GPT-5 After the Hype: Strategy, Monetization & the Next Billion Users

Speaker A: Foreign it has been weeks since GPT5 launched, which is basically a decade in AI time. But the hype fog has finally lifted and that's the perfect moment to ask what has really changed? What actually matters, and what were the actual product strategy decisions behind the release? I'm Brian Balfour, Founder and CEO of Reforge, and in this episode I'm joined by my co host Fareed Masavat and we dig into why OpenAI went all in on a one model to rule the mall consumer strategy, the quiet product moves aimed at the next billion users, and the weird tension between how they actually marketed GPT5 and what that product strategy actually says. We also go straight to the money, the gym membership economics of AI, why flat pricing turns into a house of cards, and how usage based, hybrid and outcome based models actually shake out when the cost to serve is variable and it's not actually likely that it gets 10x cheaper. If you want clear takeaways on strategy, distribution and monetization, keep listening as we're about to dive in. As Farid and I unpack in this episode, the economics of AI are forcing product teams to rethink how they work. Costs don't scale like cloud did, pricing models are under pressure and the pace of competition means you can't afford to spend months debating what to build next. You need faster ways to test, validate and move. And that is exactly why we built Reforge Build. It's an AI native prototyping tool designed not for those blank state startups starting from scratch, but for real product teams with existing users, data and design standards. We're currently in a closed beta, but here's what we're building towards being able to generate multiple product variations instantly instead of locking into one bet too early. Connecting prototypes to your actual customer feedback strategy in context, collecting rapid user feedback before you commit resources, and collaborating seamlessly with your team inside the same flow. If AI is rewriting the rules of monetization and distribution, then build is how product teams rewrite the rules of decision making. We are currently in a closed beta, free with unlimited usage. Apply to join the beta@reforge.com, it's been ages in the days of AI to talk about. 18 days, right? But that is ages and AI to talk about, uh, GPT5. But I actually think it's probably the best time to talk about it because of how much hype and noise and fog of war happens around these announcements. And so now I feel like that fog's cleared a little bit. Everybody you know has settled down a little bit, had some time to actually play with the models, get a realistic perspective. I think it's interesting to look back at GPT5, the launch, some of the you know, strategic decisions behind it, uh, as well and maybe talk a little bit about it. So why don't we start there for just you know, now that you've, now we've had a few weeks to sift through this. What are your overall kind of impressions or perspectives of what some of the decisions that were made around GPT5 and the product around it and what you're walking away with from your own learnings and impressions from the whole launch?

Speaker B: Well, number one is I am surprised but also into how all in they are on GPT5 being the default standard one model to rule them all for all chap GPT usage going forward. I think this is a ballsy move for lack of a better term. It is caused some people to freak out who have gotten very used to some of the other models. But I think it's a play to move to the next billion users for ChatGPT and I think that they are trying to really solidify themselves as the default LLM consumer product for as many people around the world as humanly possible. And they recognize that the nuances of uh, model switching, model behavior, different models doing different things is a barrier to entry for a lot of people. But I do think that what they've done is said hey, there's a single entry Point. It's ChatGPT model isn't important, we will do all of that behind the scenes. And what it's done is it's created a framework for them to improve that over time like more opaquely meaning like we can change certain things under the hood for different types of tasks and improve them without having to launch a new model or explain new things and also really just land it as a consumer product at its core. And I think there's some real trade offs here which I think are worth talking about. I'm impressed that they went for it like I think we thought of ChatGPT. GPT5 was always presented as GPT5 will be this AGI model.

Speaker A: Yeah. And I want to talk about that. Yeah for a second I want to talk about that.

Speaker B: But yeah, dead it's something different instead. It's more the framework for the future of how consumer AI will work is the way that I would frame it.

Speaker A: Yeah. So I 100, I 100% agree that if you take away like all the press, all the like user responses and even Sam Altman's own free hype up to GPT5 and you just look at the fundamental, the biggest things that were actually changed about the product. It's very clear to me that they were product strategy decisions around really in nailing down their position as the dominant uh, consumer player in this category above things like Claude. So a lot of it was about simplification around the, those models but even the slight tweaks around a little bit of personalization as well. Right. With the light color changes of the chat experience. And even since then they've continued to reinforce that. They uh, announced, I think it's called ChatGPT Go. It's their lower priced subscription plan for India I believe.

Speaker B: Yes.

