WEBVTT 00:00:00.000 --> 00:00:02.259 Imagine an AI that doesn't just, you know, assist 00:00:02.259 --> 00:00:04.320 you, but actually builds things. You give it 00:00:04.320 --> 00:00:07.360 a single prompt, and poof! Maybe a video game 00:00:07.360 --> 00:00:10.000 appears. Well, that's not quite science fiction 00:00:10.000 --> 00:00:12.970 anymore. Welcome to the Deep Dive. We're here 00:00:12.970 --> 00:00:15.250 to untack the technologies really reshaping our 00:00:15.250 --> 00:00:19.089 world. Today, we are going deep on OpenAI's GPT 00:00:19.089 --> 00:00:21.870 -5, our mission, to figure out what makes this 00:00:21.870 --> 00:00:24.329 feel less like just an update and more like a, 00:00:24.329 --> 00:00:26.609 well, a genuine paradigm shift for automation 00:00:26.609 --> 00:00:28.710 and maybe beyond. That's right. We're going to 00:00:28.710 --> 00:00:31.969 explore what GPT -5 is. You can think of it almost 00:00:31.969 --> 00:00:35.030 like a family of specialized AI brains. We'll 00:00:35.030 --> 00:00:37.549 break down some, frankly, jaw -dropping benchmark 00:00:37.549 --> 00:00:40.329 results. And uncover these surprising one -shot 00:00:40.329 --> 00:00:43.170 creation abilities it seems to have. And then 00:00:43.170 --> 00:00:44.869 we'll pivot. We'll look at the practical side, 00:00:44.990 --> 00:00:47.090 what this means for automation, for real -world 00:00:47.090 --> 00:00:49.170 applications. We want to offer you a kind of 00:00:49.170 --> 00:00:51.570 strategic playbook for using this power efficiently. 00:00:52.009 --> 00:00:54.130 And, like always, we'll wrap up by looking at 00:00:54.130 --> 00:00:56.350 the future this tech is carving out. Our approach 00:00:56.350 --> 00:01:00.310 today, calm, curious, focused on really understanding 00:01:00.310 --> 00:01:02.950 this new frontier. Let's get into it. Okay, let's 00:01:02.950 --> 00:01:04.930 unpack this then. The source material we looked 00:01:04.930 --> 00:01:08.739 at calls GPT -5 a seismic event. When we hear 00:01:08.739 --> 00:01:11.209 seismic event, I mean, what really makes this 00:01:11.209 --> 00:01:14.530 more than just, say, GPT -4 .5? What's fundamentally 00:01:14.530 --> 00:01:17.290 different here? Well, Sam Altman, OpenAI's CEO, 00:01:17.590 --> 00:01:19.730 he put it pretty well, I think. He basically 00:01:19.730 --> 00:01:23.170 described the leap from 4 to 5 as the difference 00:01:23.170 --> 00:01:25.450 between talking to a really smart college student, 00:01:25.530 --> 00:01:27.489 that was GPT -4, and talking to a PhD -level 00:01:27.489 --> 00:01:30.430 expert, that's GPT -5. He said GPT -3 was maybe 00:01:30.430 --> 00:01:32.890 like a high schooler. GPT -4, the college student. 00:01:32.969 --> 00:01:35.730 But GPT -5, it acts like a true domain expert, 00:01:35.829 --> 00:01:37.730 pretty much across any field you throw at it. 00:01:37.769 --> 00:01:40.349 A PhD -level expert. Okay, that's quite a jump. 00:01:40.469 --> 00:01:42.829 It really is. And for people actually building 00:01:42.829 --> 00:01:45.269 things, you know, building automation, it boils 00:01:45.269 --> 00:01:47.769 down to three critical improvements. First, just 00:01:47.769 --> 00:01:50.469 better reasoning and problem solving. GPT -5 00:01:50.469 --> 00:01:52.609 can work through these really complex multi -step 00:01:52.609 --> 00:01:55.930 problems with an accuracy that we just haven't 00:01:55.930 --> 00:01:58.569 seen before. Second, much better tool integration, 00:01:58.890 --> 00:02:01.409 its ability to understand and use external tools, 00:02:01.549 --> 00:02:04.349 APIs, which is vital for any real -world automation 00:02:04.349 --> 00:02:07.349 that's dramatically improved. And third, this 00:02:07.349 --> 00:02:09.490 is a big one, the cost efficiency is kind of 00:02:09.490 --> 00:02:12.610 mind -bending. The input tokens for the main 00:02:12.610 --> 00:02:16.189 GPT -5 model, they cost half of what GPT -4s 00:02:16.189 --> 00:02:19.629 did. We're talking 1 .25 per million input tokens 00:02:19.629 --> 00:02:22.669 for GPT -5 standard. That suddenly makes enterprise 00:02:22.669 --> 00:02:25.669 -grade AI really accessible. So it isn't just 00:02:25.669 --> 00:02:27.849 about raw power then, but more how it thinks 00:02:27.849 --> 00:02:29.909 and how it connects things. Exactly. It's about 00:02:29.909 --> 00:02:32.129 deeper understanding and real practical application. 00:02:32.990 --> 00:02:34.990 Hey, here's where it gets, I think, really interesting. 00:02:35.189 --> 00:02:37.789 The source mentions GPT -5 isn't just one single 00:02:37.789 --> 00:02:40.069 model. It's a family of brains. What does that 00:02:40.069 --> 00:02:42.550 actually mean for us? You know, the users, the 00:02:42.550 --> 00:02:44.509 builders. How should we think about that? It 00:02:44.509 --> 00:02:47.189 means you get choices and maybe more importantly, 00:02:47.330 --> 00:02:50.449 optimization. GPT -5 isn't this one monolithic 00:02:50.449 --> 00:02:53.050 thing. It's actually a suite of specialized brains. 00:02:53.449 --> 00:02:55.569 Each one is tuned for different kinds of tasks, 00:02:55.689 --> 00:02:58.289 different cost profiles. So you get GPT -5 standard. 