WEBVTT 00:00:00.000 --> 00:00:02.640 There is a ghost in the machine right now. Oh, 00:00:02.700 --> 00:00:05.580 absolutely. As you listen to this, tens of thousands 00:00:05.580 --> 00:00:08.820 of people are using a massive unannounced AI 00:00:08.820 --> 00:00:11.619 upgrade. It's hidden deep inside their standard 00:00:11.619 --> 00:00:15.160 accounts. And well, the strangest part, OpenAI 00:00:15.160 --> 00:00:17.600 isn't saying a single word about it. They're 00:00:17.600 --> 00:00:20.339 completely silent, which is, I mean, it's wild 00:00:20.339 --> 00:00:24.100 because we're tracking an absolute flood of leaked 00:00:24.100 --> 00:00:26.920 data surrounding what the community is calling 00:00:26.920 --> 00:00:32.060 GQC 5 .6. pro. And today, our mission is to basically 00:00:32.060 --> 00:00:34.719 cut through that noise. Right. Welcome to the 00:00:34.719 --> 00:00:36.700 deep dive. We aren't just going to list off rumors 00:00:36.700 --> 00:00:39.380 today. We are going to trace the exact mechanism 00:00:39.380 --> 00:00:41.079 of the stealth rollout. Yeah, we're going to 00:00:41.079 --> 00:00:43.700 dissect how this hidden model is suddenly generating 00:00:43.700 --> 00:00:46.640 entire playable 3D worlds from just a single 00:00:46.640 --> 00:00:49.079 prompt. And ultimately, we are asking the defining 00:00:49.079 --> 00:00:52.299 question of this new AI era. OpenAI might have 00:00:52.299 --> 00:00:54.920 solved raw mathematical logic, but can they finally 00:00:54.920 --> 00:00:58.020 teach an AI to have actual taste? Exactly. But 00:00:58.020 --> 00:00:59.979 before we look at the mind -bending stuff this 00:00:59.979 --> 00:01:01.960 model is building, we really need to understand 00:01:01.960 --> 00:01:04.379 the delivery mechanism. Because if you open your 00:01:04.379 --> 00:01:07.840 dashboard today, you will not see a GPT 5 .6 00:01:07.840 --> 00:01:10.319 Pro button anywhere. You just see the standard 00:01:10.319 --> 00:01:14.840 lineup. You see GPT 5 .5, 5 .4, 5 .3, and the 00:01:14.840 --> 00:01:17.689 03 models. There is no beta banner. There is 00:01:17.689 --> 00:01:20.250 no splashy announcement. Right. Nothing. But 00:01:20.250 --> 00:01:22.510 savvy developers started noticing a distinct 00:01:22.510 --> 00:01:26.489 pattern. If you select the standard GPT 5 .5 00:01:26.489 --> 00:01:28.849 model and you toggle the intelligence slider 00:01:28.849 --> 00:01:31.430 up to high, something weird happens. Right. Because 00:01:31.430 --> 00:01:33.329 normally, that just gives you a slightly more 00:01:33.329 --> 00:01:35.310 thorough answer. Yeah, exactly. But suddenly, 00:01:35.310 --> 00:01:38.129 the behavior diverges. Most of the time, it is 00:01:38.129 --> 00:01:40.510 business as usual. But occasionally, the prompt 00:01:40.510 --> 00:01:43.430 just hangs. The generation time stretches out 00:01:43.430 --> 00:01:45.129 significantly. Oh, yeah. It takes way longer. 00:01:45.129 --> 00:01:47.609 And when the output finally lands, it is operating 00:01:47.609 --> 00:01:50.189 on a completely different level of logical sharkness. 00:01:50.450 --> 00:01:53.569 It is a classic A -B testing strategy. They are 00:01:53.569 --> 00:01:56.170 quietly routing a small percentage of live traffic 00:01:56.170 --> 00:01:58.530 to an unannounced checkpoint. And this quiet 00:01:58.530 --> 00:02:01.269 testing phase has ignited a massive speculative 00:02:01.269 --> 00:02:04.069 fire. Oh, completely. On Polymarket, which is 00:02:04.069 --> 00:02:05.849 a platform where people place real money bets 00:02:05.849 --> 00:02:08.490 on real -world events, traders have wagered over 00:02:08.490 --> 00:02:12.490 $1 .1 million. Wow. Yeah. 1 .1 million betting 00:02:12.490 --> 00:02:15.610 that OpenAI will officially launch this model 00:02:15.610 --> 00:02:19.830 between June 22nd and June 28th. That is a very 00:02:19.830 --> 00:02:22.669 specific window. It is, which raises an interesting 00:02:22.669 --> 00:02:24.650 point about how these betting markets operate. 