WEBVTT 00:00:00.000 --> 00:00:02.339 Welcome to the deep dive. Yeah, thanks for having 00:00:02.339 --> 00:00:05.860 me. I have to say, the debate isn't about which 00:00:05.860 --> 00:00:08.839 AI coding tool is better anymore. It's really 00:00:08.839 --> 00:00:10.960 about knowing exactly which artificial brain 00:00:10.960 --> 00:00:13.140 to hire for the job right in front of you. I 00:00:13.140 --> 00:00:15.410 completely agree. I've been, um... I've been 00:00:15.410 --> 00:00:17.510 thinking a lot about this recently. We spend 00:00:17.510 --> 00:00:20.449 so much energy treating these models like rival 00:00:20.449 --> 00:00:22.969 sports teams. Oh, totally. Like it's a zero -sum 00:00:22.969 --> 00:00:25.789 game. Exactly. But today, we're taking a much 00:00:25.789 --> 00:00:28.609 more practical approach. We are examining an 00:00:28.609 --> 00:00:32.530 extensive, rigorous side -by -side test comparing... 00:00:35.739 --> 00:00:38.320 Right, and our mission here isn't to crown a 00:00:38.320 --> 00:00:41.159 singular champion. We want to unpack their unique 00:00:41.159 --> 00:00:43.820 underlying philosophies. Yeah, and run their 00:00:43.820 --> 00:00:46.259 real -world numbers on a complex research task 00:00:46.259 --> 00:00:48.719 and figure out how to combine them into one seamless 00:00:48.719 --> 00:00:50.659 workflow. It's gonna be a really fun breakdown. 00:00:51.280 --> 00:00:53.659 But before we get completely lost in the weeds, 00:00:53.740 --> 00:00:55.840 we should probably establish our baseline. For 00:00:55.840 --> 00:00:57.859 anyone listening, what is a coding agent? Let's 00:00:57.859 --> 00:01:00.380 just say it's AI that writes, tests, and edits 00:01:00.380 --> 00:01:02.719 software right on your computer. Yeah, that is 00:01:02.719 --> 00:01:05.299 the perfect distillation. Because we aren't talking 00:01:05.299 --> 00:01:07.439 about simple autocomplete anymore. You know, 00:01:07.640 --> 00:01:09.879 we are talking about a digital co -worker that 00:01:09.879 --> 00:01:13.159 can execute highly complex multi -step instructions 00:01:13.159 --> 00:01:15.939 across your entire file system. It's a fascinating 00:01:15.939 --> 00:01:19.239 shift in how we interact with our machines. But 00:01:19.239 --> 00:01:21.939 to really understand how these two specific tools 00:01:21.939 --> 00:01:24.459 diverge, we first need to look at the ground 00:01:24.459 --> 00:01:26.719 they actually share. Right. Framing them not 00:01:26.719 --> 00:01:29.579 as bitter competitors, but as distinct, highly 00:01:29.579 --> 00:01:33.319 specialized tool sets. Exactly. That is a crucial 00:01:33.319 --> 00:01:36.519 mindset shift for developers right now. So let's 00:01:36.519 --> 00:01:40.219 start by looking at Npropix Claude code. Yeah. 00:01:40.299 --> 00:01:42.180 So Claude is fundamentally designed from the 00:01:42.180 --> 00:01:44.939 ground up to act as a thinking partner. Like, 00:01:45.260 --> 00:01:48.120 it excels at complex architectural planning, 00:01:48.780 --> 00:01:50.959 intricate front -end design, and these heavily 00:01:50.959 --> 00:01:53.359 customized workflows. And what I find incredibly 00:01:53.359 --> 00:01:55.700 interesting about Claude's philosophy is its 00:01:55.700 --> 00:01:58.200 willingness to challenge you. Oh yeah, the pushback 00:01:58.200 --> 00:02:00.500 feature. Right. If you give it instructions that 00:02:00.500 --> 00:02:02.439 seem structurally flawed, it actually pushes 00:02:02.439 --> 00:02:04.879 back. It makes you pause and reconsider your 00:02:04.879 --> 00:02:07.060 approach before it writes a single line of code. 00:02:07.219 --> 00:02:09.439 Which is so valuable. And currently, it's running 00:02:09.439 --> 00:02:12.620 on their Opus, Sonnet, and Haiku models, depending 00:02:12.620 --> 00:02:14.969 on, you know, complexity of the task you give 00:02:14.969 --> 00:02:17.610 it. Right. Then on the other side of the spectrum, 00:02:17.830 --> 00:02:20.810 we have OpenAI's chat GPT codex. And we need 00:02:20.810 --> 00:02:22.409 to clarify this immediately, by the way. This 00:02:22.409 --> 00:02:25.270 is a brand new, highly advanced agent system. 