00:02:58.490 --> 00:03:00.990 That's your well -rounded genius. Good for most 00:03:00.990 --> 00:03:03.849 general stuff. Then there's GPT -5 pro. That's 00:03:03.849 --> 00:03:06.430 the real deep thinker for those super complex 00:03:06.430 --> 00:03:09.490 multi -step problems that need maximum reasoning 00:03:09.490 --> 00:03:11.650 power. Right. And what about for things that 00:03:11.650 --> 00:03:14.129 are maybe simpler, but you do a lot of. High 00:03:14.129 --> 00:03:16.909 volume tasks. Exactly. For that, you get GPT 00:03:16.909 --> 00:03:19.409 -5 Mini. It's described as a hyper -efficient 00:03:19.409 --> 00:03:22.009 workhorse. Great for high volume, maybe simpler 00:03:22.009 --> 00:03:24.949 tasks where cost is a really big factor. And 00:03:24.949 --> 00:03:26.969 finally, there's GPT -5 Nano. That's the most 00:03:26.969 --> 00:03:29.530 economical one. Perfect for really basic high 00:03:29.530 --> 00:03:32.090 frequency needs. And this new pricing structure, 00:03:32.270 --> 00:03:34.189 alongside the family, it's kind of revolutionary. 00:03:34.349 --> 00:03:37.740 It democratizes this advanced AI. Small businesses, 00:03:37.819 --> 00:03:39.539 even solo entrepreneurs who might have found 00:03:39.539 --> 00:03:42.620 GPT -4 a bit too pricey for heavy use. Now it's 00:03:42.620 --> 00:03:45.039 potentially within reach. It feels like an economic 00:03:45.039 --> 00:03:47.849 shift just as much as a tech shift. So you can 00:03:47.849 --> 00:03:49.750 really pick the right brain for the specific 00:03:49.750 --> 00:03:52.710 job you have in mind. Yes, tailoring the power 00:03:52.710 --> 00:03:56.009 and the cost to your specific needs. Okay, let's 00:03:56.009 --> 00:03:58.069 talk about the report card. The benchmarks for 00:03:58.069 --> 00:04:00.909 GPT -5 show some pretty staggering performance 00:04:00.909 --> 00:04:03.009 improvements. Let's dive into those numbers. 00:04:03.169 --> 00:04:05.229 What do they tell us about what's actually possible 00:04:05.229 --> 00:04:07.669 now? Yeah, the numbers are, they're pretty clear. 00:04:07.770 --> 00:04:10.370 We've kind of stepped into a new era here. Take 00:04:10.370 --> 00:04:13.900 SWE Bench. That uses real -world coding problems 00:04:13.900 --> 00:04:17.939 pulled from GitHub. GPT -5 hit a 75 % accuracy 00:04:17.939 --> 00:04:21.360 rate there. That's correctly solving nearly three 00:04:21.360 --> 00:04:23.639 out of every four complex software engineering 00:04:23.639 --> 00:04:25.699 challenges it was given. That's just a massive 00:04:25.699 --> 00:04:29.000 leap for code generation. Wow, 75 % on real GitHub 00:04:29.000 --> 00:04:31.339 issues. That's practically like having a pretty 00:04:31.339 --> 00:04:33.980 dependable junior developer on call. It's certainly 00:04:33.980 --> 00:04:36.160 getting there. And then there's its multi -language 00:04:36.160 --> 00:04:39.019 ability. On a benchmark called Eater Polyglot, 00:04:39.139 --> 00:04:41.990 this tests how well an AI understands. and edits 00:04:41.990 --> 00:04:44.870 code across different languages, like Python, 00:04:45.110 --> 00:04:47.209 JavaScript, SQL, you know, like a full stack 00:04:47.209 --> 00:04:50.709 dev, GPT -5 scored 88%. That means it's getting 00:04:50.709 --> 00:04:52.730 close to being a really dependable coworker for 00:04:52.730 --> 00:04:55.129 development teams. Getting it right almost nine 00:04:55.129 --> 00:04:57.250 times out of 10 on these tricky multi -language 00:04:57.250 --> 00:04:59.290 tasks, that's huge. And then there's this thing 00:04:59.290 --> 00:05:02.189 the source calls the miracle maker, single prompt 00:05:02.189 --> 00:05:04.509 creations. That sounds, well, it sounds a bit 00:05:04.509 --> 00:05:06.689 like magic. Yeah, this is where it gets genuinely 00:05:06.689 --> 00:05:10.069 kind of mind bending, which is one single well 00:05:10.069 --> 00:05:12.490 thought out. prompt, GPT -5 has apparently been 00:05:12.490 --> 00:05:15.689 shown to create complete, professionally designed 00:05:15.689 --> 00:05:19.230 landing pages. Or interactive, fully working 00:05:19.230 --> 00:05:22.250 audio step sequencers. Little music tools. And 00:05:22.250 --> 00:05:24.889 yeah, even fully playable, pretty complex spaceship 00:05:24.889 --> 00:05:27.750 video games. And the key thing is, these weren't 00:05:27.750 --> 00:05:30.089 tweaked over and over. They were supposedly generated 00:05:30.089 --> 00:05:33.160 in a single shot. Whoa. I mean, imagine scaling 00:05:33.160 --> 00:05:35.279 that kind of one -shot creation capability up 00:05:35.279 --> 00:05:37.779 to, say, a billion users. It just fundamentally 00:05:37.779 --> 00:05:39.939 changes things like rapid prototyping. You could 00:05:39.939 --> 00:05:42.259 slash development cycles, accelerate getting 00:05:42.259 --> 00:05:44.720 minimum viable products out there. Okay, stepping 00:05:44.720 --> 00:05:47.240 back from the benchmarks for a sec, how do these 00:05:47.240 --> 00:05:50.560 breakthroughs really translate into a builder's 00:05:50.560 --> 00:05:52.920 dream, especially for automation? What specific 00:05:52.920 --> 00:05:55.790 headaches does GPT -5 maybe solve? Right, for 00:05:55.790 --> 00:05:57.370 automation builders, there are three things that 00:05:57.370 --> 00:05:59.470 really stand out. First, that tool usage accuracy 00:05:59.470 --> 00:06:02.910 we mentioned. GPT -5 is significantly better 00:06:02.910 --> 00:06:05.250 at understanding and correctly using external 00:06:05.250 --> 00:06:09.310 tools and APIs. That's just crucial for complex 00:06:09.310 --> 00:06:11.269 workflows that interact with the real world. 