Download "OpenAI just dropped GPT-5.2... (WOAH)"
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FyB3ly_ZRr5o%252Fmaxresdefault.jpg&w=1139&q=75)
Similar videos from our catalog
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252F2PrYdjtoRRo%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FckFInZ2BXH8%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FeQyAzZboDbw%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FhuszaaJPjU8%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FmEahc3UbSq4%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252Fmpk4Q5feWaw%252Fmaxresdefault.webp&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252Fu1NMxMqoqtY%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi%252FvLwfguLz_qQ%252Fmaxresdefault.jpg&w=1139&q=75)
%3Aformat(webp)%2Fi.ytimg.com%252Fvi_webp%252FzkafOIyQM8s%252Fmaxresdefault.webp&w=1139&q=75)
Video tags
Subtitles
00:00:00
GPT 5.2 is here and yes, it is an00:00:02
incredible model. OpenAI published so00:00:05
much information about it. We got a ton00:00:06
of benchmarks. We have a bunch of demos00:00:08
that I'm going to show you. So, let's00:00:10
get right into it. But first, Flavio00:00:11
Adamo did his famous bouncy balls in a00:00:14
hexagon test and he really took it to00:00:16
another level. Look at this. We now have00:00:18
a 3D realistic looking hexagon. 3D00:00:22
balls. They bounce off of each other.00:00:24
The physics look real. The lighting00:00:26
looks real. When they hit each other,00:00:28
they brighten up for a moment. It is00:00:31
incredibly impressive. Here's a test00:00:33
that Ethan Mollik put together. Create a00:00:35
visually interesting shader that can run00:00:37
in twiggle.app. Make it like an infinite00:00:40
city of Neo Gothic towers partially00:00:42
drowned in a stormy ocean with large00:00:44
waves. Single shot. Look at this.00:00:47
Extremely impressive. The water physics00:00:49
looks real, although it looks low00:00:51
polygon, but that's okay. The buildings00:00:54
look incredible. And yes, it is00:00:56
infinite. And of course, I know you all00:00:58
want to see what are the benchmarks and00:01:00
how does it compare to the other00:01:02
Frontier models. So, Swebench Pro, a big00:01:05
5% jump from 5.1 to 5.2 and now GPT 5.200:01:09
thinking is state-of-the-art. It is the00:01:11
best model on the planet for the00:01:13
SweetBench Pro benchmark. We have GPQA00:01:17
Diamond, which is a science benchmark00:01:19
with no tools. 92.4%00:01:23
that is a 4% bump from 5.1 thinking. And00:01:26
yes, once again, state-of-the-art. We00:01:29
have a new 100% aced it Amy 202500:01:35
benchmark. Completely aced it. Got every00:01:37
single question right. Amy 2025 is a00:01:40
math competition. Very difficult math00:01:43
problems. And 5.2 aced it. That's00:01:46
compared to 95% with Gemini 3 Pro and00:01:48
92.8 with Claude Opus 4.5. Now, here is00:01:52
the big one. Here is the surprising one.00:01:55
Arc AGI 2. It went from 17% on 5.100:01:59
thinking to 52.9%.00:02:02
That is a stunning increase and is by00:02:06
far state-of-the-art for ARC AGI2. And00:02:08
if you're not familiar with that00:02:09
benchmark, it is a benchmark from our00:02:11
friends at ARC Prize and tests the00:02:13
model's ability to learn and generalize.00:02:17
Probably the truest definition of AGI.00:02:20
And even ARC prize posted about it. Look00:02:23
at this. We verified GPT 5.2 Pro High is00:02:26
state-of-the-art for ARC AGI2 scoring00:02:29
54.2% at $15.7200:02:33
per task. That is extremely00:02:35
costefficient. And what's even more00:02:38
impressive is the efficiency improvement00:02:40
as compared to just a year ago. Listen00:02:43
to this. A year ago, we verified a00:02:45
preview of an unreleased version of 0300:02:47
High that scored 88% at an estimated00:02:49
cost of $4,50000:02:52
per task. Today, we verified GPT 5.2 Pro00:02:56
XH High, 90.5, so a better score at $11.00:03:02
$11 per task. That is down from $45,00000:03:07
per task. That is a 390x00:03:10
efficiency improvement. It is not just00:03:13
about getting the best score. It is also00:03:16
about efficiency per token. And yes,00:03:19
both ARC AGI 1 and ARC AGI 2 for GBT 5.200:03:23
thinking both stateofthe-art as compared00:03:26
to the other frontier models, Opus 4.500:03:29
and Gemini 3 Pro. Last GDP val which00:03:32
tests real world knowledge task00:03:35
evaluations. And once again 5.2 2 is00:03:40
dominating at a 70.9%00:03:43
as compared to second place Opus 4.5 at00:03:46
59.6%.00:03:48
Incredibly impressive scores. And by the00:03:50