00:02:24.689 --> 00:02:26.849 Yeah. People aren't just throwing a million dollars 00:02:26.849 --> 00:02:29.819 at a random hunch. Right. These traders are watching 00:02:29.819 --> 00:02:32.740 global server load spikes. They are monitoring 00:02:32.740 --> 00:02:35.900 brief structural leaks, like when a candidate 00:02:35.900 --> 00:02:37.680 checkpoint accidentally appeared on the design 00:02:37.680 --> 00:02:39.819 arena platform before getting hastily scrubbed. 00:02:39.939 --> 00:02:42.960 No, I remember that. Yeah. And they even track 00:02:42.960 --> 00:02:46.060 OpenAI's historical release cadence. That cadence 00:02:46.060 --> 00:02:48.340 has actually compressed to roughly a seven -week 00:02:48.340 --> 00:02:51.479 cycle between major updates. Late June aligns 00:02:51.479 --> 00:02:53.659 perfectly with that math. I have to pause there 00:02:53.659 --> 00:02:56.080 because that stealth nature feels almost counterintuitive 00:02:56.080 --> 00:02:58.180 from a traditional software perspective. How 00:02:58.180 --> 00:03:01.520 so? Well... Why test a flagship model so quietly? 00:03:01.800 --> 00:03:03.900 It's like ordering your standard daily coffee, 00:03:04.020 --> 00:03:06.520 but occasionally the barista slips you an experimental 00:03:06.520 --> 00:03:08.759 nitro cold brew just to see if your heart rate 00:03:08.759 --> 00:03:10.659 spikes. That is exactly what they're doing. If 00:03:10.659 --> 00:03:12.819 they ask your opinion, you overthink it. But 00:03:12.819 --> 00:03:15.099 if they just watch from the kitchen to see if 00:03:15.099 --> 00:03:18.199 you finish the cup faster, they get pure untainted 00:03:18.199 --> 00:03:21.599 data on the formula. Exactly. So why the unlabeled 00:03:21.599 --> 00:03:24.479 gap? Why not just call it a beta test and get 00:03:24.479 --> 00:03:26.840 deliberate user feedback? Because deliberate 00:03:26.840 --> 00:03:30.199 feedback is inherently biased. The moment you 00:03:30.199 --> 00:03:33.840 slap a beta or GPT 5 .6 Pro label on the interface, 00:03:34.259 --> 00:03:36.520 you introduce the observer effect. People act 00:03:36.520 --> 00:03:39.599 differently. Yes. Users immediately change their 00:03:39.599 --> 00:03:42.000 behavior. They try to break the model, they feed 00:03:42.000 --> 00:03:44.500 it impossible logic riddles, or try to bypass 00:03:44.500 --> 00:03:46.159 its safety filters just to see what happens. 00:03:46.259 --> 00:03:48.960 Right. They stop using it for normal work. Precisely. 00:03:49.060 --> 00:03:51.919 By running a blind A -B test, OpenAI captures 00:03:51.919 --> 00:03:54.780 how the model handles mundane, everyday tasks. 00:03:55.379 --> 00:03:58.020 Drafting a basic email, summarizing a boring 00:03:58.020 --> 00:04:01.639 PDF. That baseline data is the ultimate ground 00:04:01.639 --> 00:04:04.199 truth before a public launch. Testing the waters 00:04:04.199 --> 00:04:06.960 quietly to get raw data before the official splash. 00:04:07.240 --> 00:04:09.680 You nailed it. That's exactly the strategy. Okay, 00:04:09.719 --> 00:04:12.400 so we know they're quietly sorting this nitro 00:04:12.400 --> 00:04:15.199 -cold brew to a fraction of users. What happens 00:04:15.199 --> 00:04:17.439 when those users ask it to actually build something 00:04:17.439 --> 00:04:20.579 complex? Oh, this is where it gets crazy. Because 00:04:20.579 --> 00:04:23.819 the leaked demos aren't your standard text summaries. 