00:06:11.470 --> 00:06:14.870 Second, long context handling. It can keep track 00:06:14.870 --> 00:06:17.290 of the conversation, the instructions, over much 00:06:17.290 --> 00:06:20.069 longer multi -step tasks. Basically, it doesn't 00:06:20.069 --> 00:06:22.550 forget what it was doing halfway through. That 00:06:22.550 --> 00:06:24.670 makes it viable for much more sophisticated business. 00:06:24.839 --> 00:06:28.920 And third, factual accuracy. There seems to be 00:06:28.920 --> 00:06:31.819 a marked improvement, a much lower rate of hallucination. 00:06:32.160 --> 00:06:34.860 You know, when the AI just makes stuff up confidently. 00:06:34.980 --> 00:06:37.339 That obviously increases reliability quite a 00:06:37.339 --> 00:06:39.319 bit. So it's not just smarter in theory, but 00:06:39.319 --> 00:06:41.620 it's also more reliable in practice. And it actually 00:06:41.620 --> 00:06:43.620 remembers what you told it earlier. Precisely. 00:06:43.620 --> 00:06:45.699 It feels like a much more robust and trustworthy 00:06:45.699 --> 00:06:48.100 AI partner. This really sounds like a perfect 00:06:48.100 --> 00:06:50.620 match for automation platforms. The source actually 00:06:50.620 --> 00:06:53.360 calls the combination of GPT -5 and NANN, that's 00:06:53.360 --> 00:06:55.920 a popular automation tool, a perfect marriage. 00:06:56.990 --> 00:06:59.910 Why is that synergy supposedly so powerful? What 00:06:59.910 --> 00:07:02.089 makes them complement each other so well? It 00:07:02.089 --> 00:07:03.490 really does feel like that. Think of it like 00:07:03.490 --> 00:07:05.850 pairing, I don't know, the world's greatest architect 00:07:05.850 --> 00:07:09.029 that's GPT -5's intelligence with the most efficient, 00:07:09.129 --> 00:07:12.029 flexible construction crew that's N8N's platform. 00:07:13.259 --> 00:07:16.459 GPT -5's huge jump in accuracy means you can 00:07:16.459 --> 00:07:19.220 now automate more complex, even mission -critical 00:07:19.220 --> 00:07:22.480 tasks with way more confidence. Fewer errors 00:07:22.480 --> 00:07:25.060 mean more reliable business processes and probably 00:07:25.060 --> 00:07:27.060 fewer late -night calls because something broke. 00:07:27.300 --> 00:07:29.620 So it smooths out the rough edges, makes automation 00:07:29.620 --> 00:07:32.980 less brittle. Yes, exactly. And beyond just accuracy, 00:07:33.279 --> 00:07:36.060 GPT -5 is more of an all -in -one AI now. It 00:07:36.060 --> 00:07:39.079 can handle text, images, code, complex reasoning, 00:07:39.120 --> 00:07:42.199 all within a single model potentially. That simplifies 00:07:42.199 --> 00:07:44.879 workflows a lot. Before, you might need to chain 00:07:44.879 --> 00:07:47.120 together multiple different specialized AI services. 00:07:47.379 --> 00:07:49.560 Now, GPT -5 might handle the whole chain right 00:07:49.560 --> 00:07:52.180 inside your RN8N workflow. And maybe most importantly, 00:07:52.279 --> 00:07:55.300 it democratizes this power. RN8N is visual. It's 00:07:55.300 --> 00:07:57.639 low code or no code. Combine that ease of use 00:07:57.639 --> 00:07:59.980 with GPT -5's incredible ability to follow complex 00:07:59.980 --> 00:08:02.339 instructions. Well, suddenly, really advanced 00:08:02.339 --> 00:08:04.259 automation becomes accessible to non -technical 00:08:04.259 --> 00:08:06.560 users, maybe for the first time ever. So this 00:08:06.560 --> 00:08:09.339 combination means more people can build much 00:08:09.339 --> 00:08:12.000 more powerful things without needing to be expert 00:08:12.000 --> 00:08:15.160 coders. Exactly. It's genuinely empowering for 00:08:15.160 --> 00:08:17.339 builders of all kinds. Okay, for anyone listening 00:08:17.339 --> 00:08:19.600 who's now itching to try this out. The guide 00:08:19.600 --> 00:08:22.639 we read provides a step -by -step for hooking 00:08:22.639 --> 00:08:25.879 up GPT -5 with NETAN. Can you give us the quick 00:08:25.879 --> 00:08:28.000 rundown? How do people get started? Yeah, it's 00:08:28.000 --> 00:08:29.579 pretty straightforward, actually. Yeah. First, 00:08:29.819 --> 00:08:32.139 you need the key to the engine room, right? That's 00:08:32.139 --> 00:08:35.620 your OpenAI API setup. Just go to platform .openai 00:08:35.620 --> 00:08:37.480 .com, set up an account if you don't have one, 00:08:37.559 --> 00:08:40.120 add a payment method because API usage isn't 00:08:40.120 --> 00:08:43.279 free, and then create a new secret API key. Now, 00:08:43.340 --> 00:08:46.080 critical step here. Copy that key immediately 00:08:46.080 --> 00:08:48.440 and save it somewhere super secure, like a password 00:08:48.440 --> 00:08:50.940 manager. Treat it like a password, maybe even 00:08:50.940 --> 00:08:52.720 more carefully, because you won't be able to 00:08:52.720 --> 00:08:54.740 see the