WEBVTT 00:00:00.000 --> 00:00:04.080 Technology usually creeps up on us slowly. Beat. 00:00:04.839 --> 00:00:08.019 But sometimes it kicks the door down. Yeah, it 00:00:08.019 --> 00:00:10.300 really does. I was pondering this pace of change 00:00:10.300 --> 00:00:13.779 recently. Yeah. We woke up on March 5th, 2026, 00:00:14.439 --> 00:00:17.059 to a very different reality. A completely wild 00:00:17.059 --> 00:00:19.679 reality. Right. OpenAI dropped an update that 00:00:19.679 --> 00:00:21.839 doesn't just answer your questions. It actually 00:00:21.839 --> 00:00:23.699 takes control of your mouse to do your digital 00:00:23.699 --> 00:00:26.239 chores. Welcome to the deep dive. I have been 00:00:26.239 --> 00:00:28.800 waiting for this one. We are unpacking the massive 00:00:28.800 --> 00:00:33.179 GPT 5 .4 chat GPT update today. It is a fundamental 00:00:33.179 --> 00:00:35.439 shift. We are moving past chatbots entirely. 00:00:35.780 --> 00:00:37.479 Absolutely. That is our mission for you today. 00:00:37.719 --> 00:00:39.979 We will cover the brand new model lineup first. 00:00:40.560 --> 00:00:43.380 And we'll explore how the system acts as a professional 00:00:43.380 --> 00:00:45.539 desk worker. We'll also look at its wild new 00:00:45.539 --> 00:00:48.119 ability to see. It can actually use your computer 00:00:48.119 --> 00:00:50.649 like a human would. Plus, we are diving into 00:00:50.649 --> 00:00:52.990 its coding chops. And finally, we will break 00:00:52.990 --> 00:00:55.570 down the real -world costs and safety implications. 00:00:55.770 --> 00:00:57.909 Let's jump right in. Let's start by meeting the 00:00:57.909 --> 00:01:00.170 new models. If you logged in recently, you probably 00:01:00.170 --> 00:01:02.549 noticed the confusing name. Oh, very confusing. 00:01:02.689 --> 00:01:04.609 The version numbers jump around quite a bit. 00:01:04.890 --> 00:01:07.489 There's no 5 .3 thinking model at all. Nope. 00:01:07.769 --> 00:01:09.650 They went straight from the old baseline to 5 00:01:09.650 --> 00:01:13.140 .4. It feels like a... Well, like a skip step. 00:01:13.900 --> 00:01:16.040 But it signals a massive architectural leap. 00:01:16.219 --> 00:01:18.120 Yeah, the lineup itself is actually very focused 00:01:18.120 --> 00:01:20.599 now. You have three main versions. First up is 00:01:20.599 --> 00:01:24.420 GPT 5 .3 Instant. Right. And that model is built 00:01:24.420 --> 00:01:26.540 for pure speed. It gives you an answer immediately. 00:01:26.739 --> 00:01:29.439 You hit Enter, and the text is just there. Almost 00:01:29.439 --> 00:01:32.140 zero latency. I find it perfect for those quick, 00:01:32.159 --> 00:01:34.950 everyday questions. like asking for a recipe 00:01:34.950 --> 00:01:37.870 or doing some basic fact -checking. It doesn't 00:01:37.870 --> 00:01:40.310 spend computing power pondering deeply, it just 00:01:40.310 --> 00:01:42.849 reacts. It's your quick reflex, but then you 00:01:42.849 --> 00:01:45.890 have the star of the show. GPT 5 .4 thinking. 00:01:46.510 --> 00:01:50.049 This model actually pauses to reason before answering 00:01:50.049 --> 00:01:53.230 complex queries. Beat, it maps out a logical 00:01:53.230 --> 00:01:55.849 path. You literally see a small box indicating 00:01:55.849 --> 00:01:59.569 its thinking. That brief pause changes the quality 00:01:59.569 --> 00:02:02.400 of the answer entirely. It stops treating words 00:02:02.400 --> 00:02:04.920 like a predictive text game. Exactly. It acts 00:02:04.920 --> 00:02:07.260 more like a strategist. And for the heavy lifters, 00:02:07.500 --> 00:02:11.180 there is GPT 5 .4 Pro. You do need a Pro or Enterprise 00:02:11.180 --> 00:02:13.879 plan to access this one. Yeah, it takes the longest 00:02:13.879 --> 00:02:17.759 to process, but it delivers absolute peak accuracy 00:02:17.759 --> 00:02:20.120 for deep research. Most of you will probably 00:02:20.120 --> 00:02:22.939 rely on auto mode, though. Auto mode is the practical 00:02:22.939 --> 00:02:25.520 choice. Think of it like an automatic transmission 00:02:25.520 --> 00:02:28.379 in a car. It shifts gears for you based on the 00:02:28.379 --> 00:02:30.710 road ahead. It takes that a step further, really. 