Your AI Bill Doubles Next Month

Everyone's drunk on AI usage right now. You've got agents, automations, workflows on top of workflows. But at the end of this month, the somewhat hidden double usage promos on Claude and Codex go away. That means you're going to hit your limits faster and your bills will go up. So today we're going to explore what options you have to stop this from impacting your business. We'll also get into the new million token context windows on Opus and Sonnet models and, and OpenAI's new mini and Nano models are here. They're cheaper, faster and not that far behind GPT 5.4. OpenAI is also pivoting hard towards business users. We have a report from a new internal meeting there and we'll talk about what that means for you, your business and everyone else using ChatGPT. Ahrefs also dropped a free web monitoring tool and opened their API to nearly everybody. So we'll look at what this means for anyone doing SEO with AI. My name is Mark Webster. I'm joined by Gail Breton, co founder, authority hacker, where we test all this AI stuff and share what business owners and knowledge workers actually need to know. How's it going Gail?

0:00

Pretty good. I've been shooting this new course for Accelerator on Vibe coding websites and it got bigger than I expected. But I think it's very good. I'm very happy with the result. I think I'll talk more about it next week when it's actually released. But yeah, it's good. I think it's going to. I've made this whole rant on like don't use WordPress. So this is kind of like the, the backed up version of that and how like it's not that bad to use websites that use code anymore. So like that's, that's it been pretty good.

1:05

Some pretty spectacular results with the examples you've shown me. You redesigned my wife's business website and it's pretty amazing. So we'll share more on that next week. This week as we record this, this is the 18th of March, probably going out on the 19th. I want to talk about the usage crunch that's coming up because I, I didn't actually know this but I'm. And everyone else who uses Claude and Codex is currently benefiting from some double usage promos which are activated behind the scenes. You don't need to do anything. So if you're using Claude, if you're using Codex, you benefit from this. On Codex I believe it is all the time. So you basically for the month of March you're getting double usage. Double.

1:35

It was two months, it was like February and March actually. So it's been, it's been a while now. That's what I was saying. Like people are used to it and a lot of people just move to codecs and so that's kind of like what feels normal to them. And GPT 5.4 is also using a lot more limits and so it's going to sting a little bit soon.

2:19

And Claude have this set up where they basically have defined peak hours as 8am to 2pm Eastern Time. And any usage outside of that you get double limits basically. So you can use twice as much before you hit those annoying limits that's going away as well at the end of the month. So I think like first of all a lot of people don't know that these promos exist. They think like the current usage is just what it normally is. So are people in for a rude awakening in a few weeks time?

2:35

Yeah, I think that's what's going to happen. But it's interesting as well as Claude launched this one later, like it didn't do two months like OpenAI. It's only like two or three weeks, like they launched it a few days ago. So it's like for Claude users it won't be too bad, it would just go back to how it was like last week for example. But for the people who use codecs, I think it's going to. I'm almost 100% sure they're releasing new plans for it at the end of that promo. So the idea is the people who are on the $20 plan now will probably get a little bit more of the same usage. On the remote upcoming $100 per month plan from Codex, you probably get maybe 40%, 30% more limits, but you need to pay five times more money to get the same thing. And then Claude cloud is okay. They changed the reasoning default on Opus from high to medium to reduce the token usage because Opus 4.6 high was like very token heavy. So fuse cloud code you might have.

3:07

Do you want to just explain what the differences are between medium high and. I think ChatGPT has extra high, extra high.

4:04

I mean it's just how many tokens the model can use to think before it answers the question. And you can set a hard limit. Like these levels are basically a hard limit number. It says you have this budget of thinking before you have to give your answer. And so that's what comes comes back. You get a better answer.

4:10

It takes Longer to get an answer, but you get a better answer pretty much.

4:26

But there is kind of like diminishing returns. So it's like, for example, for codecs, going from high to extra high is actually a pretty small increase and sometimes even a decrease. Sometimes the model overthinks and talks itself out of the right solution. And so it's not always good. But going too high, for example, it's pretty good. It makes a big difference. So Claude has been playing with that. They kind of defaulted back to medium so that it doesn't use as much and so on. But the idea is, yeah, you're just gonna end up having to pay more for the same as what you're getting right now, you know, in a few days.

4:30

And where or how can people control those medium high extra high options?

5:04

So if you are in cloud code, you can just type effort and you can type medium high, low. And then there's a max now that they add, which is like unlimited thinking, if you want. And then on Codex, I think it's when you select the, like if you're in the cli, I think it's when you select the model and then if you use tool, they just have a dropdown that you just select high, extra high, medium, et cetera. And so, yeah, but the point is like, they use your limits faster as well. And so if you're like on Opus High, for example, it goes pretty fast if you stay on that.

5:11

And it's interesting as well, because Opus 4.6 and Sonnet 4.6, they've recently released, they were trialing it before to everyone, the 1 million token context version of this. So if you're using Office 4.6, as most people will be if you're using Claude code, it now has a much larger context window. I thought this might be a good idea to actually explain a little bit about what context is, why it matters, and how it can actually affect performance, and why these kind of numbers aren't always what they seem when they get thrown out there. So I'm going to, I'm going to try something, I'm going to try and explain context to you and I want, I want you to correct me if I'm wrong, because I, I, I'm not super clear about this, but I want to see how I do here. So essentially, all models work on tokens rather than words. And for want of a better analogy, a token is roughly 3/4 of a word. It's just a piece of text, it's a few letters. All models have context sizes. Which let you input or let them output more text, essentially more tokens. And that lets you have longer conversations, it lets you create longer pieces of output, of code, of text, whatever. It lets you put in more input tokens in there. That's useful. If we go back to early days, models had really, really tiny context windows and you couldn't really do too much with it. You couldn't make long pieces of text, even let alone code. So that's changed. But the problem is, as you feed it more context, the models get worse at retrieving specific information that's buried deep in that. So if you're just throwing lots and lots of documents as the input and say, hey, figure out whatever this, this code base means or how to solve this problem, it's not going to know where to look. So you're better off giving it more specific information that's relevant to it rather than lots and lots of information, just dumping it in there. But feels like a bit of a trap because with a 1 million token context window, you can, if you want, dump more stuff in there. Am I somewhere in the right ballpark

5:40

about what happens there is, I would say so. Seven out of ten.