Speaker A: And uh, where obviously there's huge market, huge tam, tons of consumers, tons of users, where the current pricing model and structure was a little prohibitive um, to really kind of gain adoption there. So you see them making additional strategic moves to reinforce getting to what you call like the next billion users. But that's not what any of the surrounding, you know, like uh, discussion was like Sam Altman's own comments. Even if you look at how they announced the thing, it was very much focused on model improvements. Right. Like making it feel like the next big model. And so it feels like there's this weird tension between like the actual product strategy, which is to be the dominant consumer player, and the choices around that and how these things are being communicated which tends to still be rooted in how we were communicating these things a year or two ago, which was about big model upgrades. And uh, those two things I can't make sense of because I think if they would have just positioned this around the choices around simplification and enabling the next set of consumers and all of that type of stuff, I think it would have landed very differently and you wouldn't have gotten a lot of the initial user backlash which was around how people felt like, you know, what the heck, this model isn't that much better and so on and so forth and you know, all the things that kind of result from that.

Speaker B: Yeah. And I think the other thing that I find interesting is their announcements that's related is the announcements around verticals, things like health, where they're like hiring leaders and leaning in and like really approaching them. And I think this new structure of ChatGPT as the 5 as the entry point for everything and then maybe there are under the hood different switches, different models, different products. Right. That are associated with different types of, of behavior. Like I'm looking for health information so that those model M improvements can kind of happen under the hood I think is really interesting in terms of allowing it to take on wider and wider use cases at higher and higher quality and being really specific about it. But to your point around the product marketing, why didn't they talk about it that way? The first thing that came to mind was maybe it's a little appley. When Apple comes out with a new iPhone, they do still talk about speeds and feeds even though they talk about how that isn't how they wanted to compete. It's got a better camera, it's got this titanium alloy, whatever that I think some of this like model quality, it's so advanced it can do this. And that is a way of sort of building trust with everybody. This is the best I can come up with, building trust, even if you don't fully understand it, that this is the right choice for me, that it is good, it is great, it is best in class. And I think what they're trying to do is if they can continue to solidify that advanced message, it sort of sets the bar for anyone who wants to replace them. At the beginning, a product has to be just good enough. It needs to solve a hair on firearm problem. It doesn't need to be good, it just needs to be solve a problem. But then to replace it you need to be not 10% better, but 10x better. And I think what they're trying to do is kind of push on the quality enough to be, this is the default, somebody else's model can't just be 5% better and gets and like beat us on these use, on these core consumer use cases. They're going to have to be 10x better. And it's sort of a bet that that's not going to happen, which I think is interesting.

Speaker A: So if I was to ask that, I think what you're actually saying is that a big part of the consumer's choice right now in this category is still, they're still thinking about quality to some degree, right? Like output or, or model quality. I think that's what you're saying and I think I agree with that though. I think they have to find new ways to communicate what quality means in this market. Because I think the deeper and deeper we get into this market, quality means much less about performance against data benchmarks. Right. And quality to what is now the average consumer I think feels very different in their eyes. And I don't know exactly how to communicate that, but I guarantee you it's not. We're 88.7% on the SWE benchmarks, you

Speaker B: know, for the international Math Olympic ad or whatever it is.

Speaker A: Yeah. In addition it has to be paired with if you want to go. If you wanted to use Apple as the parallel. What Apple pairs, you know, that type of messaging with is also just showing very simply in a very emotional way what they can do, what the user can do with the product. And so that's, that's the piece that's still missing for me. Now I think the question that the early adopters will have why do they need to make these choices to reach the next billion users? Right. Like what is the reason behind that? And, and so I think what we should talk a little bit about is I think both of us are familiar with Bengali caba who's now@ah, YouTube but was early on Instagram wrote this amazing piece with freeforge called the adjacent user Theory. And uh, it was a framework that they kind of used at uh, Instagram and, and he sort of coined and then solidified a little bit further after his time there to think about. We have this core set of users. You can think about layers of concentric circles, almost like a bullseye target. And then what products naturally do is they start to attract the next concentric circle out from them. And those tend to be lower intent users with lower knowledge of what the product is. They're less of the early adopter type. And as a result you tend to start to see all of these signals happen where your activation rates start to drop and maybe your retention gets a little bit worse and your monetization like free to pay conversion gets worse because your, your built initially for those early adopters and as you get wider and wider those outer circles of the audience have more and more trouble with the product that was built for different audiences. So now you have to start to mold all of your product and all of those flows in order to appeal and, and really gather and retain that next concentric circle. And I think that's what some of these decisions actually really represent once again which is I'll have to come in here, educate myself about what each underlying model is especially good at. Uh, select the right one when I have a specific task. All of those pieces that the early adopters were like you and I were fine with because we live in it every single day. But 80 year old dad who started using AI every single day and is texting me about it every single day like he has zero clue what all