full key again after you create it. Got 00:08:54.740 --> 00:08:57.679 it. Guard that key. Then what's next inside NAN? 00:08:57.919 --> 00:09:01.500 Right. Then inside your NAN workflow, you just 00:09:01.500 --> 00:09:04.190 add an AI agent node. In its settings, you select 00:09:04.190 --> 00:09:07.169 OpenAI Chat Model, paste in that secret API key 00:09:07.169 --> 00:09:09.669 you just saved, and then, from the model drop 00:09:09.669 --> 00:09:12.450 -down menu, you should now see options for GPT 00:09:12.450 --> 00:09:15.330 -5 and its variants, like Mini or Pro. You just 00:09:15.330 --> 00:09:18.110 pick the one you want to use. Now, a really crucial 00:09:18.110 --> 00:09:19.950 note on billing, because this confuses a lot 00:09:19.950 --> 00:09:23.200 of people starting out. Your... monthly ChatGPT 00:09:23.200 --> 00:09:25.559 Plus subscription, if you have one, and the OpenAI 00:09:25.559 --> 00:09:28.220 API. They're two totally separate products with 00:09:28.220 --> 00:09:30.940 separate billing. ChatGPT Plus is kind of like 00:09:30.940 --> 00:09:32.720 an all -you -can -eat buffet for the chat interface. 00:09:32.960 --> 00:09:35.779 The API is a la carte. You pay specifically for 00:09:35.779 --> 00:09:38.179 what your automation workflow uses based on tokens 00:09:38.179 --> 00:09:41.639 processed. Ah, okay. So API usage is pay -as 00:09:41.639 --> 00:09:43.120 -you -go, separate from the chatbot subscription. 00:09:43.360 --> 00:09:45.500 That's important. So it's powerful, but you also 00:09:45.500 --> 00:09:47.480 need to be smart about managing those costs. 00:09:47.639 --> 00:09:50.159 Yes. Managing those API costs is absolutely key 00:09:50.159 --> 00:09:56.240 to using AI. All right. Academic benchmarks are 00:09:56.240 --> 00:09:58.779 one thing, but how does GPT -5 actually perform 00:09:58.779 --> 00:10:03.100 in, let's say, a real -world cage match? against 00:10:03.100 --> 00:10:06.279 the previous champ, GPT -4 .0. What happens when 00:10:06.279 --> 00:10:08.899 the rubber really meets the road in an actual 00:10:08.899 --> 00:10:10.980 automation workflow? Yeah, that's exactly what 00:10:10.980 --> 00:10:12.840 we wanted to find out too. So the test wasn't 00:10:12.840 --> 00:10:15.080 abstract. It was designed as a practical gauntlet. 00:10:15.200 --> 00:10:18.059 We ran both GPT -4 .0 and GPT -5 through 10, 00:10:18.259 --> 00:10:21.159 identical demanding test scenarios right inside 00:10:21.159 --> 00:10:24.240 NE10. Things like generating complex emails based 00:10:24.240 --> 00:10:26.860 on data from external tools or doing tricky database 00:10:26.860 --> 00:10:29.139 lookups and summarizing the results. And importantly, 00:10:29.139 --> 00:10:31.899 we used no custom system prompts, just raw out 00:10:31.899 --> 00:10:33.580 of the box. the box intelligence from both models. 00:10:33.700 --> 00:10:35.539 Okay. So head to head, what was the tail of the 00:10:35.539 --> 00:10:37.539 tape? How did they stack up? Right. Round one, 00:10:37.659 --> 00:10:40.919 accuracy. Here, GPT -5 landed a really heavy 00:10:40.919 --> 00:10:44.559 blow. GPT -4 -0 scored a respectable 4 .2 out 00:10:44.559 --> 00:10:47.860 of 5 .0 across the tests. But GPT -5, it hit 00:10:47.860 --> 00:10:50.559 4 .7 out of 5 .0. It's a really significant jump. 00:10:50.639 --> 00:10:52.919 It means far more reliable, much less error -prone 00:10:52.919 --> 00:10:54.720 automation right out of the gate. Round two, 00:10:54.840 --> 00:10:57.950 speed. Here, GPT -4 was consistently faster. 00:10:58.250 --> 00:11:00.750 Now, this is likely temporary, maybe due to server 00:11:00.750 --> 00:11:02.769 load from the GPT -5 launch, but for the moment, 00:11:02.870 --> 00:11:05.149 the veteran GPT -4 definitely had quicker responses. 00:11:05.470 --> 00:11:07.529 Interesting. And cost. That's always a huge factor 00:11:07.529 --> 00:11:09.590 for anyone building automations at scale. Right. 00:11:09.710 --> 00:11:12.509 Round three, cost. This was actually a bit of 00:11:12.509 --> 00:11:15.559 a surprise. While GPT -5's input tokens are indeed 00:11:15.559 --> 00:11:18.960 50 % cheaper, as OpenAI announced, it turns out 00:11:18.960 --> 00:11:21.440 GPT -5 is often so much more thorough and detailed 00:11:21.440 --> 00:11:24.059 in its responses that it frequently generates 00:11:24.059 --> 00:11:26.480 a significantly higher volume of output tokens 00:11:26.480 --> 00:11:30.360 compared to GPT -40 for the same task. So paradoxically, 00:11:30.440 --> 00:11:32.480 this can sometimes lead to a slightly higher 00:11:32.480 --> 00:11:35.159 overall cost per task with GPT -5, even though 00:11:35.159 --> 00:11:37.220 the input is cheaper. It's kind of a you -get 00:11:37.220 --> 00:11:39.100 -what -you -pay -for situation in terms of detail. 00:11:39.340 --> 00:11:41.299 Honestly, I still wrestle with prompt drift myself 00:11:41.299 --> 00:11:43.500 sometimes, you know. Finding that sweet spot 00:11:43.500 --> 00:11:45.259 between getting a really thorough answer and 00:11:45.259 --> 00:11:47.659 keeping the token count reasonable. Ah, that 00:11:47.659 --> 00:11:49.840 makes sense. More detail costs more tokens. Okay, 00:11:49.919 --> 00:11:52.679 so accuracy to GPT -5, speed to 4 .0, cost is 00:11:52.679 --> 00:11:56.200 nuanced. Was there a knockout punch? What really 00:11:56.200 --> 00:11:59.740 sealed the deal, if anything? Round four. Quality. 