I Battle-Tested GPT-5.5 In My Business

Added: 2026-05-03

[00:00:00]People treat AI subscriptions kind of like gym memberships. They think that you have to pick one, stick with it because switching feels like a huge hassle, but it's really not. We ran OpenAI's new GPT 5.5 model inside Codex and compared it against Claude Code for a full week of real business work. And the truth is they're both good, but at different things. Almost like a left brain, right brain kind of thing that's going on. And I don't want you to panic switch though. So, this is a very realistic video explaining not just which one is better in which case, but how to think about the very concept of choosing an AI model for your work. I'm Mark Webster and I'm joined today by my co-founder Gil Braton at Authority Hacker where we test all of this AI stuff inside our business so business owners like you don't have to. So, Gil, GPT 5.5 was released just last week.

[00:00:58]They really seem to be pushing not just the coding which they seem to be getting better at these days, but things like research, data analysis, and documents in a lot of their marketing material for this. So, you've been hands-on, you've tested it, you've compared it to Claude Opus. What's your initial take?

[00:01:15]>> So, I actually went more than just testing it. I kind of like tried swapping completely for almost a week now and I think that's what makes it a bit more interesting cuz like you've probably already seen the benchmarks and all of that and it's not interesting. I like this model a lot, first of all. I just want to say for some tasks I'm going to default to it, but I'm not going to default to it for everything.

[00:01:34]The high-level TLDR is it's more thorough and organized than Opus. It's like, you know, for example, when I build these advanced skills with like lots of documentation, Opus can kind of get lost and start messing up with the file structure and it becomes messy uh when it edits the instructions, that kind of stuff. GPT 5 is better.

[00:01:54]>> What's causing that though? Cuz like that's is quite a specific uh thing to say. Like is it the fact that we've talked about this before that Opus has now a million context and then you get that kind of like dilution in the middle. Is that what you mean?

[00:02:08]>> No, even at 200k context it's more thorough. Like even at the same level of context. It's like think about the way I explain it to people like in my head and please don't I'm going to play with stereotypes so don't cancel me.

[00:02:21]Uh think about like they're both engineers. You have two engineers except Opus is an American engineer that's very chatty and easy to talk to and Codex is a German engineer that is like very rule-oriented. It's just going to be like, "That's the rules. That's the rules. We do the rules and that's how we do it, right?" And so for some workflows, you want the chattier model, particularly for like marketing workflows, etc. In terms of creativity of writing, copywriting, and even front end like design of websites, etc. like Opus is by far the best model. Uh like not even close and I'm not going to switch to Codex for that. And that's also why I think let's just preface with that. I think most people listening to this podcast who do marketing, who are not big engineers, etc. They probably should stick to Opus. And the 1 million context by the way is amazing. I On Codex you don't get 1 million context and you get these compactions back and all of that and it's just a little bit less nice. Um however, when you start having, as I said, like extended documentation, heavy logic, uh math involved, and data analysis, and that kind of stuff, I think Codex is a little better. It's a little better at staying on track. Like, you know, like Claude will go on side quests sometimes and it's like, "Oh, I see this little bug. I will just kind of like forget the main task and just kind of run into that thing and be like, 'I'm going to go fix that and I'm going to do that.'" And you're like, "Wait, what was the original prompt again? Uh what what happened?" Whereas Codex is much more likely to stay on track with that. So, it's like I like this model a lot for because I develop a lot of these skills right now and so there's a lot of complicated logic that I had to restear Opus back on track quite often and quite often I would end up with a lot of slop and I would have to reformat things and re-clean stuff up, etc. which doesn't happen as much with Codex. However, when I switch back my hats to like being a marketer and running business tasks, then I'm happy to reach out for Claude Code again and I do like it better. For example, I had to do a presentation recently. I wrote it both on Codex and Claude Code and Claude Code destroyed Codex, for example. So, talking about knowledge work, I would still reach out to Claude right now. What does GPT 5.5 let you do this week though that 5.4 7 days ago did not? Okay, there's a few things. First of all, it's nicer to use because it's way faster. You know, before you'd get even with it's faster than Opus. Like if you have Opus in, let's say, high thinking mode or extra high thinking mode, you will get outputs faster out of GPT 5.5 even in high reasoning.

[00:04:44]>> Can you quantify that? Like how much faster? I mean I you know, just >> Yeah, it's going to vary obviously, but let's say like by default 20 to 35% faster maybe. But not only that, but with Codex you can actually trade your usage limits for a fast version of the model. That's 1.5 times faster. So, let's say you have 1 hour left on your limits and you know like you're fine basically. You have quota. You can just press that button and you have like the super turbo version and then that's like 50% faster than Claude basically. And so that's very handy.

[00:05:16]>> Uh that's that's very cool. Like I I can >> Now you want it. I can see [laughter] I can see this being like a big monetization play for these models.

[00:05:24]We've talked about this in the podcast a lot over the last few months. You know, all these subscriptions are heavily subsidized. They're losing money on it, etc. But if they're able to add these extra features like, "Hey, don't wait.

[00:05:35]Get it now. Go faster." Like a lot of people are going to pay money in some cases for that. So, it's a nice little order bump upsell, I guess, in the >> But but in this case you get it. So, like Anthropic also has that for Opus, but it only uses extra usage. So, it's like you will just spend extra dollars on top of your subscription. OpenAI lets you spend your current subscription.

[00:05:55]Just it will deplete 2.5 times faster.

[00:05:57]So, that's quite a lot. However, talking about that, you get I would say like I'm on a $100 plan right now on Codex. And they are running a 2x promo so they give you two times more usage currently, but I still feel >> Until when? Until the end of May. So, another month. Okay.

[00:06:14]>> And I feel like I get more than three times more usage than I get out of a $100 Claude subscription. You get way more usage out of the model. Uh things that would take 10% on my Claude Code session often take like 3% on Codex, for example, on a normal mode.

[00:06:31]>> that bonus and because of model efficiency?

[00:06:34]>> I think you still get more usage even when the bonus will go away, you'll still get more usage basically. So, the base usage is quite a lot higher and so if you're like frustrated with limits on Claude, especially with the Opus 4.7 right now, it's kind of nice to just feel the breathing room. And then you can trade these limits for speed if you want. So, it's like if you're like, "I actually don't need more usage, but I get speed instead." You can do that, too. And so that's kind of yeah, the experience is feels like you have more control. One of the things that people ask us a lot on when we're doing like sales calls and things like that, you know, people that haven't been through our program yet is like they feel like they need to make a lifetime commitment to a single LLM or a single ecosystem, right? It's like, "Oh, even if it's like I want to use Claude Code." It's like, "Do I choose the Claude Code app?

[00:07:22]website?"

[00:07:23]>> They feel they have to. It's like the impression that that's like the a lifetime commitment. Yeah. Um but what's more the case that like deciding between Codex and Claude Code, that is not a permanent decision. In fact, it's not even a decision.

[00:07:36]>> Cuz you use both.

[00:07:37]>> Yeah, why not? I think one thing that people I'm going to say that for non-developers. People really struggle with the idea that the knowledge, the data is not in the app that you're using. It's in the files it's reading.

[00:07:49]These files are plain text files.

[00:07:51]Usually most people work with markdown when they work.

[00:07:54]>> Stored on your computer.

