Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
Codex's new ultrafast mode delivers real speed, but only small tasks justify its usage cost. It runs at up to 300 tokens per second and cut one resource guide task from 13 minutes to under 2 minutes for roughly triple the weekly usage, with near-identical output. On a heavy 145GB video task the advantage shrank to 2.5x while usage jumped 4x, which is why the presenter's verdict is to treat ultrafast as an emergency lever rather than a default setting.
Key points
Ultrafast mode is available only on Codex's new $500 per month plan.
Ultrafast requires GPT-6 Astra and is coming to GPT-6.1 Soul soon.
Marketing claims 8x standard speed and up to 300 tokens per second.
A 145GB showreel task used 17% of weekly usage with ultrafast versus 4% with standard.
A resource guide task ran in 1m43s versus 13 minutes with nearly identical 7-page outputs.
Tools mentioned
Techniques
- Agentic showreel creation from raw footage
- Resource guide generation from a YouTube video
- Head-to-head comparison of standard and ultrafast modes on identical prompts
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
Codex's new ultrafast is really fast. Here I just asked what did I talk about in this week's community Q&A call and you can see how fast it's going to go find the right call, get the transcript
and summarize that up. This was a 54minute call that it had to look through and give me a summary of and it did that in 20 seconds or sorry 19 seconds. So ultraast is only available
with GBD6 Astra and you can see we had standard fast but now we have ultraast. Now the thing about ultraast is it's only available on their new plan they just announced which is $500 a month.
Here you can see the fastest work and codecs. This isn't available on $200 a month or $100 a month. Now, what actually is ultraast? It's the fastest way to build with Astra in codecs. It's
eight times faster than standard and four times faster than the regular fast mode. So, obviously today I'm going to show you guys a few examples of me running ultra fast versus standard. And
we'll compare the quality of the outputs as well as how much it ate of my weekly usage. It basically just moves a lot faster up to 300 tokens per second, which is really, really fast. That's
faster than any human could read. But like I said, it's on the 500 bucks plan or on the API and it's also going to be coming to GBD 6.1 Soul soon, which is great because Soul is also a lot
cheaper. So hopefully the Ultra Fast won't eat as much of my usage because this thing eats usage. Okay, so here's a first example that I want to show you guys. I asked Astra to create me a
60-second showre from my recent trip to Google in San Jose. I gave it two files and these two files combined was 11 minutes of me kind of talking about my experience, reflecting on it. And then I
gave it a folder with these raw assets. And this folder contains over 145 gigabytes of footage to actually comb through and try to tell a story out of. And I told it to do this in a 9 by6
format. So something that could go on Instagram or Tik Tok or, you know, a YouTube short. Now, before we watch the outputs, let me show you guys the stats right here. When we did this Google
video, I did both of these on extra high. So extra high reasoning, and we did one with standard, one with ultraast. And here are the stats. Ultraast ran in about 19 minutes and
cost 17% of my weekly usage, which is absurd. This is the weekly usage of the $500 a month pro plan and standard ran for 50 minutes and cost 4% of my weekly usage. So, we'll dive a little bit more
into that, but first let's actually go ahead and watch the outputs. I'm first going to play the standard output and then I'm going to play the ultraast output and you guys let me know which
one you thought was actually better. Like, man, this energy here is really, really cool. So, the Google campus that we were actually touring here, it was in San Jose. So, it felt like a college
campus. Great energy. There were these Gbikes all over that people were riding around between buildings. You kind of got inside these modern looking buildings and there was so much
character. There was so much color everywhere. That's where we did like the soccer ball test. Yeah, kicking the soccer ball with all of those Google pixels up on the walls to analyze. It
gave you these custom cards. We talked a lot about how much quicker we're all able to move now, which is pretty cool. We the humans have to be in control with things like judgment and clear
boundaries, clear definitions of done and just a clear goal. It's all about how do you find what you're good at and what you're passionate about and find out how in that domain AI makes you a
superhuman. It [music] makes you faster. It allows you to do more in less time. Thank you guys so much. I appreciate it. It was super fun. hope that we get to come back and do it again sometime.
Man, this energy here is really, really cool. But that is one thing I appreciated about Google is that you kind of got inside these modern looking buildings
and there was so much character. There was just so much color everywhere [music] and everyone I talked to there loved what they did and they loved the culture which I thought was really cool.