7:53

Okay, give me the ten.

7:56

So the first thing is like you mixed output tokens and input tokens. Output tokens have their own limit. Like each model can output a certain number of tokens per response. And it's like it's irrelevant of the context window. The context window is what the model remembers, right? So AI model are stateless by default. When you make your first AI call, it knows nothing about anything and you just get access to that raw intelligence. But it really has only its training in memory, nothing else. And so the context is everything else that you put. That's your prompt, that's the system prompt, that's documents you might attach, that's anything like that. And that's also the chat history. Because again, the way the chatbot works is an illusion. It's like you think that you're having a discussion with them all, but you're not. Actually what's happening is when you send a follow up message after your first message, your chat history is fed into the prompt which allows the model to write the next answer. But in reality you're calling another stateless API call. That needs to take that into consideration. So when you make a long chat, you end up just feeding a very long context window inside the model. So essentially the context window allows you to have longer chats. That's pretty much what it allows you to do. And to attach more documents and more things into that. The only problem is, as you increase the context of a model, the quality degrades. It's kind of like your memory, right? It's like if I tell you to remember three things, you're likely to remember them all. If I tell you to remember 50 things, it's going to be a bit more fuzzy in your head and you will have seen them and you will have some memory of it. But because I've asked you to remember so many things, you kind of got overloaded and then you were unable to remember exactly the 50 things I showed you. So the models work exactly the same, which is kind of interesting when you think about it, that it's so similar to the way our biology works. And so that's kind of the challenge with one main context. It's nice in a way that when you add 200,000 contexts, quite often you would end up having the model compact halfway through the conversation. And compacting would be taking the whole chat history and summarising it and adding that instead of your full chat history, which creates a more fuzzy memory for the model and a less good experience. Now, the problem that you're facing is that, well, when you get to like, 4, 500,000 tokens, the model is going to be a little bit more fuzzy, especially about the earlier parts of your chat. The same way as the first item I would have showed you when I. With 50, you would probably remember that

7:59

there's a recency bias with this.

10:18

There is a recency bias. And so this 1 million context window means you manage your chat a little bit differently. And it also means that while you can go to 1 million, you probably should try to avoid going to 1 million. It's just a way to avoid having this kind of like you're halfway through something and the model is running out of context. You're still better off keeping going than compacting like you used to before. But you don't want to kind of keep the same thread going if you have the opportunity to start a new one, or even, as I tell to people, you can manually compact in cloud code. So you can do compact. And then at the natural post point of a task, for example, I've made a V1 of a new skill, I need to audit it, or I need to run it for the first time, et cetera. I don't really need the full context of everything that I had to build it, but it's nice to have a bit of a summary of where we're at instead of Starting a new chat, well, compacting manually on cloud code starts making sense. You had 300k tokens, you go down to 50k or something and the model performance becomes better because it's less fuzzy on your chats basically.

10:20

Is there a risk though, with increasing the context size this much that people get lazy and they start reusing chat windows and just never starting new ones? And it creates bad practices there, it

11:22

creates opportunity for good users. I think that's the way I see it. And the thing as well is you can see that they have some benchmarks. So there's these benchmarks on long context retrieval. So let's say I feed the model like an entire book and I ask it like a specific detail and then it's like we'll see how good the quality of the detail is when it surfaces it back. And you can see that at 1 million actually Opus is far ahead of the competition. Even GPT5.4 that just released that is also 1 million context. I mean Opus outperforms it at 256k context, which is like, you know, a lot of people will get to that point. It's basically the max context window you had before, but it stays pretty good at 78.3%. At 1 million now it still drops and they all drop apart from solid 4.5 for some reason that seems to get better.

11:36

Is there a chance that seeing as this is published on Claude's website, that they've maybe cherry picked a specific benchmark

12:23

that shows up favorably a thousand percent? They all do that and it's like all this marketing material is often like Apple marketing material or they would just present the best thing and then they would just put like a very small asterisk at the end of the page with all the edge cases, et cetera. It happens sometimes that entropic benchmarks against OpenAI and then OpenAI outperforms them. For example, there was a benchmark where they said, ah, but the OpenAI model didn't understand the instructions, therefore we removed it. And then when the cloud model started to kind of not behave because it understood it was tested, they surfaced it as like, ah, the model is so smart. I understood it was tested and essentially the OpenAI model did it before and understood it was tested before. But Entropic dismisses it and says it's a mistake from the model so we don't report the results. And when Claude does it, it's like emergent behavior that shows how smart the model. So yeah, be a bit careful, take these things with a grain of Salt. But Opus is still probably a pretty good model at long context, just the quality drops. And anecdotally, a lot of people I've talked to since the 1 million has dropped report a drop in quality, probably because they extend their context window further and then as a result it's not as good, basically.

12:29

Okay, so that's an interesting potential takeaway. If you are perceiving a drop in quality, maybe you're using more of your context window and therefore you're further along that x axis of that graph.

13:43

Yeah, I mean, potentially, or what many people believe as well is that these companies quantize the model. So they make a slightly cheaper version to run, but slightly dumber version. And a version bump to 1 million is essentially a new version of the model. So that would be an opportunity for them to quantize the model and make it a little bit lower quality as well.

13:55

So you mean we've kind of gone to like 4.5, 5.5 instead of 4.6, so it's worse. But higher context window, they have a

14:15

compute crunch, you can see entropic. They're even going down quite often these days. And so they try to find ways to run the models more efficiently. And it's kind of like you're selling soup at a restaurant and you just add a pitcher of water in your bowl of soup.

14:24

Is this just insertification of AI then?