way, I've started including some00:03:52
mentions of prediction markets like Khi00:03:54
and Poly Market in my videos. Let me00:03:56
know if you like those in the comments00:03:57
below. I find them pretty interesting.00:03:59
So this prediction says which AI company00:04:01
will have the best coding model on00:04:02
January 1st, 2026. And if we look all00:04:05
the way up until about December 5th or00:04:08
6th, Anthropic was by far in the lead00:04:11
and then all of a sudden it switched. So00:04:13
definitely seems like someone knows00:04:15
something. But as of right now, there's00:04:17
an 86% chance that OpenAI will have the00:04:20
best coding model on January 1st, 2026.00:04:24
And we're doing a major giveaway. Let me00:04:27
tell you about it. We just launched the00:04:29
high taste AI bundle with Zapier,00:04:32
GitHub, Microsoft Copilot, Windsurf,00:04:34
Replet, Crew AI, Gemini, and so many00:04:38
more. There are over $15,000 in total00:04:41
prizes up for grabs. All you need to do00:04:43
is go sign up for our newsletter, share00:04:46
it with your friends, and the more00:04:48
referrals that you get, the higher your00:04:50
chances of winning. The grand prize00:04:51
winner wins over $12,000 in credits,00:04:54
seats, and pro plans from all the00:04:57
companies that I mentioned. We're going00:04:58
to choose a winner on December 26. So,00:05:00
make sure you enter. Links down below.00:05:03
Now, to the blog post released by OpenAI00:05:06
for GPT 5.2. They are putting a lot of00:05:08
emphasis on GPT 5.2 being incredibly00:05:10
good and a big improvement from 5.1 on00:05:14
real world tasks, economically valuable00:05:16
tasks. All right, so who gets it and00:05:18
when? Well, 5.2 2 is available right00:05:20
away for paid users. That includes the00:05:23
instant, the thinking, and the pro00:05:25
versions of 5.2 for all paid accounts.00:05:29
Let me show you some examples of00:05:31
economically valuable work that GPT 5.200:05:33
is able to do much better than 5.1.00:05:36
First, workforce planner. The prompt00:05:39
create a workforce planning model,00:05:40
headcount, hiring plan, attrition, and00:05:43
budget impact include engineering,00:05:44
marketing, legal, and sales departments.00:05:46
on the left. This is 5.1 thinking. And00:05:49
as you can see, it is a very basic00:05:51
looking Excel, which actually to be fair00:05:54
is fine. You don't really need it to be00:05:56
very pretty. It just needs to have the00:05:58
right data. But if you look at the right00:05:59
side, it was able to format the Excel in00:06:02
a much more easily readable way. Here's00:06:05
a cap table management Excel document.00:06:08
On the left, we have 5.1 thinking. On00:06:10
the right, we have 5.2 thinking. And the00:06:12
interesting thing to note is 5.100:06:15
incorrectly calculated seed series A and00:06:17
series B liquidation preferences. By the00:06:20
way, these are very sophisticated00:06:22
spreadsheets and specifically formulas.00:06:25
5.1 incorrectly calculated seed series A00:06:28
and series B liquidation preferences and00:06:30
left majority of those rows blank,00:06:32
leading to an incorrect final equity00:06:33
payout calculation. This is a big deal.00:06:36
Being able to trust an AI model to00:06:38
create these cap tables is extremely00:06:41
valuable, but only if they get it right.00:06:44
The cost of getting it wrong is00:06:46
enormous. Potentially millions and00:06:48
millions of dollars, potentially00:06:50
billions of dollars. And obviously, a00:06:51
human is ultimately going to need to00:06:53
review this and sign off on it, but it00:06:55
is still incredibly important to get it00:06:57
right. And GPT 5.2 gets it all right.00:07:00
Here's another example. You are a00:07:02
project manager at a UK-based tech00:07:04
startup called BridgeMind. Bridgemine00:07:06
successfully obtained grant funding from00:07:07
a UK- based organization that supports00:07:09
the development of AI tools to help00:07:12
local businesses. This website provides00:07:14
some background information. So, I won't00:07:16
read the whole thing, but it basically00:07:17
gives it a bunch of information and ask00:07:19
it to create a report. And so, what we00:07:21
see on the left side here, this is one00:07:23
of the slides, the report on the right00:07:25
side, as you can tell, much more00:07:27
beautiful, much more easily readable.00:07:29
And so, it definitely did a much better00:07:31
job. Now, I'm looking right here, and00:07:33