And then we got to see all these other ways that Google AI is being used in things like retail and inventory and all these different little demos. >> But we talked about how much these AI
skills are going to matter even for people that don't think they're going to be in a techf facing role. It's all about how do you find what you're good at and what you're
passionate about and find out how in that domain AI makes you a superhum. >> [music] >> Huge shout out to Google for having us out. Hope that we get to come back and
do it again sometime. So, I definitely thought that the ultraast output was better. It felt more Google branded. It just felt like it told a better story. You'll also notice
it was kind of annoying in the standard version. It was kind of like cutting off what I was saying before my sentence actually finished. So, it just didn't feel very natural. I was actually very
impressed by how it was able to comb through all this stuff and actually make a story out of it. I mean, this is a lot of B-roll and a lot of stuff to sort of pull together, but I like the story that
it told here. And that was interesting to me because you kind of would assume, okay, it's the same model. It's the same reasoning effort. They're probably going to have a similar type of output, but it
doesn't feel similar at all. It even feels like for this one, I didn't even use Astra. It felt like I kind of maybe I used GBD6 Soul or something like that. Like it felt like a different model, a
worse model, even though it ran for a lot longer. But this is interesting because in this case, ultra fast was about, you know, 2.5x faster than standard. So, not the kind of marketed
8x faster, but I would never really care about the speed if it was going to eat that much more of my weekly usage. That's, you know, 4x more, 4.25x more of my usage. I would not really be going
for that. Anyways, that was kind of a complex use case. This one's a little bit more, I would say, normal or maybe something I would do more frequently, where I'm asking it to create a YouTube
resource guide from a YouTube video, which was the one I just did with Open Eyes Dots versus Meta's Muse. And I did this once again, same exact prompt with standard and ultraast. And you can see
here, standard took 13 minutes and ultraast took 1 minute and 43 seconds. So, this is actually closer to that 8x speed that was marketed. It's still a little bit less. This was more around
seven, but still, this one was definitely much faster. And it was only a little bit more on the usage. So 3% more than 1%. So it feels like to me after a couple prompts that I've been
testing, you know, difference between standard and ultra fast, it is definitely faster. There's no doubt about it. And I would probably use it if I'm in a pinch and I need to do
something actually fast, maybe 5 minutes before a meeting or something like that. Because when you do these smaller tasks that maybe don't require so much input and so much thinking and back and forth
and verification, then you know you're probably going to be in a situation where the amount of usage that you're actually increasing is maybe two to 4x. But of course, as that multiple scales
up, it's going to just eat more and more. And normally when I'm working in codeex or cloud code, I don't really need things to happen that fast, especially as they get more complex. I
like to iterate. I like to, you know, let it run. I don't really care about speed. But if you do, then obviously it's really nice to have this ultra fast lever that you can pull. Just be aware
that you're going to be paying a lot more for it. But that is ultimately how the world works. So it makes perfect sense. But let's actually go look at these two outputs real quick. Okay, so
here are the two resource guides. I'm not going to yet say which one's which because I want you to think like is one of these noticeably better and worth the money that we paid for it or not really.
So OpenAI Dots versus Muse. They both have a link up here at the top. This one has a longer kind of like summary up top. We have number one, what these personal agents do. They both link to
the YouTube video up at the top. As you can see, we've got a scope. We've got a decision rule. Number two is memory, context, and personalization. Memory and personal context. Number three is
communication and ecosystem. Number three over here is channels and ecosystem. So ultimately, you guys can see that these are very, very similar. They're similarly branded because they
both use the same skill and they have similar structure. Number six is dots, strengths, friction, and verdict. Over here is working inside of dots. And then we get over here to some key takeaways
and we have the end of the doc. And they both link back to AIS+ at the bottom. These both ended up being seven pages. As you can see, the key takeaways is longer over here. But the one on the
right also has these useful links which is the YouTube video as well as getting started with dots from OpenAI and a doc from introducing Muse. Do we have any other links on this left-hand side? I
don't see any other links besides just the actual YouTube video it looks like. So anyways, this one on the left, this one was the ultra fast, and this one on the right was the standard. So
obviously, I don't think one is meaningfully better than the other. And I think that's a really good argument to say maybe even this one was a little bit better. But ultimately, they're both
about, you know, 9 out of 10 or whatever. But I wouldn't care to wait 13 minutes and use 1% for an output that's pretty much just the same as me waiting 2 minutes and using 3% of my usage. And
ultimately guys, I do use fast mode a lot, 1.5 speed, and I don't really notice too much of it using a crazy amount of usage. Obviously, it does use more than if you just use it on
standard, but I like to use fast a lot. And then I also like to use fast when I'm on like 6.1 soul and things like that as well. So, if you're constantly hitting your usage, then yeah, it's
probably worth getting up to the $500 a month plan and then maybe playing around with ultraast and just knowing, okay, cool. I have that lever if I need to pull it. But ultimately, I will probably
never be using ultra fast unless I really, really need, you know, my back's up against the wall for some reason. But then I would argue that maybe you should be building other systems to help you
not get in a situation where you're in a time crunch rather than having to pull that as like the last resort. But anyways, I know this one was super quick, but just wanted to show you guys
real quick what that actually looked like. Hopefully, you enjoyed, you learned something new, and if you did, please give it a like. It helps me out a ton. And as always, I appreciate you
guys making it to the end of the video, and I'll see you all in the next one. Thanks everyone.