00:02:30.889 --> 00:02:33.129 It's not just shifting gears. It's deciding if 00:02:33.129 --> 00:02:35.509 you need a bicycle or a freight train. Right. 00:02:35.669 --> 00:02:38.150 You ask a simple math question. It uses instant. 00:02:38.349 --> 00:02:40.710 You ask for a 10 page mark analysis. It shifts 00:02:40.710 --> 00:02:43.229 into thinking. I am looking at these three tiers 00:02:43.229 --> 00:02:46.590 and I have to ask, does auto mode actually save 00:02:46.590 --> 00:02:48.830 you time in the long run? Yes, it takes the right 00:02:48.830 --> 00:02:50.969 brain power so you don't have to guess. That 00:02:50.969 --> 00:02:53.750 friction removal is key. Let's shift into how 00:02:53.750 --> 00:02:57.060 this impacts professional work. OpenAI focused 00:02:57.060 --> 00:02:59.340 heavily on knowledge work with this release, 00:02:59.699 --> 00:03:02.400 managing data, rating emails, building presentations. 00:03:02.759 --> 00:03:05.780 They ran a rigorous benchmark called the GDPVAL 00:03:05.780 --> 00:03:08.740 test. Right. This test measures how well the 00:03:08.740 --> 00:03:11.539 AI performs jobs that human professionals do. 00:03:11.840 --> 00:03:14.240 We're talking about complex accounting or sales 00:03:14.240 --> 00:03:17.479 roles. The results were genuinely striking. GPT 00:03:17.479 --> 00:03:22.120 5 .4 matched or beat human pros 83 % of the time. 00:03:22.620 --> 00:03:25.199 Two -sex silence. That isn't just a slight improvement. 00:03:25.280 --> 00:03:27.960 No, that is a structural shift in corporate capability. 00:03:28.280 --> 00:03:30.699 It changes the hiring landscape entirely. Think 00:03:30.699 --> 00:03:33.280 about spreadsheets. If you spend half your week 00:03:33.280 --> 00:03:35.740 formatting raw data, this changes your life. 00:03:36.039 --> 00:03:39.360 The source highlighted a specific, highly complicated 00:03:39.360 --> 00:03:42.879 example. You feed it a messy list of sales data. 00:03:43.099 --> 00:03:46.400 Data that is unformatted, missing columns, and 00:03:46.400 --> 00:03:48.680 full of weird date strings. Yeah, the worst kind 00:03:48.680 --> 00:03:51.099 of data. You ask it to find the trends. You ask 00:03:51.099 --> 00:03:53.159 for the best -selling product. Then you ask for 00:03:53.159 --> 00:03:55.520 a prediction for next month based on a 5 % growth 00:03:55.520 --> 00:03:58.000 rate. It handles the math flawlessly. It runs 00:03:58.000 --> 00:04:00.199 the regression without complaining. But here's 00:04:00.199 --> 00:04:02.280 the kicker. It formats it as an Excel -ready 00:04:02.280 --> 00:04:05.500 table. Plus, there is now a direct chat GPT for 00:04:05.500 --> 00:04:08.379 Excel plugin. That plugin is huge. Context switching 00:04:08.379 --> 00:04:11.199 kills productivity. If you leave Excel to use 00:04:11.199 --> 00:04:14.360 a chatbot, you lose focus. Now, you use the model 00:04:14.360 --> 00:04:16.720 directly inside your sheets. The days of endlessly 00:04:16.720 --> 00:04:19.339 copy -pasting back and forth are over. It transforms 00:04:19.339 --> 00:04:22.279 you from a data entry clerk into a manager. And 00:04:22.279 --> 00:04:25.350 it isn't just raw data. It creates full presentations 00:04:25.350 --> 00:04:28.170 in minutes. You ask it to research a topic like 00:04:28.170 --> 00:04:31.009 the future of solar energy. You prompt it to 00:04:31.009 --> 00:04:33.709 generate a 10 -slide PowerPoint presentation. 