14:38

Well, it's not proven and AI company will never confirm it. But go on X or on social media and people who are heavy users and you will see that there's always chatter about the model got worse. So there's kind of like this whole new period when they release it when it's amazing for three weeks, and then they start to kind of slowly just add a glass of water into the bowl of soup, and then it just kind of slowly gets worse. And when a new model comes up, it feels amazing again because not only do you get a jump in performance, but you also get, you know, all this, it's not diluted yet, therefore it's like it feels very good. So there's these kind of periods in AI that you can identify, which is the first month of a model, of a new model always feels great, and then it just kind of feels worse. Now there is an argument to say as well that we get used to the performance. And so therefore the perceived quality feels like it drops and it doesn't feel as amazing anymore.

14:41

This could be a kind of like conspiracy theory.

15:32

It could be, it could be, it could be. But like you Know, quantizing and distilling models is a very real thing. Like Google for example for Gemini 3 Flash just said we distilled Gemini 3 Pro. Like it's like this is how we do this. And it's like we got 80%.

15:35

We talked about that last week on the show on distilling models and that actually kind of brings us on to our next topic. So GPT 5.4, Mini and Nano are two new models which have been released. So GPT 5.4 has been around for a few weeks now. Pretty good model, does very well. They have released, I guess these are distilled versions of GPT5.

15:49

That's exactly what they are.

16:11

So Mini is your kind of middle of the road version and Nano is your super fast, super cheap model which in their words should be used for kind of categorizing and you know, more simple tasks like that. But interestingly they also said you can use it for like as a sub agent for simple coding tasks now as well. And they got some benchmarks there. Again, it's benchmarks on their site. So take this with a pinch of salt. But it actually shows that it's really not. The mini model especially is really not that far behind GPT 5.4.

16:13

Yeah, I mean it's going to be used in codecs the same way, like cloak code. For example, let's say you select opus sometimes it's like, oh, let me go read some files. And it starts a subagent that reads files, right? This sub agent is not using OPUS most of the time it's using Haiku, it's using the small model from Entropic, reads files, summarizes them to the main model that then makes a decision. And so that's what this is going to be used for in codecs. So they also announced sub agents on codecs the day before these models released. And that was the idea. It's like you can save a little bit on these usage limits that are going to sting next month by delete. The main model that costs a lot of credits is going to be able to delegate some tasks to the mini models like summarizing, reading big documents, et cetera, so that you don't eat the expensive tokens and you take cheap tokens instead. And so I think it's going to be a good model. It's nice to get some competition because honestly, IQ is not amazing for API. So the only mini model that we've been using always was Flash through the API. Like, you know, if you use it in an A10 or something like this. It's kind of nice to get another model and GPT 5.4 mini I think is going to be a pretty good one. I haven't tested it much, it just released, but I think it's going to be a good competitor. And for example they released this model in Notion agents where you pay per credit when you use them. So it's like if you are building on other platforms like N8N like Notion, etc. It's nice to have a capable model that doesn't cost too much and can run like large volume operations without spending hundreds of dollars on it. So I'm going to test it.

16:45

I think especially for small businesses who you know, don't have massive margins to play around with when you're doing a lot of marketing tasks. I'm thinking like, you know, cold outreach or even content creation, like long form content creation, like you're being efficient with your token usage is like it's so critical. Like you know, it's the difference between spending $10 a day and $200 a day on project that you don't necessarily know is going to return revenue or how long it's going to be before you can make that money back. So yeah, it seems like you know, to get back to the original topic where you know, over usage and people getting used to the gravy train of

18:10

cheap chains as it were.

18:50

If we fast forward a year or two, we don't know how fast, you know, processors or GPUs are going to improve or help things. But it feels like all of these companies are subsidizing, essentially subsidizing the cost of the massively. I mean if you just look at how much you pay for your cloud code subscription versus what those same tokens would cost in the API, thousands of dollars. That's quite a big difference. So you know, and they may not even be making money on the API right now.

18:51

No, they're profitable on the API. Like the cost of producing the tokens is definitely. It's like they make a profit on it. The problem is like there's a huge R and D cost and infrastructure cost to be able to deliver these tokens in the first place. So while the companies may not be profitable like, like it costs them much less to produce the tokens than they charge you for it once everything's in place, still it's cheap and it's like the prices are going to go up and the first spike is next month. That's what we were saying in the first story. It's like you're going to start filling it from next month, it will still be subsidized, it will still be cheaper. But if you are a service user, I don't think you'll be able to use it properly on $20 a month anymore from next month. I just don't think it's going to be a thing.

19:18

All that's to say that as a business owner, model usage and being efficient with this stuff is something you absolutely need to pay attention to. So make sure you subscribe to this podcast because we'll keep you updated on all the latest models and as they come out and as we test them and what's worth paying attention to as

19:56

I've done a lot of things for. When you come in a business, it was quite frequent to have them use Sonet for everything, for example, and it's even summarizing calls, et cetera. And it's quite often I switch to a Gemini Flash or something like that, you could cut the bill by 80% and the result is exactly the same. Like prompt adjustment for these simple tasks is more important than the model, actually.

20:12

Yeah, we've seen it a few times actually. You go into particularly a small business where you have like a founder or CEO who's quite headstrong is like, no, I care about quality. Let's use the best models, like let's do Opus for everything. And then you actually show them the difference for 90% of their tasks and it's like, yeah, let's just cut your cost by 90% here for no meaningful reduction in quality. And it's probably also going to be faster as well.

20:33

For example, in the article generation and in AI accelerator, the writing model by default is like Gemini Flash. I'm not even using an expensive model. But the point is it's so detailed, the structure that comes before and then I use an expensive model before to prepare the article, et cetera. But it's like one core, but then all the other cores I can use much cheaper models. And it's like people can't tell the difference. The prompting matter more actually. And even the mini models now, they're just smarter than the best models a year ago. They're not bad models, they're quite smart. And it's like simple marketing tasks. Most of the time they can handle already.