[00:07:54]>> Yeah, they are files that any app can read, right? It's like you can open them in your finder window on your Mac and read them. And so whether you open these files with the Codex client or the Claude Code client or with the VS Code client or with Warp or with any of these hundreds of editors that exist, it's the same folder. And therefore, if Claude Code writes something, Codex can read it and vice versa. So, often I have both working at the same time on the same files. Not exactly on the same files, but in the same repository because it's not an issue. And even then, right? Even the app you use to use Claude Code, for example, I use Claude Code both in a desktop app and in VS Code when I do knowledge work. Like and I reach out for the one that makes the most sense for the task at hand, right? If I want a more robust text editor and I want to collaborate with the model, I reach out for VS Code. If I want just a chat experience that is plus plus plus with all my MCP's knowledge, etc. I'll just jump on the Claude Code app because it's just like less hassle, less interface, etc. And it's like I don't feel like I have to pick one or the other. They're both open right now on my computer and why not? So, it's like yeah, because people are used to SaaS program that kind of lock their data, they think it's the same and it's not the same. Like that's kind of the value of these agentic workflows is that you are actually owning your data back and you're free to use it within whichever interface you want. Uh as I I know for developers it's going to be like super obvious and they'll be like, "Wow, okay." But for non-developers, we see a lot of people who struggle with that concept.

[00:09:24]>> All right. So, have you got any examples that you can show us?

[00:09:28]>> Yeah. Um so, we can see the difference here.

[00:09:30]>> So, I mean I run three skills. Actually did not run them. Codex ran them for me.

[00:09:33]So, again, with Codex like you're asking me like what can you do now that you can't do before? I've been able to build a much more robust testing system with Codex for like all the skills I build for AI Accelerator. And so it's able to run instances of Claude Code and Codex in parallel for me and AB test and do stuff like that and test the outputs and basically I have this whole thing and I said, "Hey, can you take three of our skills and then run them with the same prompt on Codex and on Claude code and then you run a judge that decides, you know, what's best for each version and make me a notion page with all the output. And that's the notion page you're looking at, right? It's like Codex. It says Claude code, but Codex did it because it's just integration I have on notion.

[00:10:13]And so like let's talk about them and actually the test that I picked they actually favor Opus a little bit more because they're more like creative writing task, but I took task that most people would do, right? So the first one was actually my social manager, the one that writes my social posts. Uh I did just the text, right? And so this is my GPT 5.5 output. So maybe if you check the LinkedIn post that talks about GPT 5.5 it decided to talk about that. Uh you know, it picked the test. But the interesting part about GPT 5.5 is not that it writes better answers, it's that OpenAI is positioning for works that happens outside the chatbot. Coding, research, analysis, documents, spreadsheets, software use, multi-tool workflows. That is the real shift. AI is moving into your desktop browser, code base, CRM, meta account. Uh which is also where the risk starts. The same agents can clean up your spreadsheet can also touch your live account. The same browser can research competitors can also log into tools with payment access.

[00:11:04]The same publishing workflow that saves your team 6 hours can publish the wrong thing while everyone is asleep. My rule for business owners is boring on purpose. Start with read-only access.

[00:11:15]Require approval, etc. etc. You get the idea. The post is not too bad. And that's one of the things I want to say about Codex is that it's not as good as Opus 2, right? But it's much better than previous GPT models.

[00:11:26]Yeah, so historically really for quite a long time now Claude Opus specifically, but all the Anthropic models have typically been regarded as being better for writing, for content, for blog posts, social media posts, those types of things. But you're saying that that gap has narrowed.

[00:11:43]>> I think it's still lower. I think it's still not as good. And we'll check the Opus answer in a second, but I think this is an acceptable post.

[00:11:50]It's a little bit dry, I would say. It has the the German-ness of it, you know.

[00:11:54]>> [laughter] >> Yeah, could you articulate why it's not so good? Like is it It's lacking emotion.

[00:12:00]It's lacking emotions. It's like usually the the way I prompted this skill it was like I want my social post to trigger an emotion in a reader. And so it's kind of like how you get higher engagement, basically. And so I feel this is very fact-based.

[00:12:17]It plays a little bit on the fear, but it doesn't kind of like twist the knife at the right time with like the right extra word adjective or something like that. It's very very dry. And so like that's when I'm like this is okay, but it is like a seven out of 10, you know?

[00:12:31]So and then I have the Opus output here and it's the same thing, the same topic, right? The chatbot era is ending. GPT 5.5 is being sold as a smarter assistant. OpenAI is positioning it as a model for real work. Coding, research, spreadsheet, multi-tool workflow. Codex now control your desktop, drives browser, holds memory across sessions.

[00:12:48]So you can see this kind of like this thing, like this sentence. Codex now controls your desktop, drives a browser, holds memory into sessions, runs multiple terminals, etc. Like this enumeration, this is more stylistic than what Codex did as a bullet list, for example.

[00:13:04]It has a little bit more flair in terms of literacy and like it just feels better, right? Does that basically mean that it feels more human?

[00:13:12]>> Yeah, it's just like you're more likely to read this sentence, this enumeration, the way Opus did it in a book. You know, that like if I read a a story or something, that's kind of how it would be written rather than that. And it's like, you know, it has this literary vibe, Claude, that you don't necessarily find as much in Codex.

[00:13:30]>> to push back on that a little bit though, like are you looking at this purely through the lens of like, you know, posting on social media and like that kind of content? Cuz like a lot of people would use this for, you know, emails or customer support or, you know, more like different types of content which So for social media this is important. For like support, Codex would be fine. The answer would be fine, right? Would it be fine or would it be better?

[00:13:53]>> It might be better because it's kind of sticking a little bit more. It's a little bit less stylistic in a way.

[00:13:57]Sometimes you want that, you know?

[00:13:59]Uh and they have the same instructions here, right? So it's like really the degree of how much flair they put into the writing is really inherent to the model cuz they got the same prompt. Is that the model or is that cuz there's a few variables you used have have control over this and like ChatGPT back in the day, things like top P and top K which like >> These are not really important anymore.

[00:14:17]Like people like you don't use them as much anymore. Even temperature, you don't use that much anymore. Like it's some models even have removed it in the API, right?

[00:14:25]Yeah, they kind of got rid of it, but have they got rid of that because they think they know how to do it better and their model only performs >> automatically. It's just it's done for you. But yeah, it's like I would say the Opus one like same like the same capabilities that make this useful make it dangerous. An agent that can publish the wrong thing. And it's like yeah, the sentences are slightly better. Like if I had to pick one, if you ask me pick one, I would pick the Opus one, actually. So it's like yeah, writing is good. Now the next one I did is I actually ran my new version of my meta campaigns skill that creates uh Meta ads campaign.

[00:14:58]>> yeah. It creates campaigns. It creates like the ad set. And again, I run it on the same prompts for GPT 5 and Opus. It basically tried to promote AI accelerator. I went for like stop chasing AI news, ship AI workflows. I made the prompt for the images, which is the more interesting part, I think. So GPT 5.5 made these images. So it made this first one, stop reading, start shipping AI workflows for your real business.

[00:15:20]>> Just to be clear, did it make the images or did they both use a different image model?

[00:15:24]>> They used the same image model. They prompted it themselves. So they write the prompts.