Speaker B: of these are and actually that's a huge barrier for him but it's not just sophistication. So I think this is the other piece of the adjacent user is that these adjacent users have clusters of other commonalities that you can then target and try and like build for. So I think one of the examples Bengali and I have talked about or and I think he talks about in his article is for instance, and this is much more common for products with high word of mouth or virality, right? Because you can't control who's hearing about your product. You can't narrow on an ICP like you could if you were a B2B company doing outbound, uh, sales. Right? Is that people are coming to your product who aren't part of your original ICP or product market fit. They're trying to use the product and failing. And there's some amount of time where that's okay. Like for instance, I've been at many companies who are acquiring customers from outside the US and they're like, cool. We'll just exclude them from our metrics. They're not part of our growth story. But eventually you have to start to think about them. So for instance, it's not just sophistication, it's lookout. What does their Internet connection look like, what does their device look like, what's their willingness to pay look like, et cetera. And they tend to cluster in some way. What I think is interesting about AI products is that the use cases are so horizontal. Like it's mega horizontal in a way that like, even what we used to call a horizontal product isn't horizontal.

Speaker A: It's times 100.

Speaker B: I can use it to do meeting notes and collate them all the way to ask about what this, like, spot on my face is. What happens is those adjacencies are really complicated too, by language, by culture, et cetera. And so in order to hit the next billion users, you have to build your system in a way where you can adjust for all of that without changing the behavior for the clusters that are working. And I think this is why they need to get away from the, hey, here's a model. And, um, we're going to switch the model. And you can choose which model you want for which tasks. And the reason for that is the models aren't just improvements on existing tests. They're actually totally different outputs. Like, it's a little magical what happens. And I'm riffing on an article I just read. I don't know if you've seen this. I think I shared it with you called Building AI Products in the Probabilistic era.

Speaker A: I have not had a chance to read that article.

Speaker B: Okay. The author, uh, talks about a bunch of different things. It's a really fun and really deep read. But one of the ideas is that like, each model is really a new product and so you can't just swap the whole product on people. Right. Like, you need to have a system where you can adjust over time is one of the things I've been thinking about here. And so I think the pricing move in India is one of these things is like a lower willingness to pay. We need to do that. Very common in things. I think we'll see.

Speaker A: I think you're mobile. Yeah, mobile will be a thing. Yeah, voice will be a thing. I think your point about, uh, the vertical focus around healthcare, I'm sure they're looking at usage data saying how many people that this is just like a major underlying use case, you know, for xyz. And I think going back to the communication part of that, it's, you know, maybe the communication thing going forward is less about the specific technical differences every time they release a model, but it's more use case driven to say, you can now do. It's now X, Y, Z better. When you ask it a healthcare question. My question also was it feels like there was a lot of this communication about how better it was at coding. Right? That, that was, that was a huge part of this. And I initially looked at that and I was like, man, it feels like they're in a tit for tat with Anthropic, which feels to me a bit like a distraction because they are best positioned to replace search, the most lucrative business model and category that's ever existed. They are best positioned that and it's their game to lose at this point. Despite, you know, not having the initial distribution and history and stuff like Google and Facebook and so this tit for tat on the part with Anthropic, that, uh, felt a little distracting to me. But maybe to your point earlier, it's like they feel like they can need to continue to make moves to say, others in this space. They're not going to separate 10x from us. Maybe we don't end up 10x better to them on this. It's important dimension, but this is more of a defensive move. But I just kind of look at where they're placed right now to replace like one of the most lucrative things and I'm like, oh my gosh, how is this, how, how are the laser beams not like fully focused on, on that specific Focal point.

Speaker B: I'm curious, Brian, ignoring the theoretical and like, you know, positioning framework of it, what's your experience with GPT5 been?

Speaker A: Personally, that's a good question. So I think it, the context here is I mostly use this, mostly use AI for writing and essentially like research. Those are my two main use cases. I am not doing a ton of coding or app development, though others on my team have, have been been experimenting with it quite a bit. And I would say that I, I was using three before and I would say on those tasks I find it good, but I do not find it fundamentally better than what I felt like I was getting out of O3 or what I get out of Claude. Either, uh, Sonnet or. Okay, there are some like edge use cases like, uh, I'm redoing the office, so none of you have to look at my crooked basketball hoop in the background that everybody loves to keep commenting on.

Speaker B: You also have a crooked blind. That's been driving me nuts.