20:58

Let's talk now about Cowork Dispatch. So this is a new essentially a remote control functionality for your phone to be able to control Anthropic's cowork, similar to how you could use Telegram or WhatsApp to control a Claude bot when that was all the rage a few weeks ago. Seems to have died down and still

21:32

popular, but it's not as big. Yeah, but I actually set it up on mine, on my cowork, actually, and you can see it's pretty easy. So it's like, literally I just had to press a few buttons and now if I just open the session on my phone, I can talk to it, but it's connected to my mcps. It's connected to everything. And it's like I have this little dispatch thing on my phone and if I open it, you can see that, and you can see it read my email. So you can see I am going in Montpellier on March 25th. That's fine. So if you're there, send me a message. But the point is, you can see. I don't know if you can see, but I literally have the same chat on my camera as I have here. If I just say hi, for example, on my phone, I can send it and you can see that now it starts walking on its own and it will just kind of go. It's kind of weird because of one thread. So I guess they're just using Compaction or they're just using opus 1 million and not giving a shit. But the point is that, yeah, it's actually very, very easy to set up. Much easier to set up than remote cloud code, in a way. And Cowork can connect to your file system. It can even open a browser. It can use the Chrome extension of Claude and open a browser and do something on the browser and respond to you on your phone, for example. So it's easy to set up and it works quite well, basically.

21:54

Does your computer need to be on to use it?

23:06

It does. It does need to be on. But there's this optional Mac where if you go in your battery settings, you can say, prevent sleeping. And all you need to do is just keep your laptop plugged. It's not going to ruin your battery because Macs just gets. Get the power directly without transitioning through a battery.

23:09

I'm just thinking, though, like, take that example of you're going on that trip, like when you're in the airport, you probably bring your laptop with you for.

23:24

Yeah, that's the problem. That won't work. Yeah, yeah.

23:32

I mean, I guess you could just open your laptop and connect to WI Fi and use it there. But it's like, yeah, but for example,

23:35

if I go to a hotel, like, I just plug it in the hotel connect it to the WI fi and then I just leave it open and it will work, even if the screen will turn off, et cetera. Then I can be all around in the streets or whatever and I'll just message it and it will just connect to the hotel WI fi and do it for me. So it's like. It's not perfect, but it's okay.

23:41

Are you worried if someone steals your phone while it's unlocked? That could be a lot more problematic.

23:59

By the time the person that steals my phone figures out the setup, I think I will have it disconnected. I'm not super worried. I'm already worried if my phone is unlocked. Like, there's a lot already. I think that's the main thing. It's already connected to my email, et cetera. Think about it. If my phone's unlocked, it already has access to more than Claude has. So is it really important? No, not really. It's just. Yeah, it's fine. And also, if I can control the computer, I can just disconnect it, right? It's like as I go back to my computer, I press a button. It's not connected anymore. So it's not too bad. And I think it's a lot easier than the cloth code version that they have or that you can do on the terminal. So I think this one people should pay attention to it. And you can use cloud code on your computer and switch to Cowork for remote, for example. It's not a big deal in the same folder. So, yeah, pretty good.

24:06

Who should be using Cowork over Claud code?

24:54

The way we help people right now is like a lot of people now. They're like, cloud code is amazing. I want my team to benefit from this. And it's not easy. It's easy as one person, but it's kind of hard to deploy, etc. But. But now Entropic has these enterprise settings where you can upload skills, you can connect mcps, et cetera, and they kind of deploy to different departments. You can even set the access per department in your company. And so the way it works is we have people who use cloud code inside the company. We call them the builders. They build skills, they do all of that and so on. Then they deploy them on the company's dashboard so that you can give access to a skill to a specific department, for example. And then that department uses the same skills within Cowork. And it's like it's all the people who. There's a lot of people who do not want to open a terminal that don't want to deal with the technical stuff, etc.

24:58

Or you might not want them editing those, just essentially reading those instead.

25:43

But it's quite nice because they can still fork the skills. So let's say you have a company skill to do support, for example, let's say they have an opinion or they want to personalize it for their voice so that it's kind of them, et cetera. They can fork it. So they can ask, ask cowork to duplicate the company skill into a personal skill. It will bring everything that there is, but then they can customize that duplicated version without affecting the one that everyone else has access to. It gives them a really strong starting point that will have all the company's guidelines while still allowing them to customize it and quote, unquote, Vibe code their skills a little bit. Like they're not going to completely go away from the workflow, but they can change it a bit. And what's really powerful with these skills at the company level is that if you connect them to a GitHub repository, you can, as a builder, you can just push updates and it's going to update on everyone's workstation for the skills. So it's like it's starting to take shape. The environment, it's still very new. And to be honest, it's not what gets the most spotlight because people are more excited for the shiny features than the sharing features. But it's getting better. And the one thing is, if you really want to get most of the value of this, you need to learn GitHub to be able to make the plugins marketplace for your company.

25:48

That's the one thing, and that's something which you're teaching in the AI accelerator in this new course, right?

26:57

I've started teaching GitHub in the vibe coding your website because you need it as well. And so the idea is that's kind of how I introduce it through something practical. And it's not just Vibe coding your website, by the way. You can make landing pages. You can do all of that if you want. You don't have to rebuild your whole site. But the point is that once you understand how this works, then it's much easier for you to kind of translate that knowledge into something like creating a plugins marketplace. So, yeah, when I released this course, one of the first lessons I'm going to do next is like, I'm going to make a module in a cloud code course on sharing with your team and I'm going to piggyback Right. On all the GitHub stuff I've done.

27:04

Yeah, something a lot of people have been asking for and it's a very kind of new frontier. So there hasn't until recently been a established way of doing this. So yeah, we are going to be covering that. I know there's a lot of business owners in the community that are interested in that. So that is coming soon. If you are a business owner, marketer, knowledge worker, and you also want to upskill yourself, you want to get into cloud code, or maybe you're already using cloud code and you want to see how we're using it. You want some of our skills, our processes, our workflows, our automations. Head on over to authorityhacker.com AIaccelerator and you can learn more about our program, what's involved, what you get, the community that we've got in there, which is really cool. And yeah, hope to see some of you in there as well. This podcast, not sponsored. That's the only thing we sell. This is our promo. We're done with it now, so. So let's move on to the next topic.