00:02:25.650 --> 00:02:28.370 This is not the original codex model from way 00:02:28.370 --> 00:02:31.469 back in 2021. Right, right. This modern iteration 00:02:31.469 --> 00:02:34.389 of codex is engineered for pure unadulterated 00:02:34.389 --> 00:02:37.469 speed. Yes, absolutely. It doesn't want to debate 00:02:37.469 --> 00:02:40.050 architecture with you. It focuses entirely on 00:02:40.050 --> 00:02:42.789 rapid execution, designed to take well -defined 00:02:42.789 --> 00:02:45.310 tasks straight to the finish line as fast as 00:02:45.310 --> 00:02:48.129 possible. It is all about momentum. It runs on 00:02:48.129 --> 00:02:51.189 the primary GPT codecs model, but it also offers 00:02:51.189 --> 00:02:54.520 a lighter, highly responsive GPT codec spark 00:02:54.520 --> 00:02:57.060 version for when you just need blistering fast 00:02:57.060 --> 00:02:59.659 simple execution exactly So if we think about 00:02:59.659 --> 00:03:02.039 it in human terms Claude is like the thoughtful 00:03:02.039 --> 00:03:04.419 Architect asking you a dozen questions about 00:03:04.419 --> 00:03:07.000 the foundation while codex is the speedy form 00:03:07.000 --> 00:03:08.840 and just getting the bricks laid Oh, that's a 00:03:08.840 --> 00:03:11.120 great way to put it. That's exactly it. But despite 00:03:11.120 --> 00:03:13.759 those vastly different personalities They actually 00:03:13.759 --> 00:03:16.400 share a massive foundational baseline, right? 00:03:16.400 --> 00:03:18.960 For instance, both of them feature incredibly 00:03:18.960 --> 00:03:21.639 robust local code editing capabilities Yeah, 00:03:21.680 --> 00:03:25.000 and they both offer dedicated standalone desktop 00:03:25.000 --> 00:03:27.580 apps for Mac and Windows environments. They also 00:03:27.580 --> 00:03:30.280 integrate flawlessly into your existing setup. 00:03:30.780 --> 00:03:34.039 Both plug right into VS Code via dedicated extensions. 00:03:34.960 --> 00:03:37.259 And they both handle standard command line interface 00:03:37.259 --> 00:03:39.560 workflows without breaking a sweat. They both 00:03:39.560 --> 00:03:42.219 allow you to delegate those heavy compute -intensive 00:03:42.219 --> 00:03:44.580 tasks up to the cloud. And they share an open 00:03:44.580 --> 00:03:46.990 architecture for extensibility. meaning they 00:03:46.990 --> 00:03:49.689 both feature diverse marketplace plugins and 00:03:49.689 --> 00:03:52.349 utilize highly reusable markdown skills. They 00:03:52.349 --> 00:03:55.189 also both utilize MCP support. For those listening, 00:03:55.370 --> 00:03:58.289 MCP support is a protocol letting AI tools securely 00:03:58.289 --> 00:04:00.409 connect to your local files. Yeah, it essentially 00:04:00.409 --> 00:04:03.050 acts as a secure bridge between the AI's brain 00:04:03.050 --> 00:04:05.389 and your local hard drive. So with all these 00:04:05.389 --> 00:04:07.389 different versions, do their basic environments 00:04:07.389 --> 00:04:10.189 actually overlap? They really do. The underlying 00:04:10.189 --> 00:04:13.139 architecture is surprisingly compatible. Because 00:04:13.139 --> 00:04:15.300 they both hook directly into your standard desktop 00:04:15.300 --> 00:04:17.720 environment and utilize intelligent sub -agents 00:04:17.720 --> 00:04:20.579 for task management, moving between them isn't 00:04:20.579 --> 00:04:22.660 nearly as jarring as you might expect. So yes, 00:04:22.740 --> 00:04:26.220 both handle local editing, desktop apps, and 00:04:26.220 --> 00:04:29.459 sub -agents seamlessly. Exactly. The shared foundation 00:04:29.459 --> 00:04:32.740 is incredibly solid. The true divergence only 00:04:32.740 --> 00:04:35.160 really happens in how they build their specific 00:04:35.160 --> 00:04:38.300 workflows on top of that foundation. Claude's 00:04:38.300 --> 00:04:40.360 obsessive control over your workflow is great 00:04:40.360 --> 00:04:42.740 for planning, so I want to explore exactly how 00:04:42.740 --> 00:04:45.079 it gives you that granular, almost surgical control 00:04:45.079 --> 00:04:47.339 over your daily projects. This is where we have 00:04:47.339 --> 00:04:49.939 to talk about custom workflows. Yeah. Because 00:04:49.939 --> 00:04:52.420 this is the area where Claude code really separates 00:04:52.420 --> 00:04:54.639 itself from the rest of the pack. Right. To