00:04:34.050 --> 00:04:36.509 You ask for professional tone. And you make sure 00:04:36.509 --> 00:04:39.769 it includes rigorous academic citations. It doesn't 00:04:39.769 --> 00:04:42.490 just give you the text. It builds the actual 00:04:42.490 --> 00:04:45.430 file for you to download. The designs are incredibly 00:04:45.430 --> 00:04:48.069 sharp now, too. They use clean, modern colors 00:04:48.069 --> 00:04:51.430 and solid typography. And if you hate the design... 00:04:51.500 --> 00:04:53.819 You don't have to start over. Nope. You just 00:04:53.819 --> 00:04:55.579 tell it, make it more modern and minimalist. 00:04:55.959 --> 00:04:58.319 It rebuilds the entire thing in minutes. Yeah. 00:04:58.699 --> 00:05:00.560 But dealing with neat examples is one thing. 00:05:01.040 --> 00:05:03.879 Can it handle genuinely chaotic, unformatted 00:05:03.879 --> 00:05:06.079 data? Absolutely. It cleans it up and formats 00:05:06.079 --> 00:05:08.800 it perfectly for Excel. This brings us to what 00:05:08.800 --> 00:05:11.699 might be the real breakthrough. TwoSec Silence, 00:05:12.100 --> 00:05:15.310 AI with built -in computer use. This is the mind 00:05:15.310 --> 00:05:16.829 -blowing part of the update. It doesn't just 00:05:16.829 --> 00:05:18.790 give you advice anymore. It actually acts on 00:05:18.790 --> 00:05:21.389 your behalf. Right. It writes code to run your 00:05:21.389 --> 00:05:24.290 machine. Or it visually navigates screenshots. 00:05:24.509 --> 00:05:27.350 Think about that ISWorld verified score. It hits 00:05:27.350 --> 00:05:32.129 75 .0 % on desktop use. Beating the human score 00:05:32.129 --> 00:05:36.160 of 72 .4%. and absolutely crushing the old GPC 00:05:36.160 --> 00:05:40.839 5 .2. Which sat at just 47 .3%. It can look at 00:05:40.839 --> 00:05:43.100 a messy desktop and find an invoice perfectly. 00:05:43.300 --> 00:05:45.699 When you look at web browsing. The Mine2 web 00:05:45.699 --> 00:05:48.519 test is fascinating. It finished complex tasks 00:05:48.519 --> 00:05:51.139 just by looking at browser screenshots with 92 00:05:51.139 --> 00:05:54.060 .8 % accuracy. It doesn't need a special API 00:05:54.060 --> 00:05:55.980 to talk to a website. It just looks at the screen 00:05:55.980 --> 00:05:58.180 like we do. We also have to mention agentic web 00:05:58.180 --> 00:06:00.560 search. On the browse comp test, the Pro model 00:06:00.560 --> 00:06:03.959 scored 89 .3%. It excels at those painful needle 00:06:03.959 --> 00:06:07.300 -in -a -haystack tasks. Finding that one tiny 00:06:07.300 --> 00:06:09.620 buried piece of information on a chaotic website. 00:06:09.860 --> 00:06:12.100 It works because its upgraded vision is incredible. 00:06:12.279 --> 00:06:16.079 The MMU Pro test proves that. Scoring 81 .2 % 00:06:16.079 --> 00:06:18.459 on complex photos and scientific charts. And 00:06:18.459 --> 00:06:21.860 on Omnidoc Bench, it reads PDFs with a tiny .11 00:06:21.860 --> 00:06:24.579 error rate. It can process images at an original 00:06:24.579 --> 00:06:27.779 detail level of 10 .24 million pixels. That is 00:06:27.779 --> 00:06:30.800 a 6 ,000 pixel dimension. It spots tiny 10 pixel 00:06:30.800 --> 00:06:33.480 buttons it used to miss entirely. I still wrestle 00:06:33.480 --> 00:06:36.480 with the idea of letting an AI freely click around 00:06:36.480 --> 00:06:38.800 my personal desktop. I completely understand 00:06:38.800 --> 00:06:41.420 that hesitation. Handing over the mouse feels 00:06:41.420 --> 00:06:44.459 unnatural. Very unnatural. But there are serious 00:06:44.459 --> 00:06:46.779 safety rules built into the architecture. You 00:06:46.779 --> 00:06:49.579 can guide it with specific, limiting messages. 