27:37

If you just skipped 30 seconds, welcome back.

28:29

This is OpenAI. So OpenAI is now apparently refocusing entirely on business use cases and coding. So the Wall Street Journal somehow got access to an internal meeting. I don't know, it was a leak or something, but basically OpenAI leadership is doing this big strategic shift and is going all in on coding, on business users, on enterprise users, because shock horror, that's where the money is, right? Who would have thought that your grandma looking for a lasagna recipe isn't going to be profitable for OpenAI. But a business trying to invest in content and marketing and coding and spend lots of money and lots of tokens would be profitable. So is this a case of like, you know, I feel like I could have told them this like many, many years ago. Like didn't most?

28:32

It's like a year plus ago they had a study that showed that only 5% of people are willing to pay for ChatGPT. And now for example, GP 5.4 mini is plenty good enough for most people already the consumers at their usage level was like, where's the recipe for lasagna? Or replace web search or very simple things. The mini models are good enough. There's no need for state of the art expensive models and so on. And so it's almost like now because the costs are decreasing to serve these users. These mini models cost a lot less to to run than even a GPT4.0, for example, back then, they get opportunities to use these models for less money. For example, Gemini gives you a lot of free usage. Claude started giving a lot of free usage, actually, finally. And so the case for paying $20 per month for everyone paying $20 per month is actually evaporating because they're getting the same thing for free almost. And they don't value the increase in quality of responses because their queries are so simple that the models can handle them anyway. But when we get to coding, when we get to complex use cases, et cetera, every time a new Opus model releases, you are excited because you're like, finally, it can do these things without me having to tell it five times to correct it. It's like, maybe I'll tell it three times only. And so we're willing to pay much more money, like 100, 200 bucks per user, probably more. As I said in a previous episode, soon, I wouldn't be surprised if eventually it's like the $200 plans are $500 or $1,000 a month, and people will still be happy to pay for them as the models get better.

29:26

I think there's like, a psychological shift that people need to go through here because as a business, you have no qualms about spending thousands of dollars a month on an employee to, like, do one task. Yeah, obviously depends on your size. But, you know, that's a normal thing for businesses to do. Spending money on freelancers for editing videos or graphics or writing content. That was. I mean, one of our biggest costs over the last 10 years were those activities. But. But now the alternative, or the option is to spend it on tokens or a higher pricing tier on Claude or OpenAI or Gemini, there's a bit of friction or emotional resistance to spending this on something that's not human, even though demonstrably you get much more output faster, better and cheaper, which are all good things.

30:55

I don't think it's always better sometimes. And video editing is not always better, for example, and so on. So it's like, I think that's what they're doing now. They're subsidizing the cost so that you change your usage, and then you experiment at risk.

31:50

And then once they've got you, they just ratchet everything up.

32:04

Once you've fired your staff and you rely on these things, it's like it's easier to just pay more for the same thing that you already have set up than to rehire people and refigure all your processes and so on. And so that's kind of the bet they're making. And that's.

32:08

And that's also why, like, their valuations and the stock market in general is through the roof. Because the bet is that that will. Will pay off. But as a small business, that could be. I mean, there's huge upsides to it if you get locked in. That's problematic because that could be another just like, oh, we just tripled our costs. Therefore, you have to make a more profitable business or just shut down because you have no other option and no one knows how to do anything anymore.

32:22

I think it makes me think of streaming services, right? It's like all like, Uber, right? Uber is probably the better example. Uber came out. It's like, oh, my God, it's like, so much better with the app. And it's cheaper and it's easier and everything. It's like. And now Uber is just taxis. It's not better. It got more expensive. And it's like, now they need to make some profit, etc. It's literally, we're just back to taxis. And it's the same for streaming services, right? It's like, oh, my God, cable television, so difficult. You need to subscribe to all these things, and then eventually you just end up not having what you want, et cetera. The one thing Netflix brought is like video on demand, right? But we already had video on demand with other services, and then there was all the competitors, and now you subscribe to five streaming services that replace the five cable services that you used to pay for. And it's like, we got better shows, we got better TV shows. So there is some progress, but there's also, like, as markets mature, there's always a regression back to how things were before. And I would be surprised if AI doesn't go through the same cycle where you kind of end up. You'll pay, like, 30% less than you used to pay your employees, but not 5% like you do now. And you'll end up with some frustrations as well. And it will be better, but it will not maybe have changed things as much as you want, but in between, Uber was amazing at the beginning, Netflix was amazing at the beginning. And there is this kind of golden age for like, five to seven years for these services to establish themselves. We are there. This is where we are. The thing where it's going to be different is unlike Netflix, where they can't exponentially increase the quality of the shows. They have increased, but not exponentially. I think the AI models will keep improving, and so There will be a case for increasing the price as the models take on more and do more as well and do more autonomously. So you might not feel the sting as much when the prices end up going up, because compared to what the models do today, they will do like five times more, for example, and be like, okay, well, okay, I'm paying more. But. But if I look back at 2026 and what I was doing with Opus 4.6, this is a big step up, basically. So that's kind of the use case for that. And that's why people are bullish on it. I think it makes sense. I think the consumer market, nobody can beat Google. We're in Europe, so we don't get all this stuff. But Gemini is in Chrome right now. The integration of Gemini in Android is looking incredible. They're going to announce everything at I O in May, but it's going to be incredible. Like, it can order Ubers for you and order pizzas and everything. You just talk to it. It's crazy.

32:50

And so the reason they're able to do that is because they have the distribution.

35:16

Distribution, exactly. OpenAI has nothing. And as I think every user earns is like so much more difficult. And I don't think anyone will beat Google. And Google, you know, Gemini Street Flash is a good example. Very cheap model, but also very capable and very good for your mom. And for people who just need casual use of AI, there's no need for a better model. They will get better models, but, like, it won't change very much how they use it. Whereas OpenAI was trying to both play that market and the enterprise market, and Entropic beat them to that. And so they were kind of like eternal number two. They were behind Google in distribution for the consumer, and they've started to fall behind Entropic in terms of code that started taking out force, et cetera. They've done a lot of work on codecs, but obviously they can't do both. And it's like, if they want to be worth their variation, they need to be number one at something. And so they think.