give 00:04:54.639 --> 00:04:56.620 you an idea of the scale, Codex offers about 00:04:56.620 --> 00:04:59.870 six basic hook events. Claude drops in with around 00:04:59.870 --> 00:05:03.009 30 different hooks. Wow, that is a staggering 00:05:03.009 --> 00:05:05.949 difference in operational capability. Hups, for 00:05:05.949 --> 00:05:07.949 anyone unfamiliar, are essentially automatic 00:05:07.949 --> 00:05:10.329 triggers. Yeah, little scripts that fire off 00:05:10.329 --> 00:05:13.269 invisibly whenever a very specific event happens 00:05:13.269 --> 00:05:15.350 within your coding session. And Claude gives 00:05:15.350 --> 00:05:18.170 you immense power here. It really does. For example, 00:05:18.290 --> 00:05:20.920 they have a pre -tool use hook. This mechanism 00:05:20.920 --> 00:05:23.379 actually intercepts a command before Claude even 00:05:23.379 --> 00:05:26.000 attempts to execute it. So you can use it to 00:05:26.000 --> 00:05:28.779 validate a destructive command or enforce strict 00:05:28.779 --> 00:05:31.420 security rules before the AI touches your active 00:05:31.420 --> 00:05:33.899 files? Precisely. And then on the flip side, 00:05:34.019 --> 00:05:36.139 you have the post tool use hook, which reacts 00:05:36.139 --> 00:05:38.699 immediately after a tool finishes its execution. 00:05:39.129 --> 00:05:41.329 I have to play devil's advocate here though. 00:05:41.910 --> 00:05:45.069 Why would an average developer need 30 different 00:05:45.069 --> 00:05:47.850 hook events? It sounds like you need a PhD in 00:05:47.850 --> 00:05:50.730 prompt engineering just to configure your workspace. 00:05:50.949 --> 00:05:53.339 I totally get that reaction. It sounds like absolute 00:05:53.339 --> 00:05:55.720 overkill at first glance. But think about the 00:05:55.720 --> 00:05:58.600 practical daily application of that power. OK, 00:05:58.699 --> 00:06:00.579 give me an example. Well, you could set a simple 00:06:00.579 --> 00:06:03.220 post tool use hook to automatically format your 00:06:03.220 --> 00:06:06.579 code through Prettier or ESLint. That means the 00:06:06.579 --> 00:06:09.060 formatting happens instantly and invisibly after 00:06:09.060 --> 00:06:12.310 every single file edit the AI makes. Oh, wow. 00:06:12.370 --> 00:06:14.269 So you don't even have to think about it. I still 00:06:14.269 --> 00:06:17.149 wrestle with prompt drift myself. So having a 00:06:17.149 --> 00:06:19.829 hook automatically pull the AI back on track 00:06:19.829 --> 00:06:22.149 sounds incredible. It really is. It just keeps 00:06:22.149 --> 00:06:25.689 the agent hyper focused on the actual goal. You 00:06:25.689 --> 00:06:28.850 can literally force Claude to request an entirely 00:06:28.850 --> 00:06:31.959 separate test suite from a sub agent. preventing 00:06:31.959 --> 00:06:35.519 it from trying to haphazardly write its own tests 00:06:35.519 --> 00:06:38.199 in the same breath as the feature code. Exactly. 00:06:38.720 --> 00:06:41.199 And speaking of subagents, both of these tools 00:06:41.199 --> 00:06:44.180 support them, but Claude Code ships with three 00:06:44.180 --> 00:06:47.180 distinct built -in types right out of the box. 00:06:47.480 --> 00:06:49.759 Yeah, Anthropic gives you the Explore Agent, 00:06:50.040 --> 00:06:52.540 the Plan Agent, and the General Popis Agent. 00:06:52.720 --> 00:06:54.939 Right, and the Explore Agent is designed to just 00:06:54.939 --> 00:06:57.680 quietly read and analyze your entire massive 00:06:57.680 --> 00:07:00.250 code base without altering anything. While the 00:07:00.250 --> 00:07:02.629 plan agent gathers all that deep context to present 00:07:02.629 --> 00:07:05.310 a highly structured architectural strategy. And 00:07:05.310 --> 00:07:07.170 then the general purpose agent steps in to handle 00:07:07.170 --> 00:07:09.790 the actual tasks requiring a mix of both reading 00:07:09.790 --> 00:07:13.129 and writing. And the cool part is Claude intelligently 00:07:13.129 --> 00:07:15.189 automatically delegates between these three. 00:07:15.430 --> 00:07:17.430 So you don't have to manually micromanage the 00:07:17.430 --> 00:07:20.730 handoffs. Not at all. Beyond the subagents Claude 00:07:20.730 --> 00:07:23.569 code features some incredibly powerful slash 00:07:23.569 --> 00:07:27.540 commands. The standout is slash ultra plan. Right. 