The Open AI team is almost ready to drop

GPT Astra with a new base model code

named Bell and this model is quite

competitive. And not only that, it looks

like the Google DeepMind team is ready

to drop Gemini 4 in early September. So,

let's get into it. So, it looks like

Open AI is ready to release GPT 6 Astra,

which is probably going to be a model

that we might be getting very soon and

this model is supposed to be much

stronger than we had all expected.

[music] Originally, this model was using

a model base that was code named Doug.

Now, this was a competitive base model,

but it looks like Open AI continued

doing some further RL training on the

model and now are using a new model,

which is called Bell. Now, Bell is the

successor to Doug. So, think about Doug

and a stronger version of that. That is

going to be Bell. And this is going to

be the base for Astra and basically GPT

6. Now, Bell is a giant pre-chain. We

saw originally in Open AI owned blog

posts and couple of leaks earlier on how

they were training something supposed to

be much more competitive. They're

training something that they said that

they hadn't done in a long time, a much

larger pre-train and that is what has

produced Bell. And Bell is supposed to

be a model with over 10 trillion total

parameters. Now, that seems like a lot,

but this is going to be similar in size

to GPT 4.5 and I think that was probably

the last pre-train that did if my memory

serves me correct. And this is supposed

to be the starting point, if I

understood this correctly, for a AGI

level model that is expected sometime

end of this year. So, I can tell you one

thing that Open AI is clearly

accelerating development of this new

model and this is going to set the tone

for what's coming at the end of this

year and people are saying the model

that's supposed to come at the end of

this year is going to be AGI level. Now,

AGI is something that a lot of us

probably have heard of. A lot of us

probably have debated what does AGI mean

and currently still there is no singular

definition from it, but if people are

expecting this model to be AGI

threshold, one thing I can tell you for

sure, which is going to sound like a

no-duh moment, is that it's going to be

something that is nothing like what we

have at the moment. Something much

stronger than Fable 5, something much

stronger than Opus 5, something much

stronger even than the next upcoming

model from OpenAI, which is GPT-6. Now,

this has put some pressure on Anthropic

because OpenAI believes Anthropic may

have no good response to the upcoming

model GPT Astra, and that is why there

is a lot of competition and a little bit

of pressure on both sides of the party

because OpenAI believes that Anthropic

does not have a model that, you know,

they'll be able to compete with, and the

same goes for Anthropic. But at the

moment, based on what I know so far, it

looks like OpenAI has some sort of a

lead because Anthropic doesn't seem to

have any sort of model like Bell being

in the works. Obviously, they both

released blog posts where they said that

their models are much more capable, and

because of that, they had some concerns

about their security levels, and because

of that, there have been some delays

from both labs, OpenAI and obviously

Anthropic. Anthropic is working on a new

model, Fable 5.1, but it's not we

haven't heard anything like, you know,

Bell or any news about them kind of

releasing a model that is going to match

that AGI threshold level. We haven't

heard any of that from Anthropic. The

only lab we have heard about that from

is OpenAI. But now, what's really

interesting is that this lead that

OpenAI seems to be taking over Anthropic

is not because they have more research

or better scientists, it's because the

biggest constraint in AI at the moment

is compute, and Anthropic seems to be

struggling with compute, and because of

that, it seems to me that OpenAI has

been able to take the lead at the

moment. Now, Anthropic is obviously not

going to stay silent. They have a deal

as well with the SpaceX team, and they

are using the supercomputers from the

SpaceX team, but obviously it's still

early because we haven't seen the

results or the benefits of that deal

translate yet cuz at the moment they are

still kind of bracing for the end of the

year as this person mentioned because

they were expected to IPO in September,

but the way things have been moving at

the moment, I don't know if that's going

to be a successful IPO because clearly

at the moment in my opinion and probably

yours as well, Open AI has an advantage

over them. Before we continue, we just

launched the Universe of AI newsletter.