00:12:00.200 --> 00:12:03.220 Yeah, this is where GPT -5 delivered the decisive 00:12:03.220 --> 00:12:05.379 blow, I'd say. The sheer quality of the output 00:12:05.379 --> 00:12:08.309 was just in a completely different league. The 00:12:08.309 --> 00:12:10.490 responses felt dramatically more detailed, more 00:12:10.490 --> 00:12:13.450 personal, more nuanced, just more human -like. 00:12:13.789 --> 00:12:16.110 It didn't feel like just an incremental improvement 00:12:16.110 --> 00:12:18.629 over GPT -4 .0. It really felt like a generational 00:12:18.629 --> 00:12:20.850 leap in the quality of the interaction. And what 00:12:20.850 --> 00:12:23.289 about the little guy, the mini version? Did the 00:12:23.289 --> 00:12:26.110 featherweight contender make an impact in these 00:12:26.110 --> 00:12:28.470 tests? Yeah, the featherweight. GPT -5 mini was 00:12:28.470 --> 00:12:30.669 also put through the same 10 tests. It scored 00:12:30.669 --> 00:12:33.909 3 .6 out of 5 .0 for accuracy, respectable, but 00:12:33.909 --> 00:12:36.889 clearly below the bigger models. But here's the 00:12:36.889 --> 00:12:39.769 kicker. Running all 10 evaluations cost approximately 00:12:39.769 --> 00:12:42.889 $0 .03 in total API charges. Just $0 .03. So 00:12:42.889 --> 00:12:45.330 the verdict there is pretty clear. For high volume, 00:12:45.570 --> 00:12:47.909 relatively simple tasks where cost is absolutely 00:12:47.909 --> 00:12:50.809 paramount, GPT -5 Mini looks like the undisputed 00:12:50.809 --> 00:12:53.009 king of efficiency. So yeah, the overall judge's 00:12:53.009 --> 00:12:55.590 decision from the cage match. GPT -5 is the clear 00:12:55.590 --> 00:12:57.210 champion for accuracy and especially quality, 00:12:57.450 --> 00:12:59.750 but there's currently a tradeoff in speed and 00:12:59.750 --> 00:13:01.850 potentially slightly higher overall cost because 00:13:01.850 --> 00:13:04.070 its output is so much richer. So it sounds like 00:13:04.070 --> 00:13:06.649 for business -critical tasks, that boost in quality 00:13:06.649 --> 00:13:09.429 and reliability probably outweighs the slightly 00:13:09.429 --> 00:13:12.129 slower speed or the nuanced cost difference. 00:13:12.450 --> 00:13:15.389 Absolutely. In the real world, quality and reliability 00:13:15.389 --> 00:13:17.610 are usually what drive the most value. Okay, 00:13:17.669 --> 00:13:20.490 moving beyond those structured benchmarks, how 00:13:20.490 --> 00:13:23.710 does GPT -5 handle more complex reasoning or 00:13:23.710 --> 00:13:26.110 even creative challenges? The source material 00:13:26.110 --> 00:13:28.429 talks about testing it as an ultimate assistant 00:13:28.429 --> 00:13:32.200 or even an AI art director. This starts to sound 00:13:32.200 --> 00:13:34.720 like the JRVIS test, right? Like from Iron Man. 00:13:34.860 --> 00:13:36.639 It absolutely does feel like that. We simulated 00:13:36.639 --> 00:13:39.080 this ultimate assistance scenario. The mission 00:13:39.080 --> 00:13:41.799 was intentionally high level. Achieve a business 00:13:41.799 --> 00:13:43.620 goal that required coordinating multiple steps, 00:13:43.779 --> 00:13:45.500 web research, looking things up in a database, 00:13:45.700 --> 00:13:47.960 checking a calendar, drafting an email basically, 00:13:48.120 --> 00:13:49.919 orchestrating several different tools seamlessly. 00:13:50.200 --> 00:13:52.820 And what we saw was really a master class in 00:13:52.820 --> 00:13:55.659 AI orchestration from GPT -5. What was the most 00:13:55.659 --> 00:13:57.480 impressive part of that orchestration? What really 00:13:57.480 --> 00:13:59.659 stood out? The most striking thing was what the 00:13:59.659 --> 00:14:03.320 source called a self -healing workflow. During 00:14:03.320 --> 00:14:06.360 the test, we intentionally sabotaged the API 00:14:06.360 --> 00:14:08.679 key for the primary web search tool it was supposed 00:14:08.679 --> 00:14:11.399 to use. Now, an older model probably would have 00:14:11.399 --> 00:14:14.059 just failed, thrown an error, and stopped. But 00:14:14.059 --> 00:14:16.940 GPT -5 correctly identified the specific error 00:14:16.940 --> 00:14:19.779 it saw it was an authentication error. It automatically 00:14:19.779 --> 00:14:22.340 retried the tool once, just in case it was a 00:14:22.340 --> 00:14:24.879 temporary glitch. And when the error persisted, 00:14:24.899 --> 00:14:27.000 it intelligently switched to its backup plan. 00:14:27.340 --> 00:14:29.779 It used a different research tool, perplexity 00:14:29.779 --> 00:14:31.659 in this case, to get the information it needed 00:14:31.659 --> 00:14:34.100 and complete the mission successfully. That ability 00:14:34.100 --> 00:14:36.580 to recognize a problem, try to fix, and then 00:14:36.580 --> 00:14:39.860 execute a plan B. That's huge for building robust, 00:14:40.039 --> 00:14:42.019 reliable automation that doesn't break easily. 00:14:42.299 --> 00:14:44.820 Wow, okay. It can adapt to failures. What about 00:14:44.820 --> 00:14:46.940 the creative side, the AI art director test? 00:14:47.120 --> 00:14:49.000 How did that go? Right, so on the creative front. 