00:04:23.939 --> 00:04:26.850 Not even close. Yeah. We are talking about fully 00:04:26.850 --> 00:04:30.430 interactive, playable environments. And the critical 00:04:30.430 --> 00:04:33.069 detail here is the architecture. These environments 00:04:33.069 --> 00:04:35.550 are running in single files, generated from a 00:04:35.550 --> 00:04:38.959 single prompt. A single prompt? Wow. Let's dive 00:04:38.959 --> 00:04:42.040 into the voxel in Rocket Scene. This is a complete 00:04:42.040 --> 00:04:45.800 3D house, and it is generated entirely inside 00:04:45.800 --> 00:04:50.779 one HTML file using WebGL2. And for anyone unfamiliar, 00:04:51.259 --> 00:04:54.399 WebGL2 is a tool for rendering 3D graphics directly 00:04:54.399 --> 00:04:56.300 inside your web browser. And what makes this 00:04:56.300 --> 00:04:59.420 specific voxel scene a breakthrough is its structural 00:04:59.420 --> 00:05:02.899 coherence. In previous AI models, one -shot 3D 00:05:02.899 --> 00:05:05.120 generation suffered from terrible object permanence. 00:05:05.160 --> 00:05:07.490 Right, it would just fall apart. Exactly. The 00:05:07.490 --> 00:05:09.709 moment you rotated the digital camera, the illusion 00:05:09.709 --> 00:05:11.970 shattered. The back of the house wouldn't exist, 00:05:12.189 --> 00:05:14.290 or the geometry would collapse into a mess of 00:05:14.290 --> 00:05:16.649 intersecting polygons. But with this hidden checkpoint, 00:05:16.930 --> 00:05:19.589 the structure actually holds. The architectural 00:05:19.589 --> 00:05:22.149 proportions remain mathematically sound. You 00:05:22.149 --> 00:05:24.089 can actually walk through the generated scene 00:05:24.089 --> 00:05:26.649 live. It's wild. And they didn't stop at simple 00:05:26.649 --> 00:05:28.939 houses either. No, they didn't. Testers built 00:05:28.939 --> 00:05:33.279 a Boeing 747 using 3 .js, which is a popular 00:05:33.279 --> 00:05:36.199 3D coding library. The spatial reasoning required 00:05:36.199 --> 00:05:39.800 to write code for a 747 is immense. It is. The 00:05:39.800 --> 00:05:42.459 AI isn't just painting a flat picture of an airplane. 00:05:42.560 --> 00:05:45.959 It has to deeply understand z -depth, aerodynamics, 00:05:46.339 --> 00:05:48.720 and structural spatial relationships. They also 00:05:48.720 --> 00:05:51.069 fed prompts into Blender. you know, the professional 00:05:51.069 --> 00:05:54.350 3D modeling software. And the AI generated a 00:05:54.350 --> 00:05:56.329 robot scene where the lighting and materials 00:05:56.329 --> 00:05:58.850 look like a panstakingly finished studio render. 00:05:59.069 --> 00:06:00.389 But here's where it gets really interesting. 00:06:01.029 --> 00:06:04.069 The crown jewel of this leak isn't a static 3D 00:06:04.069 --> 00:06:06.790 model. No, it's not. It is a functioning simulation 00:06:06.790 --> 00:06:11.430 game built in one HTML file. in 48 minutes. The 00:06:11.430 --> 00:06:14.149 SimStyle game is a phenomenal benchmark. It has 00:06:14.149 --> 00:06:16.410 working character movement. It has granting dialogue. 00:06:16.509 --> 00:06:19.069 It actively tracks the state of the digital world 00:06:19.069 --> 00:06:21.290 over time. It is wiring together an entire game 00:06:21.290 --> 00:06:23.850 loop, game physics, and user interface logic 00:06:23.850 --> 00:06:26.250 without a single human developer touching the 00:06:26.250 --> 00:06:30.110 code. Whoa. I mean, imagine generating an entire 00:06:30.110 --> 00:06:32.810 functioning simulation game, complete with physics, 00:06:33.069 --> 00:06:35.949 in a single file in 48 minutes. It really is 00:06:35.949 --> 00:06:38.089 hard to wrap your head around. The leap in context 00:06:38.089 --> 00:06:40.949 window management there is staggering. To hold 00:06:40.949 --> 00:06:43.250 the logic of a game state for nearly an hour 00:06:43.250 --> 00:06:45.430 without forgetting the rules