00:07:27.740 --> 00:07:29.740 When you trigger this, it doesn't just process 00:07:29.740 --> 00:07:32.740 locally, it actually sends the entire deep planning 00:07:32.740 --> 00:07:35.439 stage up to a heavy cloud session. It builds 00:07:35.439 --> 00:07:37.860 a massive structural map and presents it to you 00:07:37.860 --> 00:07:40.759 in a browser UI, allowing you to review the architecture 00:07:40.759 --> 00:07:43.439 before committing. Yeah. And there is also the 00:07:43.439 --> 00:07:46.759 slash ultra review command. This runs a deep 00:07:46.759 --> 00:07:49.759 multi -agent code review across your entire project, 00:07:50.160 --> 00:07:52.980 returning incredibly detailed security and structural 00:07:52.980 --> 00:07:55.420 findings. It is worth noting that Pro and Max 00:07:55.420 --> 00:07:58.160 users only get free runs of this before it becomes 00:07:58.160 --> 00:08:00.899 a build feature, because it is so computationally 00:08:00.899 --> 00:08:03.480 expensive. Very true. We also have to highlight 00:08:03.480 --> 00:08:06.319 the slash loop command. This is fascinating. 00:08:06.360 --> 00:08:08.399 It runs recurring prompts on a set automated 00:08:08.399 --> 00:08:10.699 schedule. It essentially keeps Claude code in 00:08:10.699 --> 00:08:12.899 a state of continuous maintenance mode. Yeah, 00:08:12.899 --> 00:08:15.139 it just sits in the background, handling lingering 00:08:15.139 --> 00:08:18.699 PR comments, resolving frustrating merge conflicts, 00:08:18.959 --> 00:08:20.879 and cleaning up technical debt while you sleep. 00:08:21.199 --> 00:08:24.439 It's essentially an always on tireless developer 00:08:24.439 --> 00:08:27.290 assistant. Claude also recognizes that developers 00:08:27.290 --> 00:08:29.569 aren't always sitting at their keyboards. It 00:08:29.569 --> 00:08:31.810 securely connects to your mobile device via Telegram, 00:08:31.970 --> 00:08:34.899 Discord, or iMessage. Oh, that's wild. Yeah, 00:08:34.899 --> 00:08:37.220 you can literally just text your agent from your 00:08:37.220 --> 00:08:39.539 phone to check on a bill's progress while you're 00:08:39.539 --> 00:08:41.299 standing in line for coffee. It also provides 00:08:41.299 --> 00:08:45.200 an official agent SDK for both Python and TypeScript, 00:08:45.559 --> 00:08:47.580 which is a massive advantage for teams looking 00:08:47.580 --> 00:08:50.539 to build entirely custom internal systems. And 00:08:50.539 --> 00:08:53.419 for massive enterprise users, it integrates securely 00:08:53.419 --> 00:08:57.039 with Bedrock, Vertex AI, and Microsoft Foundry. 00:08:57.159 --> 00:08:59.740 Hooks give you surgical control to automate tasks 00:08:59.740 --> 00:09:03.179 like formatting code instantly. Precisely. It 00:09:03.179 --> 00:09:05.720 fundamentally changes how you trust the AI to 00:09:05.720 --> 00:09:08.539 run autonomously. Now, Claude's obsessive control 00:09:08.539 --> 00:09:11.019 over your workflow is great for planning, but 00:09:11.019 --> 00:09:12.779 sometimes you don't want a planner. Sometimes 00:09:12.779 --> 00:09:14.919 you just want to move as fast as humanly possible. 00:09:15.039 --> 00:09:18.320 Which brings us to OpenAI's philosophy with Codex. 00:09:18.480 --> 00:09:21.039 Right. It is built to move fast without breaking 00:09:21.039 --> 00:09:24.240 your existing work. Codex is fundamentally engineered 00:09:24.240 --> 00:09:28.559 for rapid, fearless execution. And the primary 00:09:28.559 --> 00:09:31.559 way to achieve this incredible speed safely is 00:09:31.559 --> 00:09:33.860 through highly isolated development environments. 