35:20

It feels like they were just throwing a bunch of shit at the wall to see what would stick.

36:09

Yeah, the Sora app, the browser, the Journey, I've devices that are supposedly coming out this year as well, et cetera. But now they've basically decided we have much more chances to beat Anthropic than we have to beat Google. And so they decide we want to be number one thing and that thing is going to be enterprise because we Think that's the bigger market and the thing that, that hopefully will not make us go bankrupt, basically.

36:13

Which makes sense, right? I agree with them, but again, it

36:37

took them a long time.

36:40

I feel like anyone who's run a business for like any amount of time could have told them this at the beginning.

36:41

Send us your offers. OpenAI talk about it. Yeah. But overall, I think it's a good thing. I think it's going to prevent Claude from costing a thousand bucks per month by the end of this year. It might delay it by two or three years, basically.

36:46

Let's talk about search traffic from AI, because Ahrefs have released an interesting new website based on all the data that they get from their crawlers called chatgpt versus google.com and it's kind of like an interactive.

37:01

I did not use my website design skills because it looks like Mega Lab Coded.

37:14

Yeah, it's okay.

37:21

But no, it's the purple gradient. Like, by default, all these models do a purple gradient. Always, forever.

37:22

The interesting part in it, though, is that they're showing traffic drops over time across 75,000 websites that they have analytics data for. So, I mean, that's a pretty significant sample size. The drop isn't actually that significant as I thought it would be. It's 20% or almost 20%.

37:27

That's a lot. That's quite a lot.

37:49

It is, but, you know, in some places, you know, we've talked to people who have 50, 60, 70, 80, 90% traffic drops in Google over the last couple years. Yeah, yeah. Not even including content sites and those affected in years before it. So this is actually a little bit less than that. Like, if I had to guess what the number was foreseeing this, I would have said maybe 40%. So 20 or 19, 20. Whatever it is, it's not that bad. But I think the interesting point is that they're trying to make this comparison. Well, how much has AI traffic grown in that same time? And the answer is very, very little. The graph as you're showing on screen for anyone watching a YouTube version, basically the X axis has a green. A flat green line along the bottom that you think is the X axis, but it's actually the AI traffic. It's at like 2, 3 million and the scale goes up to like 600 million. And it just, it hasn't really increased significantly at all. If you look at just North America, DuckDuckGo sends more than double the amount of traffic that all of AI does to websites per month. And I can't remember the last time I Heard anyone talk about DuckDuckGo.

37:51

I think you need to also consider that the AI chatbot usage in that time period has massively increased. And so that needs to be taken into consideration. What that tells me is there's a drop in CTR actually, because the impressions from AI has been increasing significant, like probably 3x or 4x, like we're talking June 2025 to February 2026. Maybe not 3x, maybe 2x, let's say. But the point is, compared to the increase in clicks, which is only 300k clicks, it's like these chatbots are starting to do what all social platforms do, which is reduce clicks to the Internet.

39:02

Yeah. Now, anyone who's selling AEO or GEO will be fast to say, oh, but the value of the traffic from AI is so much higher. They're potentially right. But are they right? Is it whatever that is, 200 times higher? Probably not.

39:36

Right? Yeah, yeah.

39:57

I mean, all that's to say is that Google traffic is going down. AI traffic isn't really making a dent in things as things are. So you might. If you're heavily reliant on SEO and think GEO is the way out, while it is worth doing, you know, it's not going to replace it looks like, or it's not replacing the traffic drop that you may have seen from, from

39:58

SEO, I think we can read this. So if I play devil's advocate, it's like we can read this a little bit differently as well, because I think a lot of the SEO traffic dropping is probably top of funnel content, like, you know, how do I cook pasta? Or like, stuff like that. Right. Because that's the kind of queries that probably will go to a chatbot now and so on. But like, when you're looking for a solution, a service, et cetera, you kind of end up probably going to a website. And so probably the drop in SEO traffic was traffic that was very low value anyway and low conversion. And probably the clicks from chatbots, again, they're probably high value because that discovery is happening on the chatbot. You don't leave. So when you click, you kind of really want to go deep into that thing. I think there is a drop. I think there's a case for like, it's not like the drop in SEO traffic is not made up by AI traffic. There is effectively traffic disappearing from the Internet at this point and it's not being replaced anywhere. None of these other channels is really up or down. Clicks that convert may not drop as much as is suggested by this graph. As well, so it's like it's one perspective, there is many other ways to slice and dice the data. So it's like I still think there is some value in all that stuff, but I also think it's probably a little bit overhyped. And this idea that AIO Geo is a one to one replacement to SEO is starting to fade away a little bit.

40:21

Anecdotally speaking to agency owners, many of them are getting into this because they're getting asked about it. There's demand. It's because people in businesses, small business owners, but also teams in larger businesses, they're asking, hey, how can we get featured on an AI? You know, this is the next wave. Like they feel it's something they should be doing, something they're not currently doing. They want to get into it. So there's demand for that service. I just don't think too many people are kind of pushing back, saying that's not actually going to generate you a massive return. Now some companies might not necessarily care about that. It's like, you know, branding and positioning and market share and there's other factors that play there. I accept that. But it feels like a bit of a bit hypey, shall we say?

41:41

I think it's just people are trying to, they're like, oh, SEO traffic is disappearing, something's replacing it. What is replacing it? Oh, AI is growing. AI traffic is replacing SEO traffic. Oh, I need to do AI SEO instead of. And it's kind of like simplistic and people don't assume that the traffic that disappears may just disappear and may just not reappear somewhere else in the universe. And so I think that's the thing, it's working like social media and we're at a point now where, I mean, there are some ads on ChatGPT, but honestly not many. From what I've seen, it's not really running. But when this is going to get bigger, this organic traffic from AI will also be shared with ads and the ads will have images and everything. And then your organic link is just going to be like a tiny little pill at the end of a sentence that is clickable. Maybe there's two or three of them as well. And so I'm wondering if the organic traffic from AI is not going to decrease significantly maybe this year or next year when AI become more ad. Sorry, become more widespread on chatbots, just because if people click on an ad, they're less likely to click on an organic link as well. Yeah.