00:14:49.519 --> 00:14:52.740 We compared GPT -4 .0 and GPT -5 again. We gave 00:14:52.740 --> 00:14:54.679 both models a really simple starting prompt, 00:14:54.799 --> 00:14:57.759 something like a shark wearing a cowboy hat on 00:14:57.759 --> 00:15:00.460 a classic car. And the task wasn't to generate 00:15:00.460 --> 00:15:02.960 the image itself, but to generate a much more 00:15:02.960 --> 00:15:05.480 detailed prompt that an image generation AI, 00:15:05.679 --> 00:15:09.019 like Midjourney or Dell E3, could use to create 00:15:09.019 --> 00:15:11.080 a really great picture. And the difference between 00:15:11.080 --> 00:15:13.960 the two was, well, night and day. How so? What 00:15:13.960 --> 00:15:16.559 made the GPT -5 prompt so much better? GPT -4 00:15:16.559 --> 00:15:18.799 acted like a, you know, a competent technician. 00:15:19.039 --> 00:15:21.399 It gave a solid literal description based on 00:15:21.399 --> 00:15:24.360 the input. Serviceable. Yeah. But GPT -5, it 00:15:24.360 --> 00:15:26.799 acted like a professional art director. Its generated 00:15:26.799 --> 00:15:29.700 prompt was incredibly rich, detailed, and evocative. 00:15:29.759 --> 00:15:32.519 It specified things like photorealistic cinematic 00:15:32.519 --> 00:15:35.340 wide shot, weathered leather cowboy hat, tilted 00:15:35.340 --> 00:15:38.230 at a jaunty angle. Mint condition, cherry red 00:15:38.230 --> 00:15:40.970 1950s American classic convertible, sun bleached 00:15:40.970 --> 00:15:43.289 desert highway, golden hour light, warm tones, 00:15:43.509 --> 00:15:45.850 long shadows, low angle composition, shallow 00:15:45.850 --> 00:15:48.409 depth of field, playful yet majestic mood, no 00:15:48.409 --> 00:15:50.370 text or watermark. It was just packed with specific 00:15:50.370 --> 00:15:52.750 visual and atmospheric details. That's, yeah, 00:15:52.850 --> 00:15:54.509 that's amazing. It really paints a picture with 00:15:54.509 --> 00:15:57.129 words. Exactly. And as you can imagine, the difference 00:15:57.129 --> 00:15:59.350 in the final image you'd get from those two tromps 00:15:59.350 --> 00:16:02.250 would be astronomical. One prompt gives you a 00:16:02.250 --> 00:16:04.250 basic picture. The other gives you a potential 00:16:04.250 --> 00:16:07.210 movie post or a whole story. This enhanced ability 00:16:07.210 --> 00:16:10.389 to generate really creative, detailed, effective 00:16:10.389 --> 00:16:13.730 prompts is a massive advantage for anyone who 00:16:13.730 --> 00:16:16.129 needs to automate content creation, whether it's 00:16:16.129 --> 00:16:18.070 for marketing, social media, anything visual. 00:16:18.250 --> 00:16:20.470 So it can adapt to errors on the fly and it can 00:16:20.470 --> 00:16:23.309 be incredibly creative. It's showing. deep intelligence 00:16:23.309 --> 00:16:26.070 and artistic flair. Yes, exactly. It's a combination 00:16:26.070 --> 00:16:27.649 we haven't really seen at this level before. 00:16:27.830 --> 00:16:29.730 So let's bring it all together. What does this 00:16:29.730 --> 00:16:32.929 actually mean for practical applications? With 00:16:32.929 --> 00:16:35.690 this level of AI, what kinds of real -world problems 00:16:35.690 --> 00:16:38.289 can we actually start solving now? Things that 00:16:38.289 --> 00:16:41.029 maybe seemed impossible or just too impractical 00:16:41.029 --> 00:16:43.750 before? Well, this new level of capability unlocks 00:16:43.750 --> 00:16:45.889 a whole range of practical automation opportunities 00:16:45.889 --> 00:16:48.649 that were maybe just out of reach. Think about 00:16:48.649 --> 00:16:51.690 level two customer service agents. AI agents 00:16:51.690 --> 00:16:54.230 that can handle genuinely complex, multi -step 00:16:54.230 --> 00:16:58.009 customer inquiries, not just simple FAQs. Agents 00:16:58.009 --> 00:16:59.830 that are fully integrated with your CRM that 00:16:59.830 --> 00:17:01.889 can look up order histories, process returns, 00:17:02.289 --> 00:17:04.670 troubleshoot issues, and importantly have a clear, 00:17:04.730 --> 00:17:07.329 seamless escalation path to a human agent when 00:17:07.329 --> 00:17:09.829 they hit their limit. Or imagine a content factory 00:17:09.829 --> 00:17:12.809 workflow. automating your entire content production 00:17:12.809 --> 00:17:14.769 pipeline from doing the initial research and 00:17:14.769 --> 00:17:17.230 fact -checking to generating drafts in multiple 00:17:17.230 --> 00:17:19.670 formats like blog posts, social media updates, 00:17:19.950 --> 00:17:22.349 video scripts, all with built -in checks for 00:17:22.349 --> 00:17:25.289 SEO and quality assurance. You could even build 00:17:25.289 --> 00:17:27.269 sophisticated business intelligence engines, 00:17:27.450 --> 00:17:30.130 systems that automatically pull data from various 00:17:30.130 --> 00:17:32.910 sources, aggregate it, generate concise executive 00:17:32.910 --> 00:17:35.589 summaries, maybe even create data visualizations 00:17:35.589 --> 00:17:37.759 on the fly. Okay, that's a serious amount of 00:17:37.759 --> 00:17:40.859 power. Which brings us back to strategy and cost. 00:17:41.039 --> 00:17:43.420 How do we apply this power effectively? What's 00:17:43.420 --> 00:17:45.519 the new playbook for building smart, cost -controlled 00:17:45.519 --> 00:17:48.220 automations with GPT -5? Right, because with 00:17:48.220 --> 00:17:51.519 great power comes potentially a great API bill 00:17:51.519 --> 00:17:53.740 if you're not careful. So, smart cost management 00:17:53.740 --> 00:17:56.059 is absolutely critical. The playbook has a few 00:17:56.059 --> 00:17:58.759 key parts. First, strategic model selection, 00:17:59.000 --> 00:18:00.460 like we talked about with the family of brains. 