it established in 00:06:45.430 --> 00:06:47.829 minute one? Well, that is a massive engineering 00:06:47.829 --> 00:06:50.230 fate. It is a monumental technical achievement, 00:06:50.870 --> 00:06:54.490 but there is a glaring catch that all the early 00:06:54.490 --> 00:06:57.490 testers keep highlighting. Ah. The catch? Yeah. 00:06:57.709 --> 00:06:59.689 While the underlying math, the structural code, 00:06:59.730 --> 00:07:02.129 and the physics are highly believable, the overall 00:07:02.129 --> 00:07:05.870 polish still trails behind Fable 5. Fable 5 being 00:07:05.870 --> 00:07:08.389 the current reigning champion among rival AI 00:07:08.389 --> 00:07:11.149 models for purely creative tasks. Exactly. When 00:07:11.149 --> 00:07:13.730 you look at the 5 .6 Pro outputs, they feel a 00:07:13.730 --> 00:07:16.389 bit robotic. They completely lack true creative 00:07:16.389 --> 00:07:18.629 taste. That is such a fascinating distinction. 00:07:18.829 --> 00:07:21.250 If this hidden checkpoint perfectly nails the 00:07:21.250 --> 00:07:24.000 complex structural logic and the physics, Why 00:07:24.000 --> 00:07:26.000 does it still feel robotic compared to Fable 00:07:26.000 --> 00:07:28.560 5? Well, if we connect this to the bigger picture, 00:07:29.100 --> 00:07:31.459 it illustrates the deep divide between organizing 00:07:31.459 --> 00:07:33.899 complexity and possessing an aesthetic soul. 00:07:34.439 --> 00:07:37.560 Like Y .6 Pro is a master architect. It can wire 00:07:37.560 --> 00:07:40.579 up a game loop flawlessly. Yeah. But Fable 5 00:07:40.579 --> 00:07:43.500 understands visual nuance. Fable 5 understands 00:07:43.500 --> 00:07:46.300 how light should feel in a room to evoke a specific 00:07:46.300 --> 00:07:49.779 mood. I see. GPT 5 .6 Pro is solving the math 00:07:49.779 --> 00:07:52.680 of the scene. Fable 5 is solving the art of the 00:07:52.680 --> 00:07:55.209 scene. Great at drawing the blueprints, but still 00:07:55.209 --> 00:07:57.810 missing that human artistic soul. That is the 00:07:57.810 --> 00:07:59.629 perfect way to look at it. Which naturally makes 00:07:59.629 --> 00:08:02.689 me wonder about its performance in a purely 2D 00:08:02.689 --> 00:08:05.389 space. Right. If it has perfect blueprints, but 00:08:05.389 --> 00:08:08.290 no interior designer in complex 3D environments, 00:08:08.730 --> 00:08:11.110 what happens when it tries to design a flat 2D 00:08:11.110 --> 00:08:13.569 website? How does it stack up against its direct 00:08:13.569 --> 00:08:16.470 predecessor, GPT -5 .5? So the community actually 00:08:16.470 --> 00:08:18.850 ran a brilliant direct comparison to test exactly 00:08:18.850 --> 00:08:21.610 that. They used a highly detailed spaceship design 00:08:21.610 --> 00:08:24.509 prompt. and the results exposed a major paradox. 00:08:25.290 --> 00:08:28.430 GPT 5 .6 Pro generated the image, but it ran 00:08:28.430 --> 00:08:31.970 for 87 minutes. Wait, 87 minutes for one single 00:08:31.970 --> 00:08:34.289 visual prompt? Yep, 87 minutes of continuous 00:08:34.289 --> 00:08:36.309 generation time. To put that in perspective, 00:08:37.070 --> 00:08:40.889 the older GPT 5 .5 model, running on extra high 00:08:40.889 --> 00:08:43.220 intelligence, completed the exact same prompt 00:08:43.220 --> 00:08:46.000 in 34 minutes and 42 seconds. I really have to 00:08:46.000 --> 00:08:48.559 point out the paradox there. An 87 -minute runtime 00:08:48.559 --> 00:08:51.159 isn't a feature you put on a marketing brochure. 