this is a tiny minor gripe, but you can00:07:36
see that we have this rounded edge right00:07:38
here. But for the header, we have this00:07:40
squared off, and they're overlapping.00:07:42
Very minor, very subtle visual detail00:07:45
that I got wrong. Next, let me show you00:07:47
maybe one of the most impressive demos00:07:50
I've seen to date for a coding prompt.00:07:52
First, here's the prompt. Create a00:07:54
single page app in a single HTML file00:07:56
with the following requirements. Name:00:07:58
ocean wave simulation goal. Display00:08:00
realistic animated waves. Features:00:08:03
Change wind speed, wave height,00:08:04
lighting. The UI should be calming and00:08:06
realistic. And look at that. It looks00:08:09
phenomenal. I don't really understand00:08:11
why it looks like you're looking through00:08:13
a port hole or something, but you know,00:08:15
that's fine. But if we change the wind00:08:18
speed, we can see if we turn it all the00:08:19
way down, it's very calm waves. And of00:08:21
course, if we turn it all the way up, we00:08:23
have much more turbulent waves. We can00:08:25
also change the wave height just like00:08:27
so. So now we have really big waves. And00:08:30
then all of a sudden, we can move it00:08:31
down to the bottom and have much smaller00:08:33
waves. Let's turn the wind speed all the00:08:35
way up. This is the calmst ocean you can00:08:36
possibly think of. And then of course on00:08:39
the other side, there we go. And you can00:08:40
also change the lighting like so. And00:08:42
GPT 5.2 also hallucinates less than 5.1.00:08:47
It's a minor improvement, but still very00:08:49
meaningful. Any squashing of00:08:50
hallucinations is definitely worth00:08:52
noting. GBT 5.1 thinking with at least00:08:56
one error in its response. We'll call00:08:58
those hallucinations. GPT 5.2 thinking00:09:00
6.2%.00:09:02
A very welcomed improvement. It also00:09:04
sets a new standard in long context00:09:06
reasoning. So, we get 256K tokens just00:09:10
like we did in 5.1 in previous models.00:09:12
So, no big expansion of the context00:09:14
window there. But what it can actually00:09:16
do with the context window has greatly00:09:18
improved. Look at this. So, as we see,00:09:20
this is 5.1 thinking, and this is the00:09:23
MRCV2 benchmark with four needles, very00:09:26
similar to the needle in the haststack00:09:28
test. And as we can see for 5.100:09:31
thinking, it drops all the way down to00:09:33
42% at 256K tokens. Whereas nearly00:09:38
perfect at 98% for 256K tokens with four00:09:42
needles in the haststack. Now, when we00:09:44
have eight needles, it dropped off a00:09:46
little bit more, coming down to 70% as00:09:49
compared to 30% for 5.1. It's also much00:09:52
better at visual reasoning, which is a00:09:55
very important feature, something that I00:09:56
use all the time. It cuts error rates00:09:58
roughly in half on chart reasoning and00:10:00
software interface understanding. Here's00:10:02
the charchive reasoning, scientific00:10:04
figure questions, benchmark, and 5.100:10:07
thinking, 80%, 5.2 thinking, 88%. So, a00:10:10
nice improvement there. And a quick00:10:12
thanks to Dell Technologies for00:10:14
sponsoring this portion of the video.00:10:15
Dell Technologies has a family of00:10:17
incredible laptops called the Dell Pro00:10:19
Max featuring Nvidia RTX Pro Blackwell00:10:22
chips, which are portable AI workh00:10:25
horses. It comes in 14 and 16in screen00:10:28
sizes and up to 32 GB of GPU memory.00:10:31
Perfect for onthe-go AI workloads. Check00:10:34
them out. Link in the description below.00:10:36
And of course for computer use we needed00:10:39
to be able to understand what's on the00:10:41
screen specifically user interfaces to00:10:43
know where the buttons are how to click00:10:45
them and 5.2 is much improved over 5.1.00:10:49
So for the screen spot pro guey00:10:52
screenshot understanding 64% for 5.100:10:55
thinking and 86% for 5.2. And here's00:10:58
another example at just how good it is00:11:00
at visual reasoning. Here's a picture of00:11:02
a motherboard. 5.1 it was asked to00:11:04
identify all the different elements of00:11:06
it. And as you can see, not very good,00:11:08
not accurate on the actual boxing and00:11:12
only identified for elements of it00:11:14
versus 5.200:11:17
really did much better, identified many00:11:19
more ports and chips and RAM and boxed00:11:22
it much more accurately. And for tool00:11:24
use, we also saw a great improvement.00:11:27
Here's TA 2 bench for telecom use cases.00:11:29
