So DeepSeek version 4 Pro is officially out today. Now you might be confused because this model was
technically out but it was in general availability but this is the official release meaning that they
have trained the model a little bit more and produced a stronger version of the model that
is available today. Now if I were to summarize today's release in one simple sentence it is that
the price to performance is becoming a very very important thing because not only do we have a new
model from the DeepSeek team, the SpaceX team also dropped Grok 4.6 and both of these models
are competing with the best frontier labs at a fraction of their cost. Let's start with Grok 4.6
and one thing I'm going to say is that I'm genuinely surprised by the SpaceX team. I'm not
trying to glaze them or Elon Musk or anything. I was just very critical of this lab. In 2025,
they dropped Grok 4 and other models like that but I wasn't really you know mind blown with their
performance because they're pretty they're pretty subpar compared to any of the other labs out there
but in 2026 it looks like things have kind of changed a little because grok 4.5 was quite
competitive based off what it cost and grok 4.6 is actually not too bad if we take a look at the
benchmarks for example if we start with the artificial analysis intelligence index this
model achieves a 61 and fable 5 is at 62 now what's really important to remember is that once again
this model is quite cheap compared to Fable 5. This is about $2 per million input tokens
and $6 per million output tokens. And Fable 5 sits at $10 per million input tokens and $50
per million output tokens. And this is why you start to appreciate this release a little bit
more. It might not beat the performance of the best models. It's matching them. Even GPT 5.6
sold, which is a pretty capable model, and this is set at max on the artificial analysis index.
this model achieves 61 and grok 4.6 61 so it ties it and it's much cheaper and then even on all of
these other benchmarks for example the gdp valve one it actually beats fable 5 which is at 1741
and then gpt 5.6 it's at 1728 i'm not sure why they didn't choose opus 5 as well but i guess
they wanted to choose the quote-unquote strongest model lineup from each lab and they chose fable
5 for anthropic which is fair and then deep software engineering one which is a critical
benchmark this model doesn't beat fable 5 or gpt 5.6 soul but it gets close to it it's 65.9 and
fable 5 sits at 70 but if you're getting results that are pretty close and the model is five times
cheaper i wouldn't be too disappointed with this result and then same thing with cursor bench 3.2
the model achieves 69.9 funny number and fable 5 sits at 70.5 so once again closer and grok 4.6
beats GPT 5.6 soul. The same thing with the Frontier Code, it gets close to Fable 5 beats
GPT 5.6 soul. So yeah, this model is actually available in Cursor. So the partnership with
Cursor or I guess the acquisition has really helped SpaceX make some strides in the AI space
this year. And Grok build is something that, you know, maybe not a lot of us have been using so
far, but it's probably going to be another platform like Codex or Cloud Code that we start to use.
but obviously cursor is quite strong as well so you have options available for you to use them in
both and one thing to note is that they're offering two times usage inside grok build and cursor for
the first week so if you just want to try it out see what you feel about it then you know it might
be worth trying it out right now because you get double the usage in the first week now if we were
to take a look at some of the outputs that people have been generating with grok 4.6 what we're
looking at right now is a Falcon 9 booster return sequence simulation and this was done in a single
HTML file and as I said if you expected to get this type of output from Grok in 2025 you would
be kind of surprised because you wouldn't expect something like this to be generated with Grok but
now it looks like we have to start taking the Grok team a little bit more serious because
this output is quite competitive and obviously we're just looking at a simulation and we're just
basing it off of a visual representation but if the model is able to produce something like this
consistently then I would expect a lot of people to start adopting Grok because number one it is
cheaper than the other labs at least at the moment because we don't know if this pricing strategy is
going to be sustainable for the SpaceX team in the long run but at least for now their models are
definitely cheaper compared to the others and as I mentioned this model excelled at the artificial
analysis index. This model jumped to probably number four model. It's tied to GPT 5.6. It's
pretty much similar. So you could say number three as well. But the models before that are
Fable 5 and Opus 5, which are only above the model by about a 1% or a 2% difference. So yeah,
even on this intelligence index, which if you're not familiar with, has nine evaluations. So on
all of these evaluations, it's kind of matching almost Fable 5 performance, which is crazy to see.
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. And what you see
on screen right now is a racing game that Grok 4.6 built. And based off of this post, the model took
about one minute and it was a five word prompt, which was a create a simple racing game in HTML.