00:09:34.399 --> 00:09:37.419 This brings us to native Git work trees. And 00:09:37.419 --> 00:09:40.120 just to clarify the jargon here, Git work trees 00:09:40.120 --> 00:09:42.500 are separate sandbox copies of your code, so 00:09:42.500 --> 00:09:45.460 changes never collide. Right. Now Claude can 00:09:45.460 --> 00:09:47.799 technically utilize work trees too, if you configure 00:09:47.799 --> 00:09:50.940 it manually, but Codex makes this feature entirely 00:09:50.940 --> 00:09:53.789 native. frictionless. It is seamlessly integrated 00:09:53.789 --> 00:09:56.309 right into the core building process. And this 00:09:56.309 --> 00:09:58.850 matters so much when you are trying to run multiple 00:09:58.850 --> 00:10:01.409 complex tasks at the exact same time. Right, 00:10:01.409 --> 00:10:04.149 because each separate task runs in its own completely 00:10:04.149 --> 00:10:07.409 isolated copy of the repository. Exactly. Your 00:10:07.409 --> 00:10:09.789 parallel experimental work never interferes with 00:10:09.789 --> 00:10:12.269 your main stable code base. And if an experiment 00:10:12.269 --> 00:10:14.769 fails entirely, you just discard that specific 00:10:14.769 --> 00:10:16.750 individual work tree. Yeah, you haven't touched 00:10:16.750 --> 00:10:19.200 anything else in your active project. It is an 00:10:19.200 --> 00:10:21.480 incredibly safe way to move fast and break things 00:10:21.480 --> 00:10:23.379 without actually breaking anything that matters. 00:10:23.759 --> 00:10:26.399 Codex also features a highly functional built 00:10:26.399 --> 00:10:29.360 -in desktop browser. After the agent builds a 00:10:29.360 --> 00:10:31.659 new web component, you can view the rendered 00:10:31.659 --> 00:10:34.399 result right there inside the agent interface. 00:10:34.659 --> 00:10:37.179 You can leave visual, pinpoint comments without 00:10:37.179 --> 00:10:39.639 ever switching windows or opening Chrome. It 00:10:39.639 --> 00:10:43.000 keeps the entire build and review loop cleanly 00:10:43.000 --> 00:10:46.419 inside one single unified environment. The computer 00:10:46.419 --> 00:10:49.139 use feature is also notably polished here. You 00:10:49.139 --> 00:10:51.740 can literally ask Codex to physically test a 00:10:51.740 --> 00:10:53.860 local application for you. It actually opens 00:10:53.860 --> 00:10:56.019 the application on your machine. Yeah, it actively 00:10:56.019 --> 00:10:58.740 clicks through the UI elements, finds visual 00:10:58.740 --> 00:11:01.480 bugs, and returns highly structured, detailed 00:11:01.480 --> 00:11:03.659 triage reports. Those reports include the bug 00:11:03.659 --> 00:11:06.019 severity, the expected behavior, and the precise 00:11:06.019 --> 00:11:08.539 steps required to reproduce the issue. It acts 00:11:08.539 --> 00:11:11.340 as an incredibly thorough automated QA process. 00:11:11.720 --> 00:11:14.100 Codex also connects directly and cleanly to your 00:11:14.100 --> 00:11:17.019 GitHub repositories. You simply tag at codex 00:11:17.019 --> 00:11:18.759 in a pull request comment. And it immediately 00:11:18.759 --> 00:11:21.379 spins up a cloud sandbox task right from that 00:11:21.379 --> 00:11:23.759 comment thread. triggering agent work completely 00:11:23.759 --> 00:11:26.620 frictionless for distributed teams. Both tools, 00:11:26.620 --> 00:11:29.000 it should be mentioned, also now support the 00:11:29.000 --> 00:11:31.100 slash goal command. Right, where you define a 00:11:31.100 --> 00:11:34.240 macro goal with a very clear hard stopping condition 00:11:34.240 --> 00:11:36.860 and the agent just keeps iterating until that 00:11:36.860 --> 00:11:39.679 exact condition is definitively met. But another 00:11:39.679 --> 00:11:42.679 massive, unique advantage for Codex's execution 00:11:42.679 --> 00:11:47.070 speed is visual assets. It has direct native 00:11:47.070 --> 00:11:50.570 integration with GPT Image 2 right inside the 00:11:50.570 --> 00:11:52.809 workflow. So it generates images as part of its 00:11:52.809 --> 00:11:54.769 primary execution. Yeah, if you need a quick 00:11:54.769 --> 00:11:57.950 product mock -up for a UI layout or a simple 00:11:57.950 --> 00:12:00.950 placeholder game asset, Codex handles it natively 00:12:00.950 --> 00:12:02.909 without skipping a beat. Whereas Claude currently 00:12:02.909 --> 00:12:05.590 requires an entirely external third -party setup 00:12:05.590 --> 00:12:08.690 to generate any kind of visual assets. So why 00:12:08.690 --> 00:12:10.970 is making worktrees native such a big deal for 00:12:10.970 --> 00:12:13.500 execution? Because it completely removes the 00:12:13.500 --> 00:12:15.840 friction of branching and stashing. It allows 00:12:15.840 --> 00:12:18.279 the agent to build fearlessly and rapidly by 00:12:18.279 --> 00:12:20.379 running multiple development threads safely at 00:12:20.379 --> 00:12:22.659 the exact same time. So they let the agent build 00:12:22.659 --> 00:12:24.480 parallel features without breaking your main 00:12:24.480 --> 00:12:27.360 code. Exactly. It completely neutralizes the 00:12:27.360 --> 00:12:29.940 paralyzing fear of catastrophic code errors. 