If you want to stay on top of AI news

without having to hunt for it, link is

in the description. Don't miss out. Now

at the same time while all this is

happening, one thing that we got to know

is that Open AI is not only limiting

themselves to their models. They are

working in creating their first chip,

which is jalapeno, which I love the

name. I love jalapenos as well. But

anyways, that's their first custom

inference chip and they've been testing

it in the system around it. And this has

allowed them to show major advance in

couple of things. Number one, more

intelligence from every watt and faster

responses. And then number two,

delivering both higher throughput and

lower latency in one architecture

without sacrificing efficiency. Now

efficiency is something that is becoming

a big concern for a majority of the labs

because all the labs are able to now

kind of produce capable models. People

are kind of happy and satisfied with the

results that they're getting, but what's

kind of becoming the pain point for

majority of the customers of these labs

is how fast these models are, how

reliable they are. And number two, is

how cost effective they are. If you're

able to deliver more intelligence from

every watt, that usually translates into

more output per dollar. And more output

per dollar is becoming a KPI or a metric

that a lot of people have been

monitoring, including myself. If you

look at Deep Seek version 4 Flash or

Deep Seek version 4 Pro, the reason why

that model is so successful and loved by

a lot of customer is that they're able

to match performance of the Frontier

Labs at a way cheaper cost and that is

okay for them. Even though they might

not be the number one model on the

leaderboard, but if they're able to get

performance as close to the other labs,

then that is a win for a lot of

customers. So, if OpenAI is not only

working on developing the stronger

model, but is also focused on developing

a stronger infrastructure for those

models, then that allows them to not

only excel in the model development

race, but also vertically helps

accelerate in the AI space. And that is

something that we have been seeing from

OpenAI from the beginning. And now with

this new chip jalapeno, it looks like

that is definitely the case. And in the

blog post, we actually got some more

information about how capable GPT Astra

is. Now, this model is already doing

some serious work internally.

Internally, when the team was using

Codex with GPT Astra, the chip team

brought three open weight models that

weren't even in the original jalapeno

production plant. So, Astra is already

helping program OpenAI's own

infrastructure and this model is helping

OpenAI set the foundation for the next

level of models coming from the team.

Which is crazy to me because as I said,

if you are investing only in model, you

will be kind of limited with your

compute and things like that. But if you

are OpenAI who's not only investing in

producing the stronger models, but also

producing better infrastructure to

support those models, you will excel in

the long term and that is what we have

seen OpenAI doing. Now, originally in

2025, the lab that I thought that was

going to be really successful in this

space was going to be Google DeepMind

because they have a wide distribution

footprint. They had did have strong

capable models and we did expect a lot

from that team and they do have the R&D

capabilities as well, you know, the

infrastructure to support large model

scale deployment. But we have seen that

lab kind of slow down and it's mainly

because of internal organizational

structure or things like that that we,

to be honest don't have a clear

understanding of. But that lab has

slowly slowed down. But currently at the

moment there seems to be some Gemini 4

leaks and they look kind of optimistic.

So Gemini 4 has just been leaked and

this model internally is beating GPT 5.6

so plus Fable 5 on internal evaluations,

which is a huge win for Google DeepMind

because if that is the case then it

means that Gemini may once again be back

in the competition for frontier labs. At

the moment it's OpenAI and Anthropic and

then anyone of the models from China. It

could be Deep Seek, it could be Kimmy,

it could be the GLM team. Whatever the

case is, obviously those are not number

one or two. They're still number three

level type of models, but Google

DeepMind has been overtaken by many of

the labs from China. But if Gemini 4 is

actually able to beat GPT 5.6 so and

Fable 5, it comes back in the

conversation. This model is expected to

come around sometime in the first week

of September. But what we do know from

the Google DeepMind team is that there

have been many delays. So I wouldn't be

surprised if this model is once again

delayed and we don't see a model from

them. But I really, really, really hope

that they're able to release this model

because we haven't heard anything as

strong from the Google DeepMind team and

Gemini 4 could be that answer. And this

is significantly larger than anything

before, which was something that was

expected. It has 1.5 million context

window, which is kind of the standard

for many of the frontier labs now. So

this is not something that is, you know,

very different from anything else that

OpenAI or Anthropic is doing. This is

supposed to have major upgrades in

browser capabilities, terminal and tool

use. So yes, it looks like Google is

trying to ramp up their development.

They have been slow for the early part

of this year, but it looks like they

want to end up strong in 2026 and Gemini

4 just might be that answer. 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 universeofaiZ as well.

Until then, I'll see you guys in the

next video.