00:18:00.640 --> 00:18:02.539 You need to pick the right engine for the job. 00:18:02.819 --> 00:18:05.160 Use the powerful standard or pro models for tasks 00:18:05.160 --> 00:18:07.240 needing deep reasoning or complex instruction 00:18:07.240 --> 00:18:09.839 following. Use mini for high -volume, simpler 00:18:09.839 --> 00:18:12.859 tasks where cost per task is key. And use nano 00:18:12.859 --> 00:18:15.039 for the absolute simplest, most cost -critical 00:18:15.039 --> 00:18:17.619 operations. Don't use the most expensive brain 00:18:17.619 --> 00:18:20.519 for every single thought. Second, aggressive 00:18:20.519 --> 00:18:23.390 token management. Think of the AI's context window, 00:18:23.589 --> 00:18:25.569 the amount of text it can consider at once like 00:18:25.569 --> 00:18:27.950 a briefcase. Every word you put in or get out 00:18:27.950 --> 00:18:30.769 costs something. So optimize your prompts. Be 00:18:30.769 --> 00:18:32.569 ruthless about cutting out fluff and unnecessary 00:18:32.569 --> 00:18:35.250 words. Implement caching wherever possible if 00:18:35.250 --> 00:18:36.849 you're asking the AI the same question repeatedly 00:18:36.849 --> 00:18:39.710 or feeding it the same background info. Store 00:18:39.710 --> 00:18:41.390 the answer and reuse it instead of hitting the 00:18:41.390 --> 00:18:43.529 API every time. And consider hybrid approaches. 00:18:43.730 --> 00:18:45.970 Maybe use a cheap model like Mini for an initial 00:18:45.970 --> 00:18:48.890 pass, like filtering emails, and then only send 00:18:48.890 --> 00:18:50.849 the really important ones to the more expensive 00:18:50.849 --> 00:18:53.549 Pro model for detailed analysis. So being really 00:18:53.549 --> 00:18:55.430 disciplined about what goes into that briefcase 00:18:55.430 --> 00:18:58.569 and reusing information smartly is paramount 00:18:58.569 --> 00:19:01.609 for cost. Exactly. And the third piece is batch 00:19:01.609 --> 00:19:04.269 processing optimization. Think like an assembly 00:19:04.269 --> 00:19:07.690 line. Making one single API call to process, 00:19:07.750 --> 00:19:11.190 say, 10 customer reviews at once is far, far 00:19:11.190 --> 00:19:13.569 more efficient and cheaper than making 10 separate 00:19:13.569 --> 00:19:16.609 API calls, one for each review. So design your 00:19:16.609 --> 00:19:18.730 workflows to group similar operations together. 00:19:18.970 --> 00:19:21.390 Maybe implement intelligent queuing systems so 00:19:21.390 --> 00:19:23.930 non -urgent tasks can collect and then be processed 00:19:23.930 --> 00:19:26.529 in one large batch during off -peak hours. So 00:19:26.529 --> 00:19:28.950 it's really about strategic deployment, choosing 00:19:28.950 --> 00:19:31.269 the right tool, and being super smart and efficient 00:19:31.269 --> 00:19:34.619 with tokens and requests. Absolutely. Efficiency 00:19:34.619 --> 00:19:37.339 is what ensures profitability and allows you 00:19:37.339 --> 00:19:40.420 to scale these powerful AI solutions. Okay, let's 00:19:40.420 --> 00:19:42.259 zoom out one last time. If we connect all these 00:19:42.259 --> 00:19:44.759 dots, look at the bigger picture. The release 00:19:44.759 --> 00:19:46.980 of GPT -5 feels like more than just an update. 00:19:47.180 --> 00:19:49.119 It feels like a signal, maybe, of some major 00:19:49.119 --> 00:19:52.220 underlying trends in AI. From this vantage point, 00:19:52.359 --> 00:19:54.200 what does the near future look like? Yeah, it 00:19:54.200 --> 00:19:55.880 definitely feels like it's signaling at least 00:19:55.880 --> 00:19:58.920 three big shifts on the horizon. First, what 00:19:58.920 --> 00:20:01.119 the source calls the great leveling. This idea 00:20:01.119 --> 00:20:04.349 that... Truly advanced AI automation, the kind 00:20:04.349 --> 00:20:06.450 previously only available to huge corporations 00:20:06.450 --> 00:20:08.990 with massive budgets, is now becoming accessible 00:20:08.990 --> 00:20:11.710 to small businesses, even individual entrepreneurs. 00:20:11.990 --> 00:20:14.490 This allows them to genuinely compete, maybe 00:20:14.490 --> 00:20:17.230 even outmaneuver, much larger players by leveraging 00:20:17.230 --> 00:20:19.890 AI smartly. It's a democratization of capability. 