42:29

And the ads are going to be much more prominent as well, with like images and things. Whereas it's hard to even find the reference link to the paragraph that's been generated.

43:33

Sometimes I had to put the instructions in my cloth, my cloth. Now I kind of told it to link. It links like a website, it makes an underlined link and I can click on it. So it's like if you put that instructions in your clothes, it's actually quite good. And it's like I click more links that way. But if you don't, it's just these little pills at the end of the sentence and that's it.

43:42

So you updated your Claude MD file

44:00

with that instruction, even though I'm talking about the chatbot. So there's the custom instructions. There's two things I have in my cloud, custom instructions for the chatbot, like on my phone, et cetera. It's this one and the other one is make it use the Ask User Question tool. If you say that you know how in cloud code, sometimes when it needs information from you, it just pops these pop ups and you click an option, et cetera. And so my instruction is like, if the query is not very clear and you could use more detail for a better answer, use this tool. And so quite often when I do a query, it just pops and just asks me questions. But on mobile it's very nice. You just tap two or three times and you get like, it's way better than a smarter model. So I often run Sonnet on the mobile because it's fine and then I just have that and that creates like a 10 times better chatbot than anything else. And ChatGPT doesn't have it and Gemini doesn't have it. And that makes Claude by far the best chatbot. On top of the fact that they have much better memory. Actually, like the memory system on Claude is really good.

44:02

So staying on the topic of Ahrefs who had created that study, one interesting thing actually is they have announced that the API is available on all of their plans except this super, super cheap starter plan. This used to be something that was gated behind the like thousand dollar plus a month enterprise plan or you had to pay specifically for the API access previously. And it just seemed like they didn't want many people to have access to that. But that has changed now and they're like, we want to give everyone access to this now the limitations are, you know, still relatively low. I think you had 25,000 credits a month for access compared to 2 million on the enterprise. So it's still low. But the fact that people can start building more applications and use cases to integrate the Ahrefs data into their AI use cases, workflows and skills, et cetera.

44:55

Yeah, it's like, it's quite a big deal actually. It's like, like I was using data for SEO, but arguably the data is quite a bit worse than Ahrefs and it's like you still need to top up like $50 at once. So I would say now if you're going to use a fair amount of SEO data, provided you get enough access to data. I haven't tried the limits yet. Ahrefs starts to be competitive. I'm happy they're like one company that starts to understand that. Actually, I don't want another SaaS to log in onto now. I just want to connect all my things to my chatbot and it does it for me. And yeah, thanks, Ahrefs. I think that's going to be a big upgrade. I have some scale being accelerated. I'm going to upgrade based on this. Actually. I used to use the MCP and kind of like hack around. I basically made the MCP an API and I just was calling it many times in the background and so on. But now it's going to be way better actually.

45:48

It's an interesting kind of business development as well. They've kind of really, they're leaning into this whole AI revolution and making it like a core part of their tool,

46:33

better than most of their competitors. I think it's like that's why it's becoming interesting. That plus the cheaper plans, et cetera, you can see they're pivoting a lot and they're adapting a lot. And I'm excited to see where they're going. They released this Firehose as well. Like you can see, like installing Firehose that is kind of like an API for changes on the Internet. So if you want to monitor our competitors, et cetera. And they released an API first. Yeah, but like now it's like an API because you can literally just give this page to close code and it just connects to it. And even if you're not technical, it doesn't matter. I think internally they are starting to be quite adapting to this new way of marketing and the products they release show that. By the way, this site is 1000% vibe coded as well.

46:44

Yeah, it's interesting though as well. If you're a business, not just a software company, but especially a software company, it really starts to make sense to think about AI first use cases. So in the case of Firehose, the Installation instructions start with a copy paste prompt that you can give to your AI agents to set up, rather than step by step instructions for how to do that manually.

47:23

I think it's also an MVP thing now. It's like you don't need to make an interface or anything really easy like an API and a skill, which is exactly what they did here. And then if it takes off, you can make an interface for it, but otherwise you don't even bother. So that's what they're doing. They're testing the market. They invest in zero in design. This is vibe coded. And then if people use it, they will make a dashboard for it, they will make all of that and then they will make a product and they've

47:47

branded it as firehose. It's this new thing rather than part of Ahrefs. So it's not like they add a feature and take it away. It's just like, oh, firehose. Yeah, we're testing it, playing around with it. Fine, if it takes off, it takes off, but if it doesn't, like, no one's going to be super upset that they don't kind of expand on it further.

48:12

Yep, that's good. Yeah.

48:29

Cool. Final story then. And again, another person from Ahrefs, but I thought this was quite, quite interesting. So Ryan Law, who is their director of content marketing, he's been with them for a number of years now and has when we were in the SEO industry, someone who had like quite a high standard, let's say, for they have

48:31

a good blog, it's still one of the few companies that puts iPhone in their blog. Actually, if you go on there, they

48:51

do do it and that the reason for that is because their team has such high quality expectations and standards and they really like critique themselves and have quite strict processes for how to make content, but how to make content better as well. And he came out with a blog post this week where he basically says, I think AI content, it didn't used to be good enough, but now it is. And so if he's saying that, I really think we have arrived at this point where AI generated content is good enough for just about any business use case, especially on your kind of blog, your marketing side of things. Yep, it's interesting as well because I actually looked back on our YouTube channel and it was four and a half years ago in September 2022, I think it was 2021, 2022, where we did a video comparing. Don't know if you remember this one, Gail, comparing Jarvis AI to textbroker which was a Jarvis, was like pre chat GPT. It was AI content two, I believe. GPT two, yeah.

48:57

Yeah.

50:04

Chat GPT didn't exist at that point. And we compared it to textbroker, which was a platform. You could hire writer. Human writers to one.

50:04

Like, it was cheap.