This is customer support. nearly perfect00:11:32
98.7 for 5.2 thinking versus here's 5.100:11:36
thinking 47% double literally double and00:11:40
so what that actually means is 5.2 is00:11:43
much better at tool calling and long00:11:45
chains of tool calling. Here's an00:11:47
example. My flight from Paris to New00:11:48
York was delayed and I missed my00:11:50
connection to Austin. My check bag is00:11:53
also missing and I need to spend the00:11:54
night in New York. I also require a00:11:57
special front row seat for medical00:11:58
reasons. Can you help me? 5.1, it was00:12:01
only able to do tool calling over a00:12:03
handful of different iterations versus00:12:05
5.2, look how many more it was able to00:12:07
do. It is also much better at mental00:12:09
health evaluations, which is equally as00:12:12
important as everything else we're00:12:14
discussing. When people use these00:12:16
models, we need to know that they are00:12:17
safe while using them. So, let's talk00:12:19
about pricing. It is more expensive,00:12:22
unfortunately. I thought a version bump00:12:24
would be the same price, but it isn't.00:12:26
You do get a lot more for it, but still00:12:28
it is a pretty meaningful price00:12:30
increase. So, GPT 5.1 per million input00:12:34
tokens, $12500:12:37
versus $1.7500:12:39
for 5.2. That is a big increase. And $1000:12:42
per output for 5.1 versus $14 per00:12:46
million output for 5.2. And last,00:12:48
Almarina published some preliminary00:12:51
results. And GBT 5.2 too high is coming00:12:55
in at an ELO score of 1486, putting it00:12:58
at number two overall, coming in just00:13:01
under Opus 4.5, which is still the00:13:04
number one coding model in the world00:13:06
according to LM Marina. So, I'll drop a00:13:08
link to this blog post down below. They00:13:10
did put out a ton of different00:13:12
benchmarks. I would have liked if they00:13:15
included the other companies, Frontier00:13:17
models, and all of these different00:13:19
charts for the benchmarks, but hopefully00:13:21
they do that next time. Box put out some00:13:23
nice enterprise benchmarks. No, they're00:13:25
not sponsoring this video, but I just00:13:27
find this to be valuable. So, let me00:13:28
show it to you. What we're seeing here00:13:30
is time to first token for Box AAI00:13:32
Enterprise Eval. And so, lower is00:13:34
better. And what we're seeing here is a00:13:36
drastic decrease from 5 to 5.1 to 5.200:13:40
across the board. Long document complex00:13:43
extraction, analytical query, and00:13:45
multi-turn query. Then we're seeing00:13:47
accuracy scores for those enterprise use00:13:50
cases. And once again, 59 to 66. And as00:13:53
you can see pretty much across the00:13:55
board, a very nice improvement. Once00:13:57
again, I would have liked to see how it00:13:59
compares to other Frontier models in the00:14:01
same chart, but maybe next time. So00:14:04
that's it for today. It seems like00:14:05
pre-training is not slowing down. We00:14:07
have not hit a wall. Very excited to see00:14:09
these big improvements to the GPT500:14:12
series of models. If you enjoyed this00:14:14
video, please consider giving a like and00:14:15
subscribe.Description:
Check out the Dell Pro Max Workstation with the NVIDIA RTX PRO! https://www.dell.com/en-us/lp/dt/nvidia-ai Enter the AI Bundle Giveaway with over $15k in prizes!👇🏼 https://www.forwardfuture.ai/the-high-taste-ai-bundle Download The Subtle Art of Not Being Replaced 👇🏼 https://www.forwardfuture.ai/one-hundred-ways-to-use-ai-lp Download Humanities Last Prompt Engineering Guide 👇🏼 https://www.forwardfuture.ai/p/humanity-s-last-prompt-engineering-guide Join My Newsletter for Regular AI Updates 👇🏼 https://www.forwardfuture.ai/ Discover The Best AI Tools👇🏼 https://tools.forwardfuture.ai/ My Links 🔗 👉🏻 X: https://x.com/matthewberman 👉🏻 Forward Future X: https://x.com/forward_future_ 👉🏻 Instagram: https://www.facebook.com/unsupportedbrowser 👉🏻 Discord: https://discord.com/invite/xxysSXBxFW 👉🏻 TikTok: https://www.tiktok.com/@matthewberman_ai Media/Sponsorship Inquiries ✅ https://crswhd1x01z.typeform.com/to/BX2fuHIe Links: https://openai.com/index/introducing-gpt-5-2/ https://x.com/flavioAd/status/1999183432203567339 https://x.com/arcprize/status/1999182732845547795 https://x.com/arena/status/1999183339283185878 https://blog.box.com/how-openais-gpt-52-delivers-lightning-fast-specialist-level-reasoning
Mediafile available in formats
* — If the video is playing in a new tab, go to it, then right-click on the video and select "Save video as..."