[00:15:30]Uh and so again, that's how you test the text model and its creativity, basically. Uh and you can see like there's quite a bit of um It's funny because it talks about meta campaign draft on the laptop screen. So you can see there's a little bit of a meta here where it kind of like repeated its own instruction that it got.

[00:15:45]Uh but it's okay. It's a bit of a generic ad, this first one. It's okay.

[00:15:49]Then it did a before-after from AI note to ship workflow with a before, which is like a very messy desk, like lots of posted notes, and after shipped, etc. Again, it's okay. It's a little bit generic still, I would say.

[00:16:01]>> It's still really impressive. I mean, I know >> I'm used to it now.

[00:16:05]This is the Codex thing, but it's only been out a couple weeks. But GPD image 2 model like it really has enabled these like multiple screenshots inside an image to be generated.

[00:16:13]>> Yep. And the next one is a phone with a notification that says AI workflows that leave the group chat, workflow ship, meta campaign, agent SOP, content brief, open workflow. And it's like a notification from AI accelerator.

[00:16:26]And there's a little arrow that says ship today. It looks like an iPhone. The interface looks good. But again, yeah, decent. And then there's a one that is a posting note that says stop chasing AI news and has AI news crossed over. It says ship one workflow this week with a little tick mark. It's kind of like on a messy desk. You see like the side of a keyboard, etc. It's okay. And finally did I was like posted note ads. They really work well like when people are just doom scrolling. Like it's it's like really good at like stopping people.

[00:16:52]>> It's almost like a motivational poster.

[00:16:54]It's like You know, it's like that it works the same. And then this one is a little bit generic, ship AI workflows, Claude code builds, Codex fixed shipped, agent doing ops. I don't know why I got the context.

[00:17:04]Again, Codex came up with the idea of the campaign. So it's like it's two levels removed from me at this point. So that's the ads created for AI So out of 10, what do you rate that campaign?

[00:17:16]Knowing this skill, it's a 6.5 to me. Okay. Um like And let's take a look let's take a look at the >> the angles are like okay, but a little bit generic. Like even if the ads look cool and the images are cool, like it feels a little bit stock photo-y to me. And uh so let's just look at the ones uh from Opus. And before that, let's look at the angle it took because obviously it's given a a URL, but it chooses the angle as well, right? It shows also stop reading the AI news, AI workflows that run, and from 47 bookmarks to one workflow. So yeah, it's like it actually shows us the first one shows a screenshot of your bookmarks on Safari.

[00:17:54]And says read 47 or read 47, used zero, ship one workflow instead, Auto AI Hacker AI Accelerator. It doesn't have our colors, but I don't know if he gave the URL in the prompt. So but I think the concept here is maybe the idea of people bookmarking and not using things is a little bit more advanced. Like you see how people bookmark our stuff when we show on social media and you know they never use it. This is a little bit more in the head of the customer.

[00:18:18]>> of all those healthy recipes I saved in my Instagram saved posts that I've never cooked.

[00:18:24]>> But you know this this sounds a little bit closer to reality to me than the previous ones. Like the angle. Like forget the image, the angle is a little bit better. Uh it's also I I don't know if if it's just got lucky, but you know, using kind of like orange text there it's like it's a bit more visually catching and you like your eyes are drawn to it more than the kind of like black and blue from Codex.

[00:18:45]>> Yep.

[00:18:46]The next one is tab overload to workflow. Stop collecting AI tools, start building system. And again, it's kind of the same idea of like a messy desk of before, you know, this week with like lots of tasks. And then has kind of like a terminal on the right that says this week but useful. I think it's this one is okay. Again, I don't know if it was given our URL because it didn't get our logo, which normally this skill gets the logo. So but I have a feeling I prompted it wrong. Just excusing the skill here because I've never seen that whenever I write firsthand. This one was run by Codex.

[00:19:15]Okay, but I would say this one is a little bit more generic. Like again, the other the bookmark stuff felt more in the head of the customer to me than this.

[00:19:23]A bit busy as well. Lots of like small text >> the ads it's not the best.

[00:19:29]Yeah. So it's like it's a little bit less good.

[00:19:32]Uh this one I like the composition of like essentially target market, I stopped reading AI Twitter, same AI Twitter rings very close to home. That's where all the AI news are happening. We ship more AI now, AI Accelerator member. Again, it would be nice with like the real name of a member, for example, maybe their photos or something like that, which we could do now. So we could probably dig it deeper, but that's probably to blame on the way the test was conducted rather than the skill's ability. Uh if I provided the photo, it could do that, no problem. This is a decent ad. I like that. I like the AI Twitter as well. I think this is like if you run that on Twitter, that would be a great ad. Yeah.

[00:20:07]And then another kind of before after as well that says you every Sunday AI tool stack and has like ChatGPT, Perplexity, Claude, Midjourney, Copy AI, Tavily, Face I like that they brainstormed all these names actually. These are all real tools actually. And then it has the Ship Before Coffee, Niche Validated, Content Cluster, AI Content Publish, Backlink Build, Automation Live, Lead Magnet Live, Email Flow Set, Week Complete basically.

[00:20:31]Okay, not bad. I like the details. I think the details and the way it thought about the details is good. The angle is again maybe a seven out of 10.

[00:20:39]And instead of the post-it note, it did a notebook here.

[00:20:42]Uh what doing AI looks like AI ad creative agent uh internal SOP bot, Codex content drafts, maybe not, maybe Opus is better. Uh lead gen scraper, weekly ops report, auto AI hacker AI accelerator. Okay too. What do you think? Just before I give you my answer, like what would you rate this one out of 10? This campaign, the Claude one. Uh I think it's a little bit higher mostly for the testimonial one and for the bookmarks one. I think these two are hitting to me because they feel like realities of the target market. The other ones, they fall into the cliche a little bit. So, it's like It's worth pointing out like with these kind of campaigns, you're not aiming for like to use 100% of them. You probably throw away half of them and then like most of them you still probably want to edit and regenerate with a few, you know, like >> were 10 cents each, right? It's like I think we can afford to throw a few away.

[00:21:30]Yeah. And so >> So, my take is that the Claude version has in a couple of those images it's evoked an emotion in me that's like, yeah.

[00:21:42]It gets it. It gets it. Whereas it felt like the Codex one, the Open AI one that you showed first was more like, oh, I'm doing an ad because I'm told to do an ad. Like, you know when you have an employee that like they're not bad but they're just kind of like doing the work. It's fine. It's okay. It's like in many cases that's perfectly fine.

[00:22:04]Um but for something like ads where it really does make a difference when, you know, that final 5%, 10%, 1% effort at the end which is the hardest to do it makes all the difference in the world and it feels like Claude it's not always doing it, but it's doing it more than Open AI in in this specific context.

[00:22:25]>> And now you see why I still recommend Claude for marketers.

[00:22:28]Um like even though I'm happy to use Codex to like do my technical stuff, Claude for marketers is like it hits more often in my opinion. Uh and this is on Opus 4.7 which is a worse writer than 4.6. I should probably run the same on 4.6 just to see what happens.

[00:22:44]Uh but yeah, like I think this is a great example. So, all of that considered then, you know, you said you use both at the moment. Can you just explain which tasks this week you've done on Codex and which you're doing on Opus?

