cent. And Opus 5.5 cost about $1.22. And the outputs are obviously, this is a simple front-end test. So, in this test like these, the front-end side, things like that. This model is kind of matching or, you know, almost on par with the Opus 5.5. And if it’s 88 times cheaper, that’s not a bad deal, right? Now, at the same time while all this is happening, it looks like another lab from China is getting ready to drop something, which is GLM 5.5. And the reason why people are speculating this

11:52 is because Professor Tang Jie, I I hope I’m pronouncing that right, s- uh paper accidentally exposed that there’s a model that they have that it can simultaneously understand, generate images with interleave text image alignment boosting accuracy directly from 12.57% to 26.86%. So, they’re talking about a newer version of a model. So, visual generation is not just going to be an output for this model. It’s going to be a true channel for knowledge acquisition, according to this paper. All right? And we kind of had some idea

12:26 that GLM was going to drop something new, GLM 5.5. The nickname is GLM Omni. That’s what people are guessing what this new model’s code name might be, and it might come out as GLM 5.5. But obviously, this is still a little bit early. Once I get more news on this, I’ll definitely let you guys know. Now, what’s interesting is that while all of this is happening, the AI space, once again in China, is getting a little bit more active. The US Department of Homeland Security is ordering civilian federal agencies to remove DeepSeek and

12:56 third-party products built on it and stop buying them. So, it looks like the US is once again pressuring more companies in the US to stop using Chinese models like DeepSeek and using their products. And I think this is obviously going to be a response to China’s ban on the text exports. This is probably just a political play. Also, obviously, the US government is concerned that these models are going to be used to collect consumer data, used to collect, you know, information about US companies, things like that, and

13:26 going to be a security threat. So, I’ll just leave it at that. I’m not going to say what is right or wrong, but this might be happening. So, if you are in the US and if you are using deep seek models, obviously this is targeted towards civilian federal agencies. So, it’s not going to be like normal people, but who knows? Maybe this gets extended and a ban restricts even normal people to use the deep seek models. The rate of AI in 2026 has increased drastically. And it looks like many of the parties, including Anthropic, are fearing that

13:59 the rate of AI development is going to displace a lot of people, cause maybe some harm to society as it’s progressing. And because of that, it looks like they’re committing a portion of Anthropic share to a new public benefit fund, which is called the Anthropic public benefit fund. And this is going to be a entity, meaning a company basically dedicated to the public benefit. So, what they’re doing is that they’re going to start this off with 5% of their shares. And for every 50% increase in the company’s value over

14:29 a baseline that they’re going to set, they’re going to drop an extra 2% in there. So, they’re going to use that money for the public. Now, we don’t know exactly what that money is going to be used for. Uh you know, it’s quite unclear. Obviously, this is pretty new, but they are going to drop more information on this as the months come. And what this is basically saying is that internally Anthropic realizes that AI is going to change a lot of things. It might displace some people in the workforce, and in the short term there could be some hiccups that we all have to navigate. And this is why they are

15:02 gearing towards, I guess, making sure that they are not only benefiting society as a whole, but as we transition into this new world, we have some sort of support for people that are going to obviously be displaced because of this. And it looks like they are doing this through a public fund. We obviously don’t know what OpenAI is doing or any of the other labs, but I wouldn’t say that this is just a marketing gimmick. I think this is a good play by Anthropic. We’ll see in the long run if this is actually beneficial to the public. Now, what’s surprising is that on October 19th, which is today, at

15:37 3:52 a.m., Elon Musk said that next year that he’s going to give one of my SpaceX shares to every kid in America. This will be the first of many. So, it looks like even, you know, Elon Musk is working towards helping the public with the new rate of development with technology. And he’s actually the one who kind of brought up universal basic income a couple of months back, and it looks like now he’s going to give everyone SpaceX shares. So, if you identify as a kid in America, as this person said, you might be entitled to a

16:08 share of SpaceX next year. 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.