So if you're able to generate something like this easily using Grok 4.6, I think a lot of people
will be happy. And this is a more detailed analysis of what it costs to run GPT 5.6 on
the artificial analysis index and what it produced, meaning the output. Both of these models, if you
remember, scored 61 on the artificial analysis index. Now to run the whole test with GPT 5.6,
it costs about 2.8k. And then with Grok 4.6, it costs about 1.1k-ish. And this tells you that
you're getting similar level of performance at half the cost. So yeah, this is a big release for
the SpaceX team because they just proved once again that they are a lab that you seriously
start into considering especially in 2026 and I'm going to talk more about DeepSeek version for Pro
GA but basically what we're seeing today is that both of these releases kind of emphasize the fact
that performance and all above that is the price at what you're getting for that performance is
becoming more and more important for all users because we see many labs now focusing on creating
the best model at the cheapest cost. Last year in 2025, most of the labs were just focused on,
I would say, creating the strongest model. Yes, cost was important, but I think most of the times
the frontier labs, meaning OpenAI Anthropic, were kind of more lenient on that fact because they
didn't have as strong of a competition. Intelligent models that are maybe not always ahead of OpenAI
Anthropic, but match their performance at a fraction of the cost. So yes, price to performance
ratio is becoming a critical I would say indicator in 2026. Now this is the updated benchmark chart
after the release of the new model and one thing you'll see across the board is that it matches
the top level performance of many of the models. The one thing interesting over here is that they
haven't put Opus 5 here for some reason. There is Fable 5 here that we can compare this model
against but one thing you'll notice is that DeepSeek, remember this model costs 43.5 cents
per million input tokens and 87 cents per million output tokens while the other models all over here
are way more expensive than that. So the first thing if you look at Terminal Bench 2.1 the model
scores 87.9. The older version of the model was 72.1 and the flash version was
of the model was 72.1, and the Flash version, which we got last week, was 82.7. And what's
crazy is that Fable 5 is 88. Yes, 88. So this model is only 0.1% behind Fable 5. And then on
the Cyber Gym which is Cybersecurity the model actually beats Fable 5. Fable 5 sits at 83.1
while DeepSeek version 4 Pro the new one that we got today is at 83.3. Now if this tells you
something is that this model is once again really geared at Cybersecurity. We saw the Flash model
also be geared towards Cybersecurity and becoming a model that was quite capable in that area
and we're seeing the same thing today with the new version 4 Pro which sits at 83.3 outperforming
Fable 5 on the benchmark. On deep software engineering this model is not beating Fable 5
because Fable 5 sits at 70 but the model achieves 62.7 which is ahead of GLM 5.2 and a bit behind
Kimi K3 which sits at 67.5 but one thing again this is a big jump compared to the preview version
of the model which was at 12.8 so yeah they have really trained this model and have improved it on
the back end because we can clearly see in this deep software engineering benchmark. And there's
couple of other benchmarks but one key benchmark that this model excels at is the automation bench
where the model achieves 31.8 and fable 5 29.1 so once again the model outperforms fable 5 now this
is a fraction of the cost as i mentioned this is 43 cents and 87 cents for input and output
respectively versus fable 5 which is at 10.50 so yes if we look at the terminal bench for example
we are seeing this model achieve a result which is 0.1 behind the best model out there fable 5
and the pricing is insane to look at 43 cents versus 10 dollars 87 cents versus 50 dollars per
million input tokens so when we turn that into a fraction this model is 57 times cheaper and earlier
last week i think i made a video about how deep seek version 4 pro was going to achieve a result
like this because there was an investor report that leaked and when i saw the pricing where it
said that it was going to match Fable 5 or even outperform it and be at 57 times cheaper per cost.
When I was reading their investor report, I didn't really believe it. But today it is proven because
at least on the benchmark so far, yes, I'm just saying the benchmarks, we are seeing similar level
performance. Now, if this model actually performs like that in production, it's a little too early
to tell yet. We still would have to give a couple of weeks and see how it's performing in the long
run. Because sometimes on the release date, the models perform quite good. But over time,
they kind of deteriorate in their quality. So we hope DeepSeek doesn't do that. Historically,
they haven't done that. But let's just see. Just going to put it out there because right now we're
basing this off of benchmarks. 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 Universe of AIZ
as well. Until then, I'll see you guys in the next video.