[00:22:56]>> Okay. So, um like all the technical stuff of building these skills, uh documenting them like in a very thorough way, having the logic, having no contradictions, removing the fluff, that kind of stuff, I've moved to Codex for that and I like it better. Anything that requires me doing marketing, so for example, I had to do a webinar uh that I presented yesterday. I had to make a slide deck for example. I actually made the slide deck both on Codex and on Claude Code and Claude Code smashed it and Codex sucked in the output. Same for writing things, same for like uh anything where I need to write, right?

[00:23:31]But like for example, I would probably reach out to Codex if I wanted to like analyze either financial data or marketing data. Let's say I want to crunch analytics data.

[00:23:41]Then probably I'd reach out to Codex right now. And anything where it's going to write code, right? If it's going to write scripts etc. like I think Codex might be a little bit less messy. Like I have yet to have a thing that I do with code with Codex that doesn't one shot.

[00:23:55]Like I have never had something not work right out of the box with Codex. With Codex.

[00:23:59]>> So, what I'm hearing is uh there's a filter in your brain and the more logical left brain, right brain, you know >> [laughter] >> Yeah, and the more kind of uh imaginative, creative uh task you go with uh Open >> And I want to say as well for websites No, no, for websites, Opus is miles ahead in terms of like how good any website will look It's like it's not even close. Do not use Codex to build websites. So again, we have a course on building websites with Claude. Claude smashes Codex like by even more than what you see on these ads etc. on front end and in taste basically. Okay, last question on this then. Take a task which kind of sits in the middle. So, something like SEO on-page analysis, keyword research, topical map, recommend, kind of stuff. There's There's some data, there's some creativity, yeah. Yeah, yeah. So, I mean do you want to ask me how I would do it or how should the target market audience do it? Or or how you're doing it.

[00:25:01]>> Okay, how >> that's what people want to know. So, here's how I would do it. It's like I have a Codex skill on Claude Code and basically Claude Code can call Codex for help.

[00:25:11]>> [laughter] >> So, probably I would run my main thread on Claude mostly because of the 1 million token which is very nice when you're on these kind of tasks. You just have a small roomie. And then I can say like let's say I make a skill for it.

[00:25:22]Like I make a skill on like analyzing data and then just changing the website, let's say.

[00:25:26]Probably what I'll say is like I'll say, "Hey Claude Code, for the analysis part, you just give that to Codex. You give it the information." I will just basically write the prompt for Codex. Codex will process, give the answer back to Claude Code that will then do something with that information. So, Codex can do all the analysis. It will come back to Claude Code and Claude Code can do the front end edits for the website if there's something to do with the analytics data for example. And then I can get the best of both Just to be clear on that, when you're using your Codex skill from within Claude Code, that's using your subscription as opposed to API costs.

[00:26:01]>> Yep. Okay, that's interesting.

[00:26:02]>> can have two subscriptions. And I think a lot of people it makes sense to have the $20 ChatGPT subscription because you get some Codex usage. It's kind of like enough to use it as a helper to Claude.

[00:26:13]And so, you can kind of like make up for the weaknesses and messiness of the model. So, even if I did a big Let's say I do I still still build my skills on Claude. I would probably say, "Hey, when you're done with the edits, have Codex review everything and make sure that everything's tidy and give you the feedback."

[00:26:27]And I will just kind of like review all the files, give the feedback to Claude and Claude will be like, "Oh, yeah, Codex found all these issues. Let me fix them myself." And I will basically do it. And what's really powerful is it can because there's essentially session IDs, when they talk to each other, it can go back and forth. So, think about Claude Code using ChatGPT basically.

[00:26:45]>> [laughter] >> And it can And Codex can come back with something Claude is like, "Oh, really?

[00:26:48]That makes me think of this. Let me ask it what it thinks about this." And it just goes back and it's the same discussion. And it's basically giving Codex access to Claude Code and then it can outsource tasks. Interesting. Okay, so [snorts] would this be a way of actually like not using up your Claude subscription as fast cuz like I know a lot of people have have >> Potentially.

[00:27:08]Depends on your tasks, right? And that's the thing. It's like I have a Codex skill in Claude Code and I have a Claude skill in Codex.

[00:27:15]And so, the point is I can kind of balance two subscriptions. So, right now what I've done actually, I've downgraded my Entropic subscription to the max 5X, so the $100. Uh I used to be on the 200.

[00:27:26]I'm running Codex because I do a lot of skill development, I'm kind of running a lot of Codex as a main thread. So, Codex is at the $100 too and remember has like more usage than almost the $200 plan on Claude right now. And the point is like if I hit the limits on Claude, I'm just going to re-upgrade to 200 bucks basically. Uh I'm 99% sure I will not hit my limits on Codex unless I run it in fast mode all the time. And then yeah, that's pretty much how I'm going to be setting myself up. So, when you set up Claude Code for the first time, you have like your Claude.md file and like there's a few specifics to that.

[00:27:58]As I understand As I understand uh in Codex, that same file is called agents.md. Like how do you If you're working with both of them at the same time, how do you manage that?

[00:28:08]>> You can either duplicate it. You can You could duplicate your Claude.md and just call it Copy Paste the File and Rename It. Duplicate Right click duplicate, agents.md, boom, it works. Or you can write in your agents.md, read Claude.md.

[00:28:19]Uh >> [laughter] >> then it's just Yeah, yeah, why not?

[00:28:22]Okay. Remember So, that's that progressive disclosure that you talked about before.

[00:28:26]>> Well, it's just a way to have one source of truth because let's say you duplicate your Claude.md at one point and then you make some edits to it, then they are out of sync. It's kind of nice to keep just one file. Um and then it's just one source of Does Does that definitely work though cuz like let me just explain my understanding of what the Claude.md or agent.md file um does. Like every time you send a message the contents of that file are sent behind the scenes with the message.

[00:28:50]every message you send.

[00:28:51]>> how it works. Which is one of the reasons why you don't want to have it too long.

[00:28:54]Um but if all it does is say read agents.m or read Claude.md, is it just that text being sent or is it then going to actually go and do that? Do you know what I mean?

[00:29:05]>> Yeah, it's Well, think if you typed in the chat window, if you typed, "Go read Claude.md," what would happen?

[00:29:11]It would do it. Okay, so what's going to happen when you append this to your prompt?

[00:29:16]Well, I mean I guess that's my question.

[00:29:17]Is it always going to do it when it's cuz I don't know how it's coded when it does it in the back end.

[00:29:22]>> to do it basically. It's just going to go and And just you'll see a tool call like read Claude.md basically. Instead of loading it into the context, it will load that prompt in the context and it will just have one more tool call at the beginning and do it basically. So, like there's multiple ways to do that. Also for the skills, right? Remember they're under the dot Claude folder, but you can just go in Codex and be like, "Hey, uh set yourself up so you're synced with the Claude skills." And it will create what is called a symlink which is essentially like kind of like a shortcut on your desktop. You know how a shortcut on your desktop is not really the application, it's just telling where the file is.

[00:29:52]Same thing. And so the idea is you can keep that in sync as well and keep one source of truth rather than just duplicating everything. And badabim badaboom, everything works. Badabim badaboom. It's like it takes literally 3 minutes to be able to use both. I just want to put it out there as well cuz you know, you're on YouTube or social media, you'll know that I've seen everybody saying uh you know, GPT 5.5 is amazing, it's insane, it's crazy, and you have to switch, etc. Um but I think this like constant switching back and forth can be a bit of a a kind of distraction for a lot of business users, business owners.