00:08:51.179 --> 00:08:54.019 Definitely not. That represents a massive, almost 00:08:54.019 --> 00:08:57.179 unsustainable computing cost. If an AI takes 00:08:57.179 --> 00:08:59.179 an hour and a half to think through a visual 00:08:59.179 --> 00:09:02.139 prompt, the sheer volume of GPU cycles burning 00:09:02.139 --> 00:09:05.059 in the background is staggering. How do you scale 00:09:05.059 --> 00:09:07.980 an API that takes 87 minutes to answer one user? 00:09:08.059 --> 00:09:12.159 You don't. At least not yet. This massive resource 00:09:12.159 --> 00:09:14.519 burn is likely why it is hidden behind the A 00:09:14.519 --> 00:09:16.559 -B test rather than rolled out to everyone. That 00:09:16.559 --> 00:09:18.120 makes sense. But we have to look at what that 00:09:18.120 --> 00:09:22.220 87 minutes actually bought them. 5 .6 Pro definitely 00:09:22.220 --> 00:09:24.639 won on the micro -details, the specific lighting, 00:09:24.820 --> 00:09:27.159 the metallic shading on the spaceship, the intricate 00:09:27.159 --> 00:09:29.259 detail on the captive's chairs, and the exterior 00:09:29.259 --> 00:09:32.639 hull. Okay. It also produced far fewer visual 00:09:32.639 --> 00:09:35.100 glitches or warped pixels. But it wasn't a clean 00:09:35.100 --> 00:09:38.659 sweep. Right. GPT -5 .5 actually produced better 00:09:38.659 --> 00:09:40.759 interior rooms and far more compelling planets 00:09:40.759 --> 00:09:43.419 in the background. And again, the rival model, 00:09:43.620 --> 00:09:46.620 Fable 5, still beat both of OpenAI's models on 00:09:46.620 --> 00:09:49.539 the overall cohesiveness of the spaceship design. 00:09:50.039 --> 00:09:52.879 It really sounds like 5 .6 Pro is just an incremental 00:09:52.879 --> 00:09:55.340 update here, not the Fable 5 killer everyone 00:09:55.340 --> 00:09:58.299 was hoping for. For raw blank canvas created 00:09:58.299 --> 00:10:01.340 design, yes. It is purely incremental. Right. 00:10:01.549 --> 00:10:03.970 But the testers uncovered a completely different 00:10:03.970 --> 00:10:06.629 strength when they moved to design mimicry. They 00:10:06.629 --> 00:10:10.470 handed 5 .6 Pro a single reference image of an 00:10:10.470 --> 00:10:13.070 existing e -commerce landing page, and the model 00:10:13.070 --> 00:10:16.210 recreated it flawlessly. It nailed. the grid 00:10:16.210 --> 00:10:19.690 layout, the typography, the exact stylistic vibe 00:10:19.690 --> 00:10:21.990 of the original reference. I have to admit, I 00:10:21.990 --> 00:10:24.529 still wrestle with getting AI to match a specific 00:10:24.529 --> 00:10:27.169 design template. The prompt drift is real. Oh, 00:10:27.190 --> 00:10:29.529 it really is. You ask for a minimalist blue button, 00:10:29.750 --> 00:10:31.850 and three prompts later, the AI has decided your 00:10:31.850 --> 00:10:34.279 whole website should be neon purple. So seeing 00:10:34.279 --> 00:10:36.279 a model lock onto a visual template and hold 00:10:36.279 --> 00:10:38.860 it perfectly is genuinely impressive. It's huge 00:10:38.860 --> 00:10:40.639 for front -end developers. But does its success 00:10:40.639 --> 00:10:42.480 with the e -commerce page mean its true strength 00:10:42.480 --> 00:10:44.940 is just mimicry? This raises a critical question 00:10:44.940 --> 00:10:47.899 about the future utility of these models. Right 00:10:47.899 --> 00:10:51.159 now... Its absolute superpower in the 2D visual 00:10:51.159 --> 00:10:54.460 space is strict replication. OK. It performs 00:10:54.460 --> 00:10:57.059 exponentially better when you provide rigid guardrails 00:10:57.059 --> 00:10:59.919 and clear visual references. When you ask it 00:10:59.919 --> 00:11:02.980 to create from a pure blank page, it struggles 00:11:02.980 --> 00:11:05.940 to make cohesive stylistic choices. It needs 00:11:05.940 --> 00:11:08.039 you to define the aesthetic boundaries first. 00:11:08.360 --> 00:11:10.639 Better at strictly following the instructions 00:11:10.639 --> 00:11:13.139 than inventing a brilliant design from scratch. 