Speaker A: It's become, it's become a discussion point of the Internet. When I was on Lenny's podcast, the number of comments I got about it was ridiculous. But anyways, that's coming in a couple weeks, so you all don't have to stare at that. But I'm a terrible designer and, but what I have been doing is taking pictures of some of the work that they were doing and then asking ChatGPT to generate images of other things I might be able to put in the room, which I cannot do in Claude. That's a use case where the mixed models perform much better for me.

Speaker B: I, I, you know, I've been pretty locked in on, on chatgpt as my primary thing. I don't use Claude a lot. Some of that's habit, some of that is memory, some of that is, uh, projects. I've been starting to build projects with context and using that as part of my workflow and so it's hard for me to compare except to other previous GPT models. So one, I am always in the wrong model and like by accident either because I've switched it for something and weird stuff happens with projects and context and stuff and now I don't deal with that, which is really, really nice and surprisingly good. Like the number of times I've been not in the reasoning model and wanted to be in the reasoning model and then had to switch and start over again is relatively high. And like it's really nice to take that cognitive load off, off. So that's been good. The second is at least early on, I found that GPT5 was fast, like meaningfully faster, especially for reasoning things.

Speaker A: Yeah, fair enough.

Speaker B: And that in itself was a feature for me because the way I write, interact or work with ChatGPT is very much like high velocity back and forth like I would if me and you were in the same room working together.

Speaker A: And speed is one of those consumer decisions. That's what it's going to take to get to the next billion for sure.

Speaker B: And the speed thing recently I've had some really weird slow stuff. But in general I find it's quickness to a good enough answer and its follow ups and its way of encouraging iteration has actually put me in these fast loops where I can, even if the first response isn't right, the last one is way better. Allowing model choice behind the scenes is a big piece of allowing it to go faster for certain tasks.

Speaker A: I would also say like them essentially being the first to this, like the whole, with the model switcher and stuff, they're going to get a lot of data very fast and be able to improve all of those things behind the scenes. And so there. I don't think it's a forever compounding value, but I think there's a little bit of a short term arbitrage too. If you assume a bunch of other models go in that direction as well in order to simplify, there's a little bit of a data loop there to, you know, make that more accurate, more effective. I think they had pretty much hit the limit of launching a bunch of individual models and expecting people to select from them. But now you could imagine a world where underneath the hood they're actually building um, into much more specific, much more granular long tail, like medium tail, long tail, uh, type of experiences that the end consumer doesn't necessarily have to worry about. And maybe that's the next frontier to get m. Uh, you know, a lot more improvement and a lot more unlock.

Speaker B: Yeah, couple of things trade offs. One, does this make GPT or OpenAI worse for enterprise API or building on top of use cases is one question and my thesis here is while all of these are probabilistic, going back to the article I mentioned earlier, when you're building on top of a model to solve certain problems for your product, some amount of predictability of input and output is important.

Speaker A: This is why you're talking more on the developer side.

Speaker B: Yeah, on the developer side.

Speaker A: Yeah, you're talking more, sorry, not business

Speaker B: use cases, but developer use cases.

Speaker A: Yes, I think there is a trade off on the developer side for sure. Yeah.

Speaker B: Yeah, so I think that's interesting.

Speaker A: Yeah. Clearly also a trade off against power users too. Right. Early adopters and power which, which most products, most mass consumer products have to make at, at some point. Once again going back to the adjacent user theory and kind of jumping to the that next level. And so uh, and they clearly did upset some of those power users about selecting and wanting to be able to select models. They brought that back in the more power user tiers like, like Pro.

Speaker B: This is like the classic the loudest people are not the most.

Speaker A: The largest minority. Yeah, yeah.

Speaker B: The vocal minority. Quick, before we talk about monetization models generally last thing on GPT ChatGPT is my assumption is looking at the India move that making the free tier profitable, or at least somewhat profitable feels like the neck has to be one of the Next plays for OpenAI. Does that seem like. I'm sorry, without trying to make a clear prediction, I think we're gonna see ads happen.

Speaker A: I think, I think you have, or you have some subsidize the cost there so that you can continue. I mean the compounding advantage is to be able to put more and more into the free tier to drive more and more distribution and, and usage. But to put more into the free tier then yeah, they have, they have to subsidize it in some way, shape or form. So yes, like the goal is to

Speaker B: get every user on the Internet on to make ChatGPT the default like it is for search. That feels like an inevitable next move and I'm very curious to see what evolves there.

Speaker A: So yeah, and going back to the next distribution shift, that could come in two forms. It could come in ads or it could come more in an app store type of experience where they're taking up or both. Right. It could come in either though I think once again the search use case, they're just so well positioned for that that it feels like the obvious move.