[00:30:25]>> For business owners, 99% of them ignore it, stick to Claude, and keep going. And I just want to say that I have not switched. I'm still just using Claude code cuz it's simple and easy, and I just don't have the time to like If I wasn't doing technical things, I would stick to Claude code only as well. It's just like this is nice because it's just like it's a lot less hand-holding and I'm able to just give it more work without me baby sitting it, basically.

[00:30:51]And to be honest, like I am doing some I mean, they're not super technical like website changes and landing pages, things like that. And well, not it's duplicating it, but it's relatively simple stuff.

[00:31:02]>> Yeah, it's fine. So I'm I'm like Yeah, yeah. It's That's That's my point is like you don't need to like spend the extra time to to get like a little bit extra horsepower out of it.

[00:31:11]>> I agree. It's like it's fine. It's just the different tools and they have different values and they might be good for different things. And the reality is most workflows can be done pretty well on both anyway. Mhm. Okay, so we're going to talk about Codex the app in just a sec cuz like a lot has changed there. But I first want to just briefly give a shout out to our AI accelerator.

[00:31:31]So we launched this in October last year and we've just hit 600 members in there.

[00:31:37]So thank you very much for everybody who has signed up and is a member in there, is part of that community. We have courses, training, uh loads and loads of information around how to actually use this stuff in your business. So you don't have to be technical, you don't have to be a developer. This is for business owners who don't have much time. So if you want to get started using Claude code, a lot of the skills that you've shown here, the meta ads one, you can literally download that as soon as you you sign up and it it just runs.

[00:32:06]>> You need the API. It's you need to get the API, yeah.

[00:32:08]>> magical in a way, yeah. Um but until a week ago, we only offered this program on an annual subscription. And part of the reason that we did that is we wanted to have serious people in there so that the quality of the community was good.

[00:32:23]And price is a way to filter your community. So we have just launched last week monthly plans.

[00:32:31]But we're just doing it as a trial cuz we want to see how that affects things.

[00:32:36]So if you would like to join, until So this podcast is going out today is is Thursday, the 30th. So it's until Monday the 4th of May, you can join on a monthly subscription. It's $127 per month and there's no tie-in. You can cancel after You can go in, watch all the courses, download all the skills, and cancel in your first month. Please don't [snorts] do that. It's only $127.

[00:33:02]Please don't do that, [laughter] but like if you did want to do that, that is now possible. So we really like want to like take the barrier away and make this like an absolute no-brainer.

[00:33:12]>> to see what's inside and then you decide, do I want more of these? Do I want to stay or not? It's really it, right? But at least you get >> And then you can always upgrade to the annual plan and save a bit of money a few months down the line. So yeah. So if you head on over to authorityhacker.com/aiaccelerator, all the details are in there. But you have to do it by Monday, midnight on Monday, the monthly plans will close. If we do bring them back in future at a later date, they will be more expensive as well. So you know, this is the cheapest it's ever going to be.

[00:33:43]Right. Let's talk about the Codex app now. Uh so you have been using this quite a lot and you're quite excited before this show. You were sort of telling me some of the features that it can do. I haven't used it. So for anyone else that hasn't used it, just briefly explain what this app can do and why should someone use it. I mean, initially it was a coding only app. But now it's becoming kind of like a mix of Claude Cowork and Claude Code, basically. Like think about this if you're in the Claude ecosystem, that's how it works. And so I can actually show you what it looks like and what we're talking. And so the point is you have your projects on the left. It's very similar to the Claude desktop app in a way, but more polished, less bugs, which is nice. And so you see you have your folders that you're working. You can also do normal chats. So you can just like, you know, it's like ChatGPT basically. You can just Yeah, it looks exactly like ChatGPT >> Yeah, it's fine.

[00:34:37]Um you can do automations. I have an automation that runs every day at 6:00 a.m. to send a quick message. Let's say just respond hi, nothing else. And just a way to start my limits.

[00:34:47]And so every morning at 6:00 a.m. it just kind of like starts it. It just gives me two windows in the morning instead of one, basically. Yeah, can you just explain that concept for anyone that doesn't understand how those windows work?

[00:34:58]>> Yeah, so you see for example, like I mean, just visited. But the point is like you see you have your rate limits.

[00:35:03]You have a 5-hour rate limit. And it's like how much AI you can use in the next 5 hours, right? And then weekly, how much you can use in that week. But the point is like if you run out to zero in the 5-hour window, then that's it. You have to wait until the next window to be able to use it. So let's say you start work at 9:00 a.m., right?

[00:35:21]It means that there's a five it starts from your first message. So it means there's a 5-hour window between 9:00 a.m. and you know, whenever that is, 1:00 p.m. or 2:00 p.m., where you only get one of these windows. However, if your first message is sent at 6:00 a.m.

[00:35:34]and you start work at 9:00 a.m., that means the window ends 2 to 3 hours later, which means there's another window starting in the middle of your morning, which means you essentially get twice the usage during that period of time and you don't have to, you know, walk another session for it. So the idea is to make the bot start a session several hours before you're going to start using it. So the first session is much shorter and you get to use two windows instead of one, basically.

[00:36:00]Okay, and all you have to do is just set up that automation and then leave your computer on. Pretty much.

[00:36:06]Um and so it just sends a message. It just says hi and I do it with like the cheapest model. You see I'm using GPT 4 5.4 mini, which is the cheapest model.

[00:36:13]And and that's it.

[00:36:14]>> a really good hack. Yeah.

[00:36:16]>> simple. Well, that shows you what automations are. Uh you can install plugins and MCPs in a very cool way. You can see they actually have a pretty good interface. I know Anthropic has it in their app as well, but this is actually nicer. I think this works quite well.

[00:36:27]And so they have skills. It's worth pointing out as well that if you haven't looked in the plugins or the the MCPs, even in like stock ChatGPT, let alone some Codex or Claude, like there's a lot more now than there there used to be. I remember 6 months ago there was like seven or eight in here, but there's hundreds now.

[00:36:43]>> quite a bit. So you can connect your apps pretty easily. You can manage skills as well here.

[00:36:48]Uh and you can also create skills. Like you can it's very easy to do that. And the thing that I think is interesting with this app is like, okay, it's a code editor, right? So like I had it research podcast stories, for example, and interact with Notion and make the page we just looked at. But also, for example, I was working on my presentation. It's able to make PowerPoint files, but when you click on them, you actually get and let me make this a little bit smaller so that you guys can see, but you get a full presentation reviewer. The presentation, as I said earlier, wasn't very good. I don't think the slides are very good. So I would not recommend it. Although it did prompt these images that are pretty good. But the idea is it can open spreadsheets as well and you also have a full web browser here. So if I type Authority Hacker, I can use it, but most importantly, I can start prompting on the left and it will start using the website for me if I want to. And so the idea is this is becoming an app for creating documents, reviewing them, and giving it work, basically.

[00:37:43]Uh and so it's quite interesting. It feels really nice. Like that Yeah, >> [laughter] >> it's hard to articulate. If If you're If you're listening on the audio version, it just it feels a lot less intimidating even than like the Claude code app in the Claude code section in the Claude app does.