00:06:49.699 --> 00:06:52.420 With all this capability, I have to wonder, is 00:06:52.420 --> 00:06:54.920 it safe to let it click around your private files? 00:06:55.100 --> 00:06:58.079 You set strict rules and it asks permission before 00:06:58.079 --> 00:07:01.060 taking risks. We will be right back after a quick 00:07:01.060 --> 00:07:04.800 word. Sponsor. Welcome back to the Deep Dive. 00:07:05.079 --> 00:07:08.220 Before the break, we talked about how GPT 5 .4 00:07:08.220 --> 00:07:11.089 can navigate your desktop. But what if the software 00:07:11.089 --> 00:07:14.050 you need doesn't exist yet? That is where this 00:07:14.050 --> 00:07:16.470 update fundamentally changes how we build things. 00:07:16.709 --> 00:07:18.970 It's a massive deal for developers. The main 00:07:18.970 --> 00:07:22.449 GPT 5 .4 model fully absorbed the old codex tool. 00:07:22.889 --> 00:07:24.990 It is now native. And it's just as good as that 00:07:24.990 --> 00:07:26.930 specialized model ever was. It achieves much 00:07:26.930 --> 00:07:29.100 higher accuracy. and it does it in a fraction 00:07:29.100 --> 00:07:31.060 of the time. The iteration loop for software 00:07:31.060 --> 00:07:34.379 engineering is collapsing. Jeep's T5 .2 used 00:07:34.379 --> 00:07:36.759 to take nearly 2 ,000 seconds to hit its best 00:07:36.759 --> 00:07:38.959 accuracy. Now you get better results in about 00:07:38.959 --> 00:07:42.079 half that time. It starts smarter and stays ahead 00:07:42.079 --> 00:07:45.319 as it thinks. This enables a wild concept called 00:07:45.319 --> 00:07:47.920 vibe coding. You don't need to know C++ anymore. 00:07:48.160 --> 00:07:50.300 You just describe what you want and the AI builds 00:07:50.300 --> 00:07:52.819 it from scratch. The source highlighted building 00:07:52.819 --> 00:07:56.399 a 3D highway racing game. using a single prompt. 00:07:56.519 --> 00:07:59.399 It generated a complete HTML and JavaScript file 00:07:59.399 --> 00:08:01.920 instantly. It included a car selection screen 00:08:01.920 --> 00:08:04.720 with three colors. It added moving traffic, a 00:08:04.720 --> 00:08:07.839 nitro boost, and a real damage system. It even 00:08:07.839 --> 00:08:09.939 added professional details like street lamps 00:08:09.939 --> 00:08:13.120 and trees. The code was long and incredibly complex. 00:08:13.449 --> 00:08:16.290 But the physics actually felt real. If you can 00:08:16.290 --> 00:08:19.069 do that in 30 seconds, the barrier to entry for 00:08:19.069 --> 00:08:21.290 game design just dropped to zero. And if you 00:08:21.290 --> 00:08:23.629 are using the API, there is a new slash fast 00:08:23.629 --> 00:08:27.029 mode. It makes code generation 1 .5 times faster 00:08:27.029 --> 00:08:29.430 with no quality loss. It saves you from waiting 00:08:29.430 --> 00:08:32.029 for hundreds of lines of boilerplate code. But 00:08:32.029 --> 00:08:34.570 the biggest shift is the context window. It jumped 00:08:34.570 --> 00:08:38.529 from the standard 272k tokens to an experimental 00:08:38.529 --> 00:08:42.090 1 million tokens. The context window is the AI's 00:08:42.090 --> 00:08:44.070 short -term memory for tracking your current 00:08:44.070 --> 00:08:46.490 conversation and files. Whoa, imagine feeding 00:08:46.490 --> 00:08:48.809 it a whole library of code and it remembers line 00:08:48.809 --> 00:08:51.690 one. It changes everything about debugging. It 00:08:51.690 --> 00:08:54.090 can hold your entire project architecture in 00:08:54.090 --> 00:08:56.470 its head. It sees all your files at the same 00:08:56.470 --> 00:08:58.850 time to find hidden mistakes. But with all that 00:08:58.850 --> 00:09:01.850 processing power, Does using that massive 1 million 00:09:01.850 --> 00:09:05.289 context window cost more? Yes. Requests over 00:09:05.289 --> 00:09:07.929 the standard limit eat usage at double speed. 