50:14

It. It was there. Yeah.

50:15

Does it still exist?

50:17

I don't know, but I would be amazed if it does because, like, they do just be using AI on there.

50:18

They do.

50:24

Are they pretending it's humans that write

50:25

it or rank higher and convert more with custom SEO content? Yeah, they do, actually.

50:27

So, yeah. But is it a platform or get a free consultation? No, I think it looks like it's an agency.

50:32

Yeah, I think they've pivoted a bit, but.

50:39

But yeah, I think they would have had to. Yeah. I mean, this was one of the first businesses like that would have been affected by AI. All that's to say is in that video we said, oh, it's not about to replace all human writers just yet. It might be a while. Well, it's a while and we're here. Yeah. It really feels like it hasn't so much snuck up on us, but it's kind of. A lot of people discounted it and like, oh, it'll never be as good. It's always going to be biased or get things wrong or hallucinate or whatever. And then people who dismissed it too early on and weren't paying attention to it kind of missed the boat as it got better and better and better and just got slowly.

50:41

I think it's like the problem is like, I don't think that's the problem. I think most people know AI can write, but the real problem is that the SEO industry still has some stigma around AI content. It's like, oh my God, I need to humanify my content and it needs to sound like a human, otherwise Google is going to penalize me, et cetera. No, not anymore. Sure, you might want to edit the style of writing and make it more engaging, et cetera, but this whole idea of AI content bad and then human content good, this whole thing is not a thing anymore. And even I think Google is past that as well. Everyone's passed that apart from people in SEO that are stuck in the past. And it's like, okay, that was true when ChatGPT came out, but it's not true anymore. And even Ahrefs, you can see they have a list of blog pipeline skills. And actually, if you go on their blog post, you can see the screenshot of the skills they use, which is interesting to think about how they think about it as well. And so if you are still in that mindset of like I need to humanify content, et cetera, you are behind, you can make content, including ahrefs, that's probably more successful than you at their blog with AI. And it doesn't mean you shouldn't read it, it doesn't mean you can't give feedback to it, et cetera, the same way as a human. But AI can be the main driver behind that content and you can make high quality content, provided you build the right process. And now that's these skills allow you to do is combining these APIs and processes into making AI do it the right way.

51:22

And I think he sums it up best when he says not all AI content is good, but the barriers that prevented AI content generation from being good have fallen away. So you still need to kind of create those skills, those processes, those workflows and need to know what you're trying to output. If you just ask ChatGPT to write a 500 word article on the best sports socks, it's still probably not going to be very good.

52:39

But I think what's interesting is it's completely changed the content manager creator job. Like you are now an engineer and you have to build pipelines of processes that AI will follow. And that's kind of like AI still cannot do that very well on its own without guidance. Like I don't think AI will build the pipeline. That's great. Without any guidance.

53:04

You could argue though, that was the role before as a content manager, blog manager at least, you know.

53:25

Yeah, but you don't instruct AI the same way. Like AI needs to be told more explicitly many things that high quality writers would just do on their own. Right?

53:31

Absolutely. But you could also argue that if you have a team of people, that a lot of these things did need to be explicit and written down and documented because otherwise you get significant variance in your content or it would degrade in quality over time as people got lazy or just.

53:40

It's not just that, it's also, it's changes completely. Because before, like, let's say you made like a Checklist of like 50 things to check for an article. You know how this works. Like your humans are going to run it two times at the beginning and then they're just going to do it less and less. I know what it is. Whereas if you make these processes now with AI and you force it to do it, it will go through a checklist every time. And so you have a degree of consistency that changes quite a bit how you do things. And there are things you can ask to AI that you could not ask humans to do because it was so tedious. So it's like you have no limit on how tedious you can make the work in order to increase the quality. So. So that changes things quite a bit in the way you do things. You need to be more explicit and you have no limit on how tedious the task is and also how much time it takes, because rereading, proofreading, et cetera would take a lot of time to a human, but to AI, it's just like a minute or something. And so that's the difference. And so that changes quite a bit how you do things. And I think that's a big deal, because if you are a content creator, your skillset has to massively evolve to be able to operate these things. You can clearly see on the screenshot here, he's using closed code and he's using skills, and you need to learn these things. And instead of talking to humans, you're going to be operating agents this year, basically.

53:58

So if people want to learn how to do all that stuff, I would, in a completely obviously unbiased way, suggest that you go over to authorityhacker.com AIaccelerator where this is what we teach and you can learn how to do this kind of stuff for yourself. So, Gael, any final words of wisdom before we wrap it up?

55:13

I'm glad Ahrefs is on the same page as us now. It's kind of nice to. It's like they're evolving as a company. I know we talked a lot about them. They didn't even sponsor us. If you want to sponsor us after the video, just tell us ahrefs. But the point is, it's interesting because, you know, super hardcore SEO company that was making its brand about 100 million a year plus from the SEO angle, and now you see how they're pivoting, how they're changing their point of view, how they are giving you API access, et cetera. And it's like that's the pivot that you need to do if you come from that world as well. Even the repackaging the fire hosting is super interesting because they're taking something they already have in ahrefs, but they're repackaging it for a completely different audience, which is, ah, what if you get access to the data when a page updates, which we track with ahrefs, but that's useful for all online marketers, including ppc, et cetera, not just SEO. And they've been able to repackage what they do. And I think even if you're selling, say, SEO services, you can do the same thing. You can repackage your link building outreach to cold outreach. You can repackage your content, content creations to like video scripts or whatever. And the point is you need to look at the pieces of what you have and reorder them into new products so that you will match this new market that's evolving very fast. So I think that's food for thought for many people I think listening to this.

55:32

Excellent. Thanks very much, Gael. Then thanks for you for listening. Watching this episode. If you want to ask us a question, interact with us, give us some feedback, tell us why we're wrong about Geo, head on over to our YouTube channel and leave us a comment on this video there. We do go through all them and we'll try our very best to, to respond to them if you leave something constructive. So thanks again and we'll see you next week for another episode of the Authority Hacker Podcast. Bye.

56:47