[00:37:58]It's hard to put my finger on, but just something about it feels quite >> I can make the browser full screen if I want to. So if I want to start using the browser and I can prompt here, for example. See? Does this browser have all the same functionality as Atlas, OpenAI's dedicated browser? It doesn't have extensions yet, but the rumor is they're making it a full browser. So like you will literally have a full Chromium browser built in. And the AI can take it over and use it, right? So I can be like, sign up Let me make I'll do it here. Sign up for the newsletter uh on the opened site, for example. And in principle, it just you start having a cursor. It will just take over and it will start operating.

[00:38:37]>> Does it take over your mouse and click?

[00:38:39]Like you can't use your mouse while it does it? Or does it do Does it just uh do behind the scenes?

[00:38:44]>> Uh you you will see it moving if I let it work. I didn't give it an email. Can you fight against the mouse movement though?

[00:38:49]>> Uh it doesn't take over your mouse. It has its own mouse. Like you have your mouse.

[00:38:52]>> All right, perfect. And then it will just do it. So it's asking me which email. Um but the idea is like, yeah, it's like then it's going to start operating. So you could imagine logging even in like a spreadsheet, for example, and have it start operating. And see now it's going to start taking it over and use the browser, basically. And you can do this on sites that like you need to log in with like a username and password.

[00:39:13]>> because you can just log in and do it.

[00:39:14]Log in.

[00:39:15]But Yeah, yeah. So just to be clear, you're not typing in your username and password and giving that to ChatGPT.

[00:39:21]You're just in literally in the browser, which is in the same app, and you're just typing it in in the browser. So that um I supposedly should be a more secure way of doing that.

[00:39:31]>> Yeah, I actually did not activate it on this chat. So you won't be able to see it. I'm sorry. But the point is you can do that and you can have it I had it do this.

[00:39:38]Uh but I can see it's not activated. I just took the thread of my presentation.

[00:39:42]Is this a trust me bro moment?

[00:39:43]>> Trust me bro, it does it.

[00:39:45]Uh it's like I believe you. Uh but the point is like yeah, this app is very nice and then you still I'm stopping it.

[00:39:49]But you also have your files on the right. Uh so you can see the change files, but you can also see all your files. You can just um like you know, if I take a a run for the meta campaign for example, like I see the assets. Like I can open the images that were in there.

[00:40:03]I can open the research with the brief.

[00:40:06]Uh and then I can highlight things and I can add it to the chat and drop a comment etc. So the app is actually decent to work with basically. And then you have all these tabs all of the stuff I open, right? I can switch between the tabs here between my presentation, the website.

[00:40:19]>> onto something here. I tell you like in terms of like structuring how this is laid out. You got your files on your right, your browser with all your tabs in the middle taking what 60% of the screen and then your chat at the left.

[00:40:29]Like for something about that feels much better. Uh much more natural. Especially like we have these like super wide monitors. I can see myself using more of this real estate in the screen than right now with VS Code. Like it's I actually use quite small amount of space. Um I I don't know why. Yeah, you could use more. And the thing is if you want to jump in VS Code or whatever code editor, you click this button here and that's just going to open the exact same view on VS Code including the open files.

[00:40:56]So you can jump like if you feel like oh I need some more control, you press this button and then you jump into the code editor for the same file. So it's like that's what I'm saying. It's a good app now and they're adding more and more features. It can control your computer.

[00:41:08]It can do a lot of stuff that the Claude app can do. Has a browser, the browser is pending. And so you can see how it's moving from just coding to almost knowledge work here. Feels like there's a bit of an arms race in this specific sector cuz you know Anthropic's also really made some nice updates to the Claude app recently specifically for like uh using Claude Code.

[00:41:28]>> But the Open AI app is better. Yeah, okay. Uh it's like Um but it's one of those things I think like they're just both going to copy each other's features somewhere like >> months tomorrow and then they'll probably be quite close. I give it I say give it 6 weeks and they'll be there.

[00:41:41]>> be quite close. So it's like but it's still interesting because I think this format is where things are going. And it's like you know, we started on VS Code, but this almost feels like the evolution of VS Code. Um where you get all the good stuff.

[00:41:52]>> having the file system there. Like that's you know, cuz you didn't used to get that in the Claude app when they when they released it. It's just not as good.

[00:41:58]>> it's just not as good. It's just like it's not if you click on an image for example, like it will not open that image. It will kind of show you the binary string for example. There's a lot of bugs. Uh whereas this is sleek and it works. And you see why I kind of like working on this app for some of my work actually. Um so yeah, that's the Claude app for you.

[00:42:17]All right, let's talk now about Opus 4.6 versus Opus 4.7 because I've been reading a lot lately that people are kind of hating on 4.7 and they're moving back to Opus 4.6, which is an earlier generation model. And this was kind of reinforced just this week because Anthropic now have like a a model selector where you can select 4.6.

[00:42:42]Again, you could not before. Yeah, again. Yeah, so when they had 4.7, it was like that was your only choice. Have they done that in response to this criticism? Why is this criticism valid?

[00:42:53]And what's your take on it? Uh okay.

[00:42:56]First of all, 4.7 prompts very different from 4.6. I had to update a lot of skills for example for us to work better with 4.7.

[00:43:04]And most importantly, it eats more tokens. It's like 1.35x more tokens for input tokens with the new tokenizer etc. So people are often getting worse results and they're getting lower limits as a result of the new tokenizer. And so that creates frustration for people.

[00:43:20]Let's focus on the worst results first.

[00:43:23]So is it the fact that they're using the same prompts, same skills that they use on 4.6 on 4.7 and they haven't adapted it? And to adapt it, there's like a transition guide or something that they published. Is that right? Uh yeah, they did do that. I think it depends on the task again. Like if you're doing a creative task for example, I think 4.6 might be better writing for example. Uh so if you want like pure writing, it's not very far. Like you know, we're talking like a 10% difference or 15% difference. But it like there's more nuance and texture on Opus 4.6 copy than there is on Opus 4.7. But Opus 4.7 is a little bit more diligent than Opus 4.6.

[00:44:00]Therefore Opus 4.6 can be a bit more sloppy on you know, specificity, higher detail tasks. And so Okay. So you might want to reach out for different models for different tasks, which again is a bit annoying. Yeah. What's your take in terms of like the community response? Is most of this it's not up to the task or is most of this it's using too many tokens and people are hitting their limits so they want something a bit more >> Well, I mean you hit your you started hitting your limits, right? Recently. So >> Yeah, I hit my limit for the very first time last week. That felt like a bit of a badge of honor to be honest.

[00:44:31][laughter] So I'm I'm only on the $100 plan though.

[00:44:34]But like the fact that it did not happen before. Do you feel like you have more usage or do you just feel you used more tokens?

[00:44:41]Like I I mean I'll be honest with you. I don't even monitor my usage.

[00:44:44]>> do you feel like you used it more? Do you feel like you Do you feel like oh I'm doing more things now with Claude than I did before and that's why I hit my limits. Or do you feel like oh it just No, no, no. That was a really busy day and I was doing like a lot of like parallel meta ad generation at the same time. Okay, so it's like it was still a high I mean it's yeah, it does use more limits. It's like even on the $200 plans like at the beginning it was hard to get over like 50% of a session limit. You get four times the session limit of the $100 plan than the $200 plan.