00:12:30.039 --> 00:12:31.700 I think this is a perfect time to take a quick 00:12:31.700 --> 00:12:34.080 pause. Today's deep dive is supported by our 00:12:34.080 --> 00:12:36.519 friends at TechMinds. If you're building the 00:12:36.519 --> 00:12:38.639 next generation of software, you know keeping 00:12:38.639 --> 00:12:41.809 your team aligned is harder than ever. TechMinds 00:12:41.809 --> 00:12:44.110 provides the infrastructure you need to seamlessly 00:12:44.110 --> 00:12:47.169 integrate AI tools into your daily scrums. Their 00:12:47.169 --> 00:12:49.669 platform gives you full visibility into agent 00:12:49.669 --> 00:12:52.929 workloads, resource allegation, and project timelines. 00:12:53.450 --> 00:12:56.029 You can try it free for 30 days by visiting their 00:12:56.029 --> 00:12:58.970 website. Alright, welcome back. So, theories 00:12:58.970 --> 00:13:00.870 and distinct philosophies are always great to 00:13:00.870 --> 00:13:04.330 discuss. But I love looking at hard data. Me 00:13:04.330 --> 00:13:07.409 too. How do these tools actually perform when 00:13:07.409 --> 00:13:09.909 they are tasked with doing the exact same job 00:13:09.909 --> 00:13:12.259 under pressure? Well, the author of the newsletter 00:13:12.259 --> 00:13:15.039 ran a truly fantastic, highly structured real 00:13:15.039 --> 00:13:18.080 -world test. They tasked both agents with creating 00:13:18.080 --> 00:13:21.899 a highly branded, professional research PDF analyzing 00:13:21.899 --> 00:13:24.080 small -to -medium business automation tools. 00:13:24.340 --> 00:13:26.120 And crucially, the agents had to use active web 00:13:26.120 --> 00:13:28.480 search to gather the raw market data, synthesize 00:13:28.480 --> 00:13:30.799 it, and format it. Right. Let's look at the precise 00:13:30.799 --> 00:13:33.419 numbers for Codex first. It was running on the 00:13:33.419 --> 00:13:37.769 GPT 5 .5 model set to high performance. and Codex 00:13:37.769 --> 00:13:39.769 sprinted through the complex task, finishing 00:13:39.769 --> 00:13:42.210 in exactly eight minutes and one second. Along 00:13:42.210 --> 00:13:45.029 the way, it consumed approximately 2 .8 million 00:13:45.029 --> 00:13:47.990 tokens to process and generate the data. It successfully 00:13:47.990 --> 00:13:50.649 produced a very clean, blue, and gold structured 00:13:50.649 --> 00:13:54.289 cover page. It even generated a custom SMB mark 00:13:54.289 --> 00:13:56.649 natively, since it couldn't locate the original 00:13:56.649 --> 00:13:59.379 requested logo online. Yeah, and it proved to 00:13:59.379 --> 00:14:02.799 be fantastic at structuring raw, messy data into 00:14:02.799 --> 00:14:05.259 tight, readable tables. Now let's pivot and look 00:14:05.259 --> 00:14:07.740 at the CLOD results. It was running on the massive 00:14:07.740 --> 00:14:10.940 OPUS 4 .7 model, also set to high performance. 00:14:11.340 --> 00:14:13.759 CLOD took slightly longer, clocking in at 8 minutes 00:14:13.759 --> 00:14:16.269 and 15 seconds. But the token consumption is 00:14:16.269 --> 00:14:19.289 where things get wild. Yeah. It consumed a massive 00:14:19.289 --> 00:14:22.129 4 .7 million tokens for the exact same task. 00:14:22.309 --> 00:14:24.610 However, it actually successfully searched for 00:14:24.610 --> 00:14:27.789 and placed the correct AI fire logo. It also 00:14:27.789 --> 00:14:31.129 created a beautiful, highly nuanced dark orange 00:14:31.129 --> 00:14:33.929 gradient cover. It displayed much stronger overall 00:14:33.929 --> 00:14:36.350 brand fidelity throughout the document. And instead 00:14:36.350 --> 00:14:39.429 of just raw tables, it route the entire report 00:14:39.429 --> 00:14:42.919 in a compelling, flowing narrative style. But 00:14:42.919 --> 00:14:45.480 we have to talk seriously about token efficiency 00:14:45.480 --> 00:14:47.480 and the underlying economics here. Oh, absolutely. 