00:09:08.529 --> 00:09:11.159 We had to look at the broader rivalry here. How 00:09:11.159 --> 00:09:13.379 does this compare to competitors like Clawed 00:09:13.379 --> 00:09:17.080 4 .6 and Gemini 3 .1? It's a very close race 00:09:17.080 --> 00:09:20.539 right now. GPT 5 .4 clearly wins on knowledge 00:09:20.539 --> 00:09:23.220 work, overall speed, and computer use. That built 00:09:23.220 --> 00:09:25.899 -in desktop interaction is incredibly advanced 00:09:25.899 --> 00:09:28.299 compared to the rest. But to be fair to the source, 00:09:28.539 --> 00:09:31.159 Clawed still sounds more natural. Yeah, GPT 5 00:09:31.159 --> 00:09:33.600 .4 can sometimes feel a bit robotic when writing 00:09:33.600 --> 00:09:36.379 blogs or creative essays. It lacks a certain 00:09:36.379 --> 00:09:39.559 warmth. That is a fair critique. And Gemini remains 00:09:39.559 --> 00:09:41.620 highly creative for marketing and storytelling. 00:09:42.179 --> 00:09:44.159 The constant back and forth between these companies 00:09:44.159 --> 00:09:47.179 is great for users, though. It forces rapid innovation. 00:09:47.419 --> 00:09:49.379 Let's talk about pricing. Because this power 00:09:49.379 --> 00:09:52.080 isn't free. No, it is not. For plus users paying 00:09:52.080 --> 00:09:56.559 $20 a month, you get 5 .3 instant and 5 .4 thinking. 00:09:56.879 --> 00:09:59.700 Pro requires a much higher enterprise tier. There 00:09:59.700 --> 00:10:03.120 were some launch day message limit bugs frustrating 00:10:03.120 --> 00:10:05.679 users. But those usually clear up quickly as 00:10:05.679 --> 00:10:07.779 they scale their servers. For API developers 00:10:07.779 --> 00:10:10.299 building apps, the tokens do cost a bit more 00:10:10.299 --> 00:10:15.419 upfront. It's $2 .50 in and $15 out per 1 million 00:10:15.419 --> 00:10:18.899 tokens. But you actually save up to 47 % overall. 00:10:19.259 --> 00:10:21.779 The new tool search feature is a brilliant cost 00:10:21.779 --> 00:10:24.519 saving measure. In the past, the AI had to read 00:10:24.519 --> 00:10:27.200 every single tool definition for every single 00:10:27.200 --> 00:10:29.299 request. It was like reading the entire dictionary 00:10:29.299 --> 00:10:31.720 just to spell one word. Now it only looks up 00:10:31.720 --> 00:10:34.940 the specific tool it needs. In tests, that dropped 00:10:34.940 --> 00:10:39.159 usage from 123k tokens down to just 65k tokens. 00:10:39.720 --> 00:10:42.259 So are overall API costs actually going down 00:10:42.259 --> 00:10:44.379 for developers? Smarter tool search means fewer 00:10:44.379 --> 00:10:46.820 tokens, offsetting the higher base price. On 00:10:46.820 --> 00:10:49.759 the safety front, the AI relies heavily on reinforcement 00:10:49.759 --> 00:10:51.940 learning. Which is crucial. Reinforcement learning 00:10:51.940 --> 00:10:53.840 is checking its own work and trying different 00:10:53.840 --> 00:10:57.139 paths to find answers. Exactly. It grades its 00:10:57.139 --> 00:10:59.139 own homework before showing you the final result. 