[00:45:13]>> [gasps] >> And it's like now when I'm turning like two or three threads at once, I will go 75 80% of that session in a session and that would never happen on 4.6. So it does use quite a bit more token. And then if you couple that with the fact that the output can be seen as worse especially if you don't re-prompt it. If you re-prompt it, gets closer.

[00:45:31]Especially for knowledge workers, then I can see why people are switching back.

[00:45:35]One thing that is really frustrating though is Anthropic put Opus 4.6 back on the desktop app for Claude Code. However, they put the 200,000 token version back. So if you want 1 million token, you're still kind of stuck on Opus 4.7 actually. Is that Is Is there a financial reason why they might have done that perhaps They need different server clusters, right?

[00:45:55]They're different models. So it's like they kind of like oh we have 100 servers like you know, 50 are serving 4.7 and we have like 10 for Opus 4.6 for people who still want to use it, but they cannot take another cluster of 10. So it's a hardware limitation. Yeah, this is This is every Anthropic problem right now.

[00:46:11]Every Anthropic problem is like they're really sloppy updates and they don't have enough compute. It's like once you solve these two, like apparently they wanted to release the Mythos model to more organizations this week and the White House blocked them because they're afraid that they themselves won't have enough access and Anthropic is running out of compute. So it's like it's happening right now.

[00:46:31]Apparently they were banned and then now they just can't even release their model anymore. So yeah.

[00:46:35]Interesting. Okay, let's talk now though about the Meta Ads AI connectors because I've been reading a lot this year that people have been using sort of like third-party integrations, MCPs to specifically to upload their ads to Meta cuz it's tedious.

[00:46:52]>> Look, if you are running Meta Ads yourself, like you don't do this through an agency, you know what I'm talking about here. Like it's really annoying. It takes forever to like upload an image and some text and they make you click a thousand times. It's just a little bit laggy and doesn't load properly and then like forget something and then they try and like upsell you all these hey translate my ad automatically with AI to Spanish and generates a bunch of shitty images that you don't want to use. Like it's just it's a bad experience, but they have a complete monopoly on this. So they're not really forced to make their like interface good. And this is actually a reason why a lot of people use agencies cuz you just can't be bothered yeah.

[00:47:32]Like dealing with this stuff. But I think it also costs Meta money because people aren't uploading their ads. So aren't uploading and scaling more ads and testing more things. I mean so I was really interested to hear that they are releasing two connectors themselves. So there's an MCP server and there's a CLI.

[00:47:53]Um so I would first like you Gael to explain the difference between an MCP and a CLI in this context for people that might not be aware of those things.

[00:48:01]>> Okay, MCP model context protocol, which means it is a protocol made for AI agents only, not for people.

[00:48:09]And it basically allows them to interact with the tool. It exposes the API of that tool so that they can interact with it programmatically. CLI, command line interface, is a terminal tool, which means you type a line in the terminal and it does something. And so you can do it yourself or an agent can do it. And so >> When you say does something, you mean like it does that exactly >> exact same thing. It's the API as well.

[00:48:31]There's no interpretation.

[00:48:32]>> Yeah, it's the API as well. But the point is CLIs are actually now kind of like more trendy than MCPs for connecting your tools. The reason why is because in MCPs, you need to load all the tool definition inside context. You can defer it, you can do all of that, but there's still a bit of context eaten. Whereas CLI, you can load nothing. And then what you can say is like whenever it needs to interact let's say with the Meta CLI, there's actually a command that it can type in the terminal cuz your Claude Code has a terminal, right? So it types that command and that just returns all the commands it can use and boom, it's just loaded in the memory. And so that is better. Like some of that I'm not going to lie still went over my head.

[00:49:08][laughter] Like I'm going to throw an analogy here. Tell me if this is or not. So a CLI is you give it a precise input and you get a precise output. So you're like it's like when you drive a car, you turn the steering wheel left, the wheels turn, you go left. Like it's very predictable what's going to happen.

[00:49:27]With an MCP, it's like you're sitting in the back of a a taxi say and you say hey turn to the left and the driver is going to turn to the left probably, maybe, usually. Um but like how far and exactly where, like you know, there's some interpretation there depending on what you say.

[00:49:45]>> Is that No, I mean it's like creative if you are using the terminal. So if you are using the terminal, it would do that. But I am not assuming you're going to use the terminal. I'm assuming the agent is going to use it. So it's kind of essentially the same thing. You're still asking the agent to operate the terminal. Um and so okay.

[00:50:00]>> I see what you mean. So it's like yes, if you operate the terminal, if you type in the terminal, yes, this is a good analogy. If you do not operate the terminal, it's the same thing. Uh So in that case, because you have that AI layer um that you're always interacting with, it almost doesn't matter which you use then. So why would anyone use an MCP in that case? Because it's for example, there's these nice menus in these apps that allow you to connect MCPs but not CLIs, for example. So people who need an interface uh is just easier. That's it. That's the answer.

[00:50:32]So as a Claude Code user, if I have the choice, I should always go for the CLI over the MCP. If you have a choice, yes, it's better. So for example, like um Google mostly operates through CLI. So if you want to connect to Google Drive, there's called a GWS uh CLI and then it's just like way more efficient because again, imagine all the commands you could have across all Google products for Google Sheets, Google Docs, uh Google Drive, etc. That would be like so many tools in the MCP. But through the CLI, it's much nicer because it can say, "Hey, I want to interact with a sheet. What are my commands?" It would just get that CLI command. It just gets the commands for the sheet. It doesn't load everything into context. And so it's just yeah. I know it's going to be complicated. I can see you smile. The point is it's more efficient. I was actually just smiling because the way you said sheets sounded like shits. So >> [laughter] >> okay, fine. But there you go. Fine.

[00:51:21]Okay. Completely So I I I mean all all all that's to say though that Meta has released their own version of this. So now if if your ad account has access to it with and I believe they're rolling out in stages. So I don't know um whether it's high spenders or low spenders or who's getting that first. But you will be able to have Claude Code or Codex or any AI tool actually upload your campaigns, duplicate ads. Yeah. Um you can interact with it from like a a knowledge perspective like, "Hey, which of my ads are performing well or badly today?

[00:51:52]Okay, you know, let's pause this one.

[00:51:53]Let's do this." And you know, you can make um essentially applications within your AI tool uh that will do all sorts of monkey stuff. And it can now fully take the action.

[00:52:03]>> actually reading right now that the MCP is mostly for pulling data and the CLI is mostly if you want to like post stuff, etc. So you're kind of be pushed to the CLI if you want the more powerful version.

[00:52:14]Okay, perfect. Um so I mean, this is something like this is a huge pain point for me right now. We're going to do a lot more ads, I think. uploading these.

[00:52:21]So yeah, yeah, I think Meta, you you know, you're going to make a lot more money out of us uh hopefully from releasing this. Yeah, I think so. I think this plus the ability for Claude to pull the data and analyze it again, you can use your Codex skill for that.