** — Link intended for online playback in specialized players
Questions about downloading video
How can I download "OpenAI just dropped GPT-5.2... (WOAH)" video?
http://univideos.ru/ website is the best way to download a video or a separate audio track if you want to do without installing programs and extensions.
The UDL Helper extension is a convenient button that is seamlessly integrated into YouTube, Instagram and OK.ru sites for fast content download.
UDL Client program (for Windows) is the most powerful solution that supports more than 900 websites, social networks and video hosting sites, as well as any video quality that is available in the source.
UDL Lite is a really convenient way to access a website from your mobile device. With its help, you can easily download videos directly to your smartphone.
Which format of "OpenAI just dropped GPT-5.2... (WOAH)" video should I choose?
The best quality formats are FullHD (1080p), 2K (1440p), 4K (2160p) and 8K (4320p). The higher the resolution of your screen, the higher the video quality should be. However, there are other factors to consider: download speed, amount of free space, and device performance during playback.
Why does my computer freeze when loading a "OpenAI just dropped GPT-5.2... (WOAH)" video?
The browser/computer should not freeze completely! If this happens, please report it with a link to the video. Sometimes videos cannot be downloaded directly in a suitable format, so we have added the ability to convert the file to the desired format. In some cases, this process may actively use computer resources.
How can I download "OpenAI just dropped GPT-5.2... (WOAH)" video to my phone?
You can download a video to your smartphone using the website or the PWA application UDL Lite. It is also possible to send a download link via QR code using the UDL Helper extension.
How can I download an audio track (music) to MP3 "OpenAI just dropped GPT-5.2... (WOAH)"?
The most convenient way is to use the UDL Client program, which supports converting video to MP3 format. In some cases, MP3 can also be downloaded through the UDL Helper extension.
How can I save a frame from a video "OpenAI just dropped GPT-5.2... (WOAH)"?
This feature is available in the UDL Helper extension. Make sure that "Show the video snapshot button" is checked in the settings. A camera icon should appear in the lower right corner of the player to the left of the "Settings" icon. When you click on it, the current frame from the video will be saved to your computer in JPEG format.
How do I play and download streaming video?
- in video formats, hover your mouse over "Streaming Video**";
- right-click on "Copy link";
- open VLC-player;
- select Media - Open Network Stream - Network in the menu;
- paste the copied link into the input field;
- click "Play".
- copy the video address (URL);
- select "Open Network Stream" in the "Media" item of VLC player and paste the link to the video into the input field;
- click on the arrow on the "Play" button and select "Convert" in the list;
- select "Video - H.264 + MP3 (MP4)" in the "Profile" line;
- click the "Browse" button to select a folder to save the converted video and click the "Start" button;
- conversion speed depends on the resolution and duration of the video.
For this purpose you need VLC-player, which can be downloaded for free from the official website https://www.videolan.org/vlc/.
How to play streaming video through VLC player:
To download streaming video via VLC player, you need to convert it:
Warning: this download method no longer works with most YouTube videos.
What's the price of all this stuff?
It costs nothing. Our services are absolutely free for all users. There are no PRO subscriptions, no restrictions on the number or maximum length of downloaded videos.