00:14:47.820 --> 00:14:50.000 Because tokens are essentially the agent's battery 00:14:50.000 --> 00:14:52.399 life for the day. If you drain them too fast, 00:14:52.679 --> 00:14:54.659 you're done working. That is the perfect way 00:14:54.659 --> 00:14:58.039 to visualize it. Codex usage is completely included 00:14:58.039 --> 00:15:01.259 in standard chat GPT plans. That covers the free, 00:15:01.450 --> 00:15:05.070 Plus, and Pro tiers without any convoluted extra 00:15:05.070 --> 00:15:07.470 setup. Claude, on the other hand, requires a 00:15:07.470 --> 00:15:10.990 Pro subscription at $20, or the Max 5X tier at 00:15:10.990 --> 00:15:15.549 $100, or the Max 20X tier at $200 a month. And 00:15:15.549 --> 00:15:17.649 when you are burning tokens that fast, those 00:15:17.649 --> 00:15:19.830 subscription costs add up incredibly quickly 00:15:19.830 --> 00:15:22.519 for a team. To put the raw output into perspective, 00:15:23.000 --> 00:15:26.320 Codex generally outputs roughly 16 ,000 to 20 00:15:26.320 --> 00:15:28.980 ,000 tokens per standard task. Yeah, while Claude 00:15:28.980 --> 00:15:31.600 is outputting roughly 80 ,000 to 84 ,000 tokens 00:15:31.600 --> 00:15:34.139 per task. Claude hits those strict platform usage 00:15:34.139 --> 00:15:36.659 caps so much faster because of that massive context 00:15:36.659 --> 00:15:40.480 window. Whoa, imagine chewing through 4 .7 million 00:15:40.480 --> 00:15:43.259 tokens in just eight minutes. It's crazy. The 00:15:43.259 --> 00:15:45.639 scale is staggering. It really is mind -bending. 00:15:45.860 --> 00:15:48.340 It shows exactly why Opus is such an expensive 00:15:48.340 --> 00:15:51.639 model to run continuously. It is constantly rereading 00:15:51.639 --> 00:15:55.000 its own massive system prompt to maintain that 00:15:55.000 --> 00:15:57.419 architectural perfection. We also need to note 00:15:57.419 --> 00:15:59.860 a crucial platform policy difference regarding 00:15:59.860 --> 00:16:03.679 token usage. OpenAI officially supports and allows 00:16:03.679 --> 00:16:07.059 third -party API wrappers like OpenClaw or Hermes 00:16:07.059 --> 00:16:10.299 to manage these workflows. Right, whereas Anthropic 00:16:10.299 --> 00:16:14.820 requires explicit strict approval for any third 00:16:14.820 --> 00:16:17.340 party developers trying to tap into their agent 00:16:17.340 --> 00:16:20.000 workflows. Beyond the subscription price, why 00:16:20.000 --> 00:16:21.840 does the token difference actually matter to 00:16:21.840 --> 00:16:24.940 the user? It is entirely about your daily session 00:16:24.940 --> 00:16:27.559 longevity. Higher token output means you hit 00:16:27.559 --> 00:16:30.799 your strict daily rate limits rapidly. It completely 00:16:30.799 --> 00:16:33.100 stalls your workflow when you suddenly run out 00:16:33.100 --> 00:16:35.480 of computational battery life. right before a 00:16:35.480 --> 00:16:37.659 deadline. Higher token usage drains your daily 00:16:37.659 --> 00:16:40.240 session limits much faster. Exactly. It physically 00:16:40.240 --> 00:16:42.580 forces you to stop working entirely, which is 00:16:42.580 --> 00:16:44.940 a massive friction point for professional developers 00:16:44.940 --> 00:16:47.200 trying to stay in a flow state. So we've seen 00:16:47.200 --> 00:16:49.840 the careful, obsessive architectural planner 00:16:49.840 --> 00:16:52.759 in Claude. We've seen the incredibly speedy, 00:16:52.980 --> 00:16:56.120 fearless executor in Codex. So much of the discussion 00:16:56.120 --> 00:16:58.399 online is about declaring a winner. But how do 00:16:58.399 --> 00:17:00.720 we actually apply this knowledge to our work 00:17:00.720 --> 00:17:03.980 tomorrow morning? The big idea here is incredibly 00:17:03.980 --> 00:17:06.980 liberating, honestly. Don't pledge blind loyalty 00:17:06.980 --> 00:17:10.579 to just one app or ecosystem. Code is ultimately 00:17:10.579 --> 00:17:13.440 highly portable text. Your projects are not locked 00:17:13.440 --> 00:17:16.500 into proprietary formats. Your projects are entirely 00:17:16.500 --> 00:17:19.599 completely portable. The real long -term value 00:17:19.599 --> 00:17:22.680 in this new era isn't the specific AI tool you 00:17:22.680 --> 00:17:25.160 subscribe to. No, not at all. The true value 00:17:25.160 --> 00:17:27.859 is the underlying systems and the flexible pipeline 00:17:27.859 --> 00:17:29.880 skills you build along the way. Yeah, you aren't 00:17:29.880 --> 00:17:32.480 permanently locked into one single agent just 00:17:32.480 --> 00:17:34.660 because you spent six months learning its quirks. 