00:20:20.730 --> 00:20:23.930 Second, we're getting closer to the JARVS build 00:20:23.930 --> 00:20:27.599 -me -a -workflow future. GPT -5's strong performance 00:20:27.599 --> 00:20:29.819 on coding benchmarks hints that we're rapidly 00:20:29.819 --> 00:20:31.799 approaching a point where you might be able to 00:20:31.799 --> 00:20:33.640 describe a complex automation you need in just 00:20:33.640 --> 00:20:36.200 plain English. And the AI system could actually 00:20:36.200 --> 00:20:38.460 generate a ready -to -run workflow for you, maybe 00:20:38.460 --> 00:20:41.240 in a tool like N8n. Imagine just typing, build 00:20:41.240 --> 00:20:43.460 me an N8n workflow that triggers every time I 00:20:43.460 --> 00:20:46.119 get a new email with an invoice attached, extracts 00:20:46.119 --> 00:20:48.400 the amount due and sender, adds it to my accounting 00:20:48.400 --> 00:20:50.319 spreadsheet, and sends me a confirmation message, 00:20:50.400 --> 00:20:52.109 and it just builds it. We're not there yet, but 00:20:52.109 --> 00:20:54.750 GPT -5 suggests it's getting closer. Wow. AI 00:20:54.750 --> 00:20:57.049 building AI workflows? And you mentioned a third 00:20:57.049 --> 00:21:00.190 trend, teams of AIs. Right. That leads to the 00:21:00.190 --> 00:21:02.690 third trend, the Avengers Assemble moment for 00:21:02.690 --> 00:21:05.869 AI. This is the potential shift away from trying 00:21:05.869 --> 00:21:09.250 to build one single monolithic AI that does everything, 00:21:09.410 --> 00:21:12.839 like a solo Iron Man. Towards orchestrating teams 00:21:12.839 --> 00:21:15.220 of more specialized AI agents, sometimes called 00:21:15.220 --> 00:21:18.859 agent swarms, that work together, GPT -5's much 00:21:18.859 --> 00:21:21.160 improved ability to reliably call and coordinate 00:21:21.160 --> 00:21:24.039 external tools makes these kinds of multi -agent 00:21:24.039 --> 00:21:26.599 systems far more feasible than they were before. 00:21:26.980 --> 00:21:29.259 The future might not be about building the single 00:21:29.259 --> 00:21:31.960 smartest AI, but about becoming like Nick Fury, 00:21:32.140 --> 00:21:34.339 the director, who knows how to assemble the right 00:21:34.339 --> 00:21:36.720 team of specialized Avengers, each with their 00:21:36.720 --> 00:21:39.440 own unique superpower, to tackle a complex mission 00:21:39.440 --> 00:21:42.210 collaboratively. So we're talking about AI building 00:21:42.210 --> 00:21:45.769 AI and potentially teams of specialized AIs working 00:21:45.769 --> 00:21:48.650 together on complex problems. Yes, it feels like 00:21:48.650 --> 00:21:50.609 the next frontier is really about AI collaboration 00:21:50.609 --> 00:21:53.150 and even self -assembly of solutions. Okay, so 00:21:53.150 --> 00:21:55.609 the bottom line here for you, the listener, really 00:21:55.609 --> 00:21:57.869 seems to be that GPT -5 isn't just another small 00:21:57.869 --> 00:21:59.869 step forward. It feels more like discovering 00:21:59.869 --> 00:22:02.009 a whole new continent for automation possibilities. 00:22:02.670 --> 00:22:04.349 I think that's a good way to put it. It brings 00:22:04.349 --> 00:22:06.750 dramatically improved reasoning, much better 00:22:06.750 --> 00:22:10.150 tool integration, and crucially, a more accessible... 00:22:10.160 --> 00:22:12.759 cost structure, especially with the family approach. 00:22:13.180 --> 00:22:15.519 It's fundamentally more reliable for critical 00:22:15.519 --> 00:22:18.640 tasks, and it enables more advanced, complex 00:22:18.640 --> 00:22:21.480 systems that require less constant human hand 00:22:21.480 --> 00:22:24.759 -holding. For platforms like NANN, this is massive. 00:22:25.019 --> 00:22:27.220 For years, you could argue that the flexibility 00:22:27.220 --> 00:22:29.500 of the automation platform itself was, in some 00:22:29.500 --> 00:22:32.099 ways, ahead of the intelligence of the AI brains 00:22:32.099 --> 00:22:35.480 we could easily plug into it. With GPT -5, it 00:22:35.480 --> 00:22:37.079 feels like the brain has finally caught up to 00:22:37.079 --> 00:22:40.089 the body. The potential is huge. So the race 00:22:40.089 --> 00:22:41.849 has definitely started. It seems the question 00:22:41.849 --> 00:22:43.789 is no longer if these tools will change pretty 00:22:43.789 --> 00:22:46.250 much every business, but really who will be the 00:22:46.250 --> 00:22:47.990 ones to actually build with them effectively 00:22:47.990 --> 00:22:50.529 and lead the way. Absolutely. And my advice, 00:22:50.690 --> 00:22:52.609 if you're listening and wondering where to start, 00:22:52.690 --> 00:22:55.609 start simple, but start now. Pick one important, 00:22:55.710 --> 00:22:58.529 tangible, real -world business problem you have. 00:22:58.869 --> 00:23:01.230 Try building an automation to solve it using 00:23:01.230 --> 00:23:04.430 these tools. Test it carefully. Refine it. Build 00:23:04.430 --> 00:23:07.289 one strong, reliable system first. Get comfortable 00:23:07.289 --> 00:23:10.390 with it. Then, slowly and confidently, grow your 00:23:10.390 --> 00:23:12.990 AI -powered capabilities from there. Don't try 00:23:12.990 --> 00:23:15.420 to boil the ocean on day one. That sounds like 00:23:15.420 --> 00:23:17.500 solid advice. Thank you for joining us on this 00:23:17.500 --> 00:23:20.039 deep dive into GPT -5. It's an exciting time. 00:23:20.119 --> 00:23:22.819 This new era of genuinely powerful, increasingly 00:23:22.819 --> 00:23:25.720 accessible AI is here. The time to start building 00:23:25.720 --> 00:23:27.900 is definitely now. Out to your own music.