00:10:59.340 --> 00:11:01.220 And its internal thoughts are fully transparent 00:11:01.220 --> 00:11:04.320 to OpenAI. It can't hide its reasoning from human 00:11:04.320 --> 00:11:07.580 oversight. That prevents it from executing malicious 00:11:07.580 --> 00:11:10.419 logic. Let's leave you with a few pro tips for 00:11:10.419 --> 00:11:14.080 your daily workflow. The first is mid -response 00:11:14.080 --> 00:11:17.000 redirection. This workflow is absolute magic. 00:11:17.120 --> 00:11:18.740 You don't have to hit the stop button anymore 00:11:18.740 --> 00:11:20.960 if it goes off track. It's like steering a horse 00:11:20.960 --> 00:11:22.840 while you're already galloping. If it's writing 00:11:22.840 --> 00:11:25.399 a long report and you realize you forgot an instruction, 00:11:25.600 --> 00:11:28.480 you just type it. You say, focus on environmental 00:11:28.480 --> 00:11:31.830 impact right while it is generating text. And 00:11:31.830 --> 00:11:34.610 it pivots instantly. It just adapts the text 00:11:34.610 --> 00:11:37.049 on the fly without missing a beat. You don't 00:11:37.049 --> 00:11:39.470 lose the good parts it already wrote. Do I have 00:11:39.470 --> 00:11:41.429 to wait for it to finish typing completely? Nope. 00:11:41.509 --> 00:11:43.450 Just interrupt mid -sentence and it pivots its 00:11:43.450 --> 00:11:45.809 thoughts instantly. You can also adjust its thinking 00:11:45.809 --> 00:11:48.570 effort in the settings menu. You have standard 00:11:48.570 --> 00:11:51.090 and heavy options. Leave it on standard for 90 00:11:51.090 --> 00:11:54.269 % of your daily tasks. It's fast, incredibly 00:11:54.269 --> 00:11:57.059 smart, and handles routine logic perfectly. But 00:11:57.059 --> 00:12:00.240 crank it up to heavy for deep math or nasty coding 00:12:00.240 --> 00:12:02.620 bugs. It might take five to eight minutes to 00:12:02.620 --> 00:12:05.259 answer. It's beat. Think about what an eight 00:12:05.259 --> 00:12:08.700 minute AI thought process looks like. It yields 00:12:08.700 --> 00:12:11.399 master level results. It really does. Let's take 00:12:11.399 --> 00:12:13.779 a step back and look at the big idea here. Two 00:12:13.779 --> 00:12:17.039 sec silence. This update fundamentally shifts 00:12:17.039 --> 00:12:20.500 what this tool actually is. It's no longer just 00:12:20.500 --> 00:12:22.740 a super powered search engine. It has become 00:12:22.740 --> 00:12:25.730 a true desktop partner. And the ultimate metric 00:12:25.730 --> 00:12:29.129 we are talking about is time. You are compressing 00:12:29.129 --> 00:12:31.669 an hour of tedious spreadsheet wrangling into 00:12:31.669 --> 00:12:33.610 minutes. You're turning a day of hunting down 00:12:33.610 --> 00:12:35.789 code bugs into 10 minutes of simple oversight. 00:12:36.210 --> 00:12:39.029 Exactly. If this AI can now flawlessly look at 00:12:39.029 --> 00:12:41.549 your screen, interpret the buttons, and navigate 00:12:41.549 --> 00:12:44.049 your desktop better than most humans. How long 00:12:44.049 --> 00:12:46.850 until we stop needing traditional operating systems 00:12:46.850 --> 00:12:48.950 and screens altogether? What happens when the 00:12:48.950 --> 00:12:52.279 AI becomes the entire interface? Pause for effect. 00:12:52.580 --> 00:12:54.960 That is wild to think about. It changes the whole 00:12:54.960 --> 00:12:56.899 paradigm of human -computer interaction. For 00:12:56.899 --> 00:12:59.840 now, try this out yourself. Pick one tedious 00:12:59.840 --> 00:13:02.620 task you do every single week. Just see if GPT 00:13:02.620 --> 00:13:05.019 5 .4 thinking can do the first draft for you. 00:13:05.299 --> 00:13:07.379 You might be surprised by how much your daily 00:13:07.379 --> 00:13:10.340 workflow changes. Thanks for exploring this deep 00:13:10.340 --> 00:13:12.460 dive with us. Catch you next time.