Speaker B: So yeah, I mean I listened to the acquired part one on Google and the whole time I was like this sounds familiar.

Speaker A: Like over and over again. Yeah, yeah, that's right.

Speaker B: Feels like a very similar story playing out, let's put it that way.

Speaker A: Yeah, cool. All right. Monetizing AI. So there have been a lot of different posts and articles that have been circulating essentially the same topic and that is how do startups and companies actually make money in the day and age of AI? And the big shift has been moving from cloud where the cost per user uh, was fairly minimal and did not necessarily scale in the same way as AI is Is the, in AI, the cost to serve because of the underlying compute and underlying costs of all of these models is actually meaning meaningful. And what that's doing is it's putting a lot of pressure on monetization models that were kind of tested and true for a couple decades. And we have like a couple quotes around this and. But there's been a number of articles floating around around different categories. Whether it's like the AI coding categories, uh, the AI prototyping categories. I've also seen some related to things like um, all of the AI sales and marketing tooling is that there are hypotheses out there that a lot of these companies are not negative on margins. They're spending more on the underlying models than they're bringing in in revenue due to this. And you're seeing like all of these companies start to make much more rapid pricing changes than you ever would historically. I think historically you would make maybe a pricing change per year max. But you're starting to see a lot of these players make multiple pricing moves in the first half of the year. And I think a lot of it is in response to trying to adapt to this. And so it's creating new challenges. These costs are definitely like creating uh, new challenges. And so I'll leave it at that and start there and kind of get your, your commentary and then, and then

Speaker B: jump in to start just to, you know, the way I always think about pricing. This is so simple but it's worth pointing out is there's a cost to serve. There's your price and then above that there's the willingness to pay or let's call it perceived value of the product. Right there's the true value, the perceived value. And your price has to fit somewhere in between the cost to serve. That gap is your incentive to sell and the gap between the price and the perceived value is the incentive to buy. Why would a customer buy it? Because the price is lower than the value they are going to extract or feel or perceive from that product. And I think that. But it's not just that. In the cloud era, cost to serve was effectively zero, which allowed for a ton of new interesting business models like freemium product led growth free trials like viral B2B software, all these sorts of things. It's not just that it was near zero and now it's non zero. It's actually also that the cost to serve is highly variable depending on who the customer is, what their level of engagement is and what their use cases are. And I think that's the more interesting dynamic is that what we have is, and we see this in consumer AI and we see this in B2B AI, et cetera, that there are some people who pay X and only use a uh, small ratio 10% of X and there are those who pay X and use thousand times X. You know we've seen some examples published by Claude and other companies about their power users sort of being on the far end and in flat rate business models this, the incentives are very weird. Where I would call this the gym membership model, which is you need a bunch of low activity overpaying customers to subsidize the people that show up to the gym all day every day. Right. And that leads to all kinds of weird stuff as you see with the way gyms operate their businesses. Right. Which is not where you want to head. And so I think that's the interesting dynamic is, is what. And we're seeing this of course with usage based models being very, very popular. Where my question is will there end up being like two kinds of company in every space? Like the ones for the casual users that are priced really attractively but cut off the power users and the ones for power users that are really usage oriented. I have a hard time imagining how it will evolve that you can do both.

Speaker A: Yeah, well, so I just, I want to read a quote from. I'm going to leave the name Anonymous here for, for obvious reasons but uh, I was emailing with, with a VC friend of mine and we were going back and forth on this topic and, and this is. Yeah, yeah, yeah, this is, this is, but this is what, this is what this person said. I uh, recently met with a company that exploded to greater than 10 million of ARR with a seat based model but unlimited usage. Their 12 month customer retention was less than 25% and meanwhile the top 10 of customers accounted for 80% of usage and were collectively and highly unprofitable. In that model the customers who don't use and then churn are subsidizing the power users who don't churn and which starts to feel like a house of cards. Um, which, yeah, like that. And so in the gym membership model there's still a limit there of how often somebody's going to go to a gym and create costs for you. But I think in Claude Code's announcement of their pricing change they were talking about how a small percentage of users had figured out a way to run cloud code 24. 7 in the background. So like just think about the infiniteness of how, of how much more costs and usage that they, they could create. And so I think on this is this. It's clearly unsustainable. Let, let me say that first though, I think this is a natural part of the cycle. Yeah. Is it's kind of my take on it. And, and I think when you see all of these cycles, there's basically as the new technology kind of rushes in, there's all of these new creations. We start to figure out which ones work and which ones don't, which popular, all those types of things. And then it