00:17:34.799 --> 00:17:37.180 You can and should move between them fluidly. 00:17:37.309 --> 00:17:39.829 This leads us directly to the optimal strategy 00:17:39.829 --> 00:17:42.450 for modern development. It's the dynamic hybrid 00:17:42.450 --> 00:17:45.529 workflow. Yes. You open up Clog Code for the 00:17:45.529 --> 00:17:47.789 heavy architectural lifting up front. You use 00:17:47.789 --> 00:17:50.490 its deep planning capabilities, its custom hooks, 00:17:50.569 --> 00:17:53.430 and its interactive front -end UI designed to 00:17:53.430 --> 00:17:55.869 map out exactly what you want to build. Then, 00:17:55.950 --> 00:17:58.910 once the blueprint is rock solid, you seamlessly 00:17:58.910 --> 00:18:01.369 switch your project directory over to Codex. 00:18:01.430 --> 00:18:03.950 Right. You unleash Codex for pure execution, 00:18:04.410 --> 00:18:07.910 deep web research, generating structured PDFs 00:18:07.910 --> 00:18:10.509 and rapid GitHub shipping. Moving fluidly between 00:18:10.509 --> 00:18:13.430 these powerful agents based entirely on the specific 00:18:13.430 --> 00:18:15.769 task in front of you is the optimal strategy. 00:18:16.069 --> 00:18:19.069 You literally get the absolute best of both unique 00:18:19.069 --> 00:18:22.049 worlds without compromising on speed or quality. 00:18:22.240 --> 00:18:24.460 So, the ultimate takeaway isn't picking a winner, 00:18:24.660 --> 00:18:27.319 but building a pipeline. Exactly. You don't need 00:18:27.319 --> 00:18:30.079 a singular software champion. You need a functioning, 00:18:30.500 --> 00:18:32.980 highly efficient, and flexible development pipeline 00:18:32.980 --> 00:18:35.680 that leverages the strengths of every tool available. 00:18:35.859 --> 00:18:38.339 Right. Use Claude to plan the building and Codex 00:18:38.339 --> 00:18:40.720 to swing the hammers. That's the exact mental 00:18:40.720 --> 00:18:43.180 model developers need to adopt right now. Keep 00:18:43.180 --> 00:18:46.019 your projects easily portable, use Markdown extensively, 00:18:46.380 --> 00:18:48.759 and keep a remarkably open mind as these tools 00:18:48.759 --> 00:18:52.180 continue to evolve at breakneck speed. incredible 00:18:52.180 --> 00:18:54.799 breakdown of the current coding agent landscape. 00:18:55.220 --> 00:18:58.299 It really highlights exactly how fundamentally 00:18:58.299 --> 00:19:00.480 different these underlying philosophies are, 00:19:00.660 --> 00:19:02.460 even when they are trying to solve the exact 00:19:02.460 --> 00:19:05.180 same problems. It's a wildly exciting time to 00:19:05.180 --> 00:19:07.480 be building software. The WIP abilities are expanding 00:19:07.480 --> 00:19:09.920 every single week. I want to leave you with one 00:19:09.920 --> 00:19:13.069 final thought to mull over. OK. If we are already 00:19:13.069 --> 00:19:16.190 building hybrid workflows using Claude to meticulously 00:19:16.190 --> 00:19:19.150 plan the architecture and codecs to rapidly execute 00:19:19.150 --> 00:19:21.950 the code, what happens when we inevitably get 00:19:21.950 --> 00:19:25.349 an overarching AI agent whose only job is to 00:19:25.349 --> 00:19:27.369 automatically manage the workflow between these 00:19:27.369 --> 00:19:30.509 other agents? Oh, man, that is exactly when things 00:19:30.509 --> 00:19:32.750 are going to get truly wild. Try spinning up 00:19:32.750 --> 00:19:34.950 a simple local project today. Keep your files 00:19:34.950 --> 00:19:38.069 portable. Test out a custom hook or spin up a 00:19:38.069 --> 00:19:40.289 native work tree just to see how it feels. Yeah, 00:19:40.289 --> 00:19:42.690 just see how actively changes your daily workflow 00:19:42.690 --> 00:19:44.650 and your relationship with the code base. Thank 00:19:44.650 --> 00:19:47.329 you for joining us for this deep dive. We appreciate 00:19:47.329 --> 00:19:49.450 you spending your valuable time with us today. 00:19:49.710 --> 00:19:50.210 Keep building.