00:11:13.220 --> 00:11:15.240 That is the reality of its current architecture, 00:11:15.379 --> 00:11:17.559 yeah. We will be right back after a quick word 00:11:17.559 --> 00:11:20.580 from our sponsor. Stick around. All right. And 00:11:20.580 --> 00:11:23.519 we are back. So we have established that this 00:11:23.519 --> 00:11:26.259 model needs strict instructions to design a standard 00:11:26.259 --> 00:11:29.419 website. But there is one highly specific visual 00:11:29.419 --> 00:11:32.600 format where it doesn't need to mimic, a format 00:11:32.600 --> 00:11:34.799 where it is genuinely shocking the developer 00:11:34.799 --> 00:11:38.659 community. SVG generation. This is the undisputed 00:11:38.659 --> 00:11:41.200 hidden superpower of the leaked checkpoint. Just 00:11:41.200 --> 00:11:43.899 to define that for a moment, SVG stands for scalable 00:11:43.899 --> 00:11:46.539 vector graphics. Simply put, it means scalable 00:11:46.539 --> 00:11:48.980 images drawn using mathematical formulas instead 00:11:48.980 --> 00:11:51.679 of individual pixels. Exactly. And because they 00:11:51.679 --> 00:11:54.120 are entirely mathematical formulas, generating 00:11:54.120 --> 00:11:56.259 complex lighting or shading is incredibly difficult. 00:11:56.279 --> 00:11:59.720 Right. When an AI draws a normal JPEG, it's basically 00:11:59.720 --> 00:12:02.379 just placing a dark pixel next to a light pixel 00:12:02.379 --> 00:12:04.899 based on training data. But with an SVG, the 00:12:04.899 --> 00:12:08.220 AI has to write pages of raw code to define gradients, 00:12:08.539 --> 00:12:11.100 light sources, and geometry. And the demo that 00:12:11.100 --> 00:12:14.919 broke the internet here was a BMW M4 CS. Yes. 00:12:15.340 --> 00:12:18.960 5 .6 Pro rendered an SVG of this car, featuring 00:12:18.960 --> 00:12:21.419 metallic shading, correct lighting reflections, 00:12:22.000 --> 00:12:25.299 and flawless perspective. It looked astonishingly 00:12:25.299 --> 00:12:27.379 close to a photograph, purely driven by math. 00:12:27.500 --> 00:12:30.279 They ran a direct head -to -head comparison against 00:12:30.279 --> 00:12:33.200 Fable 5. They pushed Fable 5 across its low, 00:12:33.240 --> 00:12:35.240 medium, high, and extra -high thinking levels. 00:12:35.379 --> 00:12:37.899 And what happened? Fable 5 failed entirely. It 00:12:37.899 --> 00:12:40.379 could only produce flat, cartoonish vector styles. 00:12:40.580 --> 00:12:42.340 It simply couldn't do the high -level metallic 00:12:42.340 --> 00:12:44.909 math. That is a definitive victory. And it wasn't 00:12:44.909 --> 00:12:47.429 just a car. Another tester prompted it to generate 00:12:47.429 --> 00:12:49.950 a Windows 11 interface. This one was crazy. It 00:12:49.950 --> 00:12:53.750 recreated the full operating system UI in SVG 00:12:53.750 --> 00:12:57.149 format. The file explorer, the task bar, the 00:12:57.149 --> 00:13:00.259 calculator app. all mathematically drawn. It 00:13:00.259 --> 00:13:02.620 cleanly outclassed another specialized model 00:13:02.620 --> 00:13:06.120 called Mythos. It did. But, as always with this 00:13:06.120 --> 00:13:08.700 checkpoint, there was a strange downside. There 00:13:08.700 --> 00:13:11.200 always is. It hallucinates extra interface elements. 00:13:11.980 --> 00:13:14.700 During the Windows 11 generation, it added bizarre 00:13:14.700 --> 00:13:18.279 unnecessary pop -ups and lines of text that simply 00:13:18.279 --> 00:13:21.049 do not exist in the real operating system. It's 00:13:21.049 --> 00:13:23.470 like an overly eager intern who gives you the 00:13:23.470 --> 00:13:26.049 pristine, incredibly complex spreadsheet you 00:13:26.049 --> 00:13:29.470 asked for, but then decides to add 10 confusing 00:13:29.470 --> 00:13:31.090 pie charts that you didn't need just to prove 00:13:31.090 --> 00:13:33.789 they could. Yeah. Why does a model this mathematically 00:13:33.789 --> 00:13:36.350 advanced throw in fake pop ups and random text? 00:13:36.490 --> 00:13:39.149 Because the model is heavily optimizing for extreme 00:13:39.149 --> 00:13:41.830 detail. It equates visual density with quality. 