[00:52:35]Uh if you want to do that. And I'm working on some seed and stuff right now so we can generate videos as well. So that's going to come as well. And this is all particularly relevant on Meta because there's all these stories of people getting banned this year. I On Meta ads, rumors are like a little bit of uh you know, conspiratorial sometimes, but you know, I've heard multiple sources and seen many posts on Reddit of people who are using third-party MCPs or integrations to upload their ads getting banned, you know, as soon as they started doing that. It could be a coincidence, but something you need to think about. Yeah, so finally we can manage ads with agents and that also means probably more competition. And then it's I'm wondering if that means this is the end of like uh human-managed ad accounts because you could literally have agents optimize 24/7. It's like does that mean that you have to manage with AI right now?

[00:53:26]Because >> [laughter] >> like think about it. It feels like there will be a gold rush like starting yesterday where people who are taking advantage of this will like gain significant advantage and then it will just be like everyone's doing it so you kind of have to do it. But then eventually I think Meta will will look at what what you're doing and then they will incorporate some of those things into their own system. So eventually they should like kind of automatically uh do a lot of the stuff. But whenever Meta's doing it, they're optimizing for their profit, not yours. Which is why you might trust your agent better. Uh and I think that also means that a lot of like all these software that you used to pay for, they're kind of uh in trouble right now because uh you'll be able to just have an agent scale and then it will do most of what you used to pay a lot of money for.

[00:54:12]Yep. All right, so let's talk now. Uh our last story is about DeepSeek, who we haven't really heard much of for a while, but they've been busy. So for context, it's a Chinese company or model? Um DeepSeek like who who owns it?

[00:54:27]It's That's the name of the company.

[00:54:28]China.

[00:54:29]>> [laughter] >> China. China. Yeah. Yeah.

[00:54:33]China owns it. It's funny because like whenever you you hear about DeepSeek on social media, especially on like uh a short-form content, it always starts off with someone's like, "China just released this new model." As if it's some massive massive like um relevant factor for it, but there you go. So anyway, they have a a new model.

[00:54:54]Is it V4? Yeah. V4 Flash and V4 Pro. So they took a page from uh Google's book here because it's the same names for the models. And this is relevant because it's pretty good. It's not Frontier good. It's pretty good, but it's super cheap. Um so this was the thing that if you remember 15 months ago, uh the stock market took like a 10% hit or something like that because this came out and then suddenly Anthropic, OpenAI, uh Google's like um relevance or dominance was threatened because they were able to make a Yeah. very cheap model that was very good.

[00:55:29]>> some benchmarks here. We're not going to go through the benchmarks very much, but you get the idea. The blue bar is the DeepSeek model and then this is Opus.

[00:55:36]This is GPT-5.4 high and this is Gemini 3.1 Pro high. And so you can see it's it's not winning most benchmarks, but it's competitive. And it's like usually they pick the ones where they win, right? But the point is the API price is so low. It's like input tokens is $1.74 and output And output is 3.48.

[00:55:57]>> Basically, this is around Gemini 3 Flash prices. For It's more for the input, but about the same for the output. But it's like Gemini 3.1 Pro level of intelligence almost. And so like that's the thing is like if you're using something like an A10 or you need an API or you need to mass process data that you could not do through an AI subscription, this is a very, very interesting model because you can do things for fraction of the price. And you get 1 million context window as well now. So are you using it in any of your workflows or systems or is it not really relevant when you're kind of doing everything in Claude Code? So right now, not so much because we're not, you know, we're not processing like 100,000 items, for example. That would be the case. Like someone's like, "I need to process 100,000 contacts and have some research done on them, for example."

[00:56:44]Well, the API bill for 100,000 API calls between an Anthropic model and something like this is like more than 100x probably difference. And so like it's significant. We don't really do massive scale stuff right now, so it's not really relevant. But if that happens, I would probably try these models for sure. Nice. Okay, great. Any other final words of wisdom then before we wrap up this episode? I like Codex, but that's not a reason for you to switch. I do think for most business owners/marketers, Opus is better. I showed you the presentation I did wasn't very good. We showed you even the Meta ads. I think Opus did a better job. So if that's the kind of test that you're using AI for, um I would say stick to Claude for now.

[00:57:24]And if you do more technical things, then give a shot at Codex. I think it's very good. I like how thorough the model is. So you don't have to pick a side.

[00:57:32]Just use the best model for each time.

[00:57:33]You could have a Claude subscription and a $20 subscription to ChatGPT to get a little bit of Codex if you want. And and that's the best of both worlds. Yeah, so don't get distracted. Uh and I realize the irony of this statement saying with uh you know, we just watched an hour podcast.

[00:57:47]>> We just watched an hour podcast about the ins and outs of this, but uh don't get distracted unless you really think it's going to make a difference to your business. Last thing I'll say is authorityhacker.com/aiaccelerator, those monthly plans are available until Monday, the 4th of May. So please go there uh and sign up and we'd love to see you inside our community. We've got over 600 members in there.

[00:58:11]Uh lots of interesting discussions from other business owners about how they're using AI, lots of courses, lots of skills uh which Gael's releasing weekly at the moment, including all of those that you've seen there. Can you give us a sneak peek as a what what's coming up um in May? There is I I'll tell you one.

[00:58:27]I won't tell you all of them. I have a few that are on the bench right now, but one that is really cool is the CRO one.

[00:58:33]So it connects to your Google Analytics.

[00:58:35]And if you have it installed to Microsoft Clarity, which is like a free heat map tool, I you can give it a landing page and it's going to analyze how users are interacting with it and see where people are stopping to scroll or where they are like rage clicking.

[00:58:46]It's called, you know, they click on things it doesn't work, etc. And it's going to make you a list of suggestions of tweaking that page based on that data. And then it will tweak that page for you on your website and just do it, basically. So it doesn't give you the report. It gives you the report, but then it actually changes that your site for you as well.

[00:59:03]>> Yep. So you could just have this run automatically um like once a week or whatever and it will just It keeps a log of what it did.

[00:59:11]So it's like it will understand what happened. You can then go and query the data before, after, etc. And the next time you run it, it will just be aware of what it did last time and what changed. And then it will make better decisions for the next round of changes, basically. I'm genuinely surprised. I mean, maybe this exists already, but has someone not built like a service or a tool or or something to just like, you know, connect it and there's something like this happening in the background and just auto CROs and you you know, like makes anyone that does >> done the auto on-page SEO, right?

[00:59:41]Released it this week. Um like it does that for on-page SEO. It connects to Ahrefs, it connects to GSC. It looks at where you're ranking, it looks at your competitors, it does gap analysis, it makes assumptions, it takes notes, and then 2 weeks later it might just pull the fresh data and be like, "Hey, did my changes improve the rankings or not?"

[00:59:58]And if it did, it's like, "Great, let's double down." Or maybe like what's missing still or whatever. And if it didn't work, it's like, "Ah, went the wrong direction. Let's try something completely different this time." And makes an edit and then does that, basically. So yeah, I'm building these side skills right now that do marketing work for you. It's really mind-blowing like where this is going.

[01:00:15]Uh I'm quite excited to see where this is in 6 weeks, let alone 6 months. But uh yeah, 30acorn.com/ai-accelerator and um yeah, we hope to see some of you in there. Um that's all for this week.

[01:00:28]Thanks for watching. We release new episodes of this show every single week.

[01:00:32]So if you're watching on YouTube, please uh subscribe so you don't miss that. It also really helps us out just getting those subscribe numbers up and drop us a comment as well with your thoughts on this episode. We look through all those and we'll do our best to answer them.

[01:00:46]But uh yeah, see you next week.