00:13:41.850 --> 00:13:43.889 Oh, I see. It has all this incredible processing 00:13:43.889 --> 00:13:47.149 power and structural understanding, but it completely 00:13:47.149 --> 00:13:49.330 lacks editorial restraint. It doesn't know when 00:13:49.330 --> 00:13:51.029 a design is actually finished and should just 00:13:51.029 --> 00:13:53.070 be left alone. Incredible attention to detail, 00:13:53.289 --> 00:13:55.610 but severely lacking an editor's restraint. It 00:13:55.610 --> 00:13:57.309 just wants to keep painting until the canvas 00:13:57.309 --> 00:13:59.909 is entirely full. So bringing all these bizarre 00:13:59.909 --> 00:14:03.289 technical quirks, the massive 87 -minute generation 00:14:03.289 --> 00:14:06.830 times, and these undeniable SVG superpowers together, 00:14:07.610 --> 00:14:09.909 where does this leave us in the broader AI arms 00:14:09.909 --> 00:14:13.259 race? Well, if we look at the honest benchmark 00:14:13.259 --> 00:14:17.440 comparison, 5 .6 Pro absolutely dominates on 00:14:17.440 --> 00:14:19.980 SVGs. It dominates on vision replication. And 00:14:19.980 --> 00:14:23.480 it wins heavily on deep game logic and code stability. 00:14:23.559 --> 00:14:27.000 OK. But it still trails Fable 5, as well as Claude, 00:14:27.159 --> 00:14:29.559 another major competitor, on standard front -end 00:14:29.559 --> 00:14:31.580 web generation and overall aesthetic polish. 00:14:32.240 --> 00:14:34.639 We should also caveat that Opus, another heavyweight 00:14:34.639 --> 00:14:36.860 model in the industry, wasn't benchmarked in 00:14:36.860 --> 00:14:38.960 this specific leak. And then there are the rumors 00:14:38.960 --> 00:14:40.860 floating around the edges of the technical data. 00:14:41.120 --> 00:14:43.879 The unconfirmed market noise. Pricing is a huge 00:14:43.879 --> 00:14:46.220 topic of speculation. The rumor mill strongly 00:14:46.220 --> 00:14:48.000 suggests the cost will sit somewhere right between 00:14:48.000 --> 00:14:51.340 Fable 5 and Opus 4 .8, while magically matching 00:14:51.340 --> 00:14:54.519 the price of the older GPT 5 .5 model. But the 00:14:54.519 --> 00:14:56.659 code name confusion is where the data gets incredibly 00:14:56.659 --> 00:14:59.840 muddy. Testers are tracking an entire constellation 00:14:59.840 --> 00:15:02.539 of names. You've got Iris Alpha, Ember Alpha, 00:15:02.919 --> 00:15:06.019 Beacon Alpha, Kepler, and Kindle. And strangely, 00:15:06.240 --> 00:15:08.039 some testers are reporting that the checkpoint 00:15:08.039 --> 00:15:10.919 named Kindle Alpha actually performs worse than 00:15:10.919 --> 00:15:14.179 the one named Kepler. Yet Kindle is supposedly 00:15:14.179 --> 00:15:17.220 the finalized release candidate. I really have 00:15:17.220 --> 00:15:19.360 to push back on the logic of that specific rumor. 00:15:19.980 --> 00:15:22.580 If Kindle Alpha is verifiably performing worse 00:15:22.580 --> 00:15:26.269 in testing, Why on earth would a multi -billion 00:15:26.269 --> 00:15:28.929 dollar company make that the flagship release 00:15:28.929 --> 00:15:31.250 candidate? Doesn't make sense. It really highlights 00:15:31.250 --> 00:15:33.870 how messy and contradictory these secondhand 00:15:33.870 --> 00:15:36.590 leaks really are. We have to treat the code names 00:15:36.590 --> 00:15:39.370 with extreme skepticism. The code names are a 00:15:39.370 --> 00:15:41.970 distraction, honestly. The underlying behavioral 00:15:41.970 --> 00:15:43.909 shift is the only thing that actually matters 00:15:43.909 --> 00:15:45.909 here. So zooming out and looking at the landscape 00:15:45.909 --> 00:15:48.710 today, is this hidden checkpoint the fable five 00:15:48.710 --> 00:15:51.009 killer the industry has been waiting for? Not 00:15:51.009 --> 00:15:53.750 entirely. It is closing the technical gap at 00:15:53.750 --> 00:15:56.690 a terrifying speed, especially regarding structural 00:15:56.690 --> 00:16:00.110 logic and complex SDG math. But it is absolutely 00:16:00.110 --> 00:16:02.389 not taking the creative crown across the board. 00:16:02.750 --> 00:16:04.990 Closing the distance fast, but definitely not 00:16:04.990 --> 00:16:06.850 taking the crown just yet. Exactly. So what does 00:16:06.850 --> 00:16:08.909 this all mean? Let's synthesize everything we've 00:16:08.909 --> 00:16:11.990 unpacked today. The overarching theme is that 00:16:11.990 --> 00:16:15.149 we are witnessing a live, highly quiet evolution. 00:16:15.470 --> 00:16:18.090 of artificial intelligence happening right inside 00:16:18.090 --> 00:16:21.509 our daily tools. The sheer fact that OpenAI can 00:16:21.509 --> 00:16:24.649 run this massive A -B test on live accounts without 00:16:24.649 --> 00:16:27.990 a single announcement shows how fluid and continuous 00:16:27.990 --> 00:16:30.309 this technology has become. And the specific 00:16:30.309 --> 00:16:33.909 capabilities of GPT 5 .6 Pro prove that a major 00:16:33.909 --> 00:16:36.490 historical milestone has been reached. Complex 00:16:36.490 --> 00:16:39.769 logic, deep physics, structural object permanence, 00:16:40.230 --> 00:16:42.509 building a one -shot HTML game that maintains 00:16:42.509 --> 00:16:45.480 state tracking for 48 minutes, these are no theoretical 00:16:45.480 --> 00:16:47.500 challenges, these are now solved problems for 00:16:47.500 --> 00:16:50.080 AI. The structural foundation is built, which 00:16:50.080 --> 00:16:52.820 means the new ultimate frontier for artificial 00:16:52.820 --> 00:16:56.019 intelligence isn't just raw computational capability 00:16:56.019 --> 00:16:58.940 anymore, it is taste. It is editorial restraint. 00:16:59.360 --> 00:17:01.480 It is knowing how to make a digital environment 00:17:01.480 --> 00:17:04.299 feel distinctly human rather than just functionally 00:17:04.299 --> 00:17:06.599 correct. It is the classic difference between 00:17:06.599 --> 00:17:10.039 building a house and creating a home. The AI 00:17:10.039 --> 00:17:12.819 can build the house perfectly now, but making 00:17:12.819 --> 00:17:16.619 it feel lived in? That is the next great technological 00:17:16.619 --> 00:17:18.740 leap. I highly encourage you to check your own 00:17:18.740 --> 00:17:22.059 dashboard. Switch your model over to GPT 5 .5. 00:17:22.140 --> 00:17:24.799 Set your intelligence slider to high. See if 00:17:24.799 --> 00:17:27.140 you can spot the slower, significantly sharper 00:17:27.140 --> 00:17:29.740 responses of that ghost checkpoint for yourself. 00:17:30.380 --> 00:17:33.140 experience that untainted A -B test firsthand. 00:17:33.319 --> 00:17:34.720 It's definitely worth trying. I want to leave 00:17:34.720 --> 00:17:36.660 you with the lingering thought to chew on. We 00:17:36.660 --> 00:17:38.619 talked about that ghost in the machine. If an 00:17:38.619 --> 00:17:41.799 AI can now quietly build complete physics -based 00:17:41.799 --> 00:17:44.180 3D worlds and functional simulation games in 00:17:44.180 --> 00:17:47.380 a single file, just by us asking, what happens 00:17:47.380 --> 00:17:49.559 to the value of human coding when that ghost 00:17:49.559 --> 00:17:52.619 finally develops real taste? That is the million 00:17:52.619 --> 00:17:55.220 dollar question. Until next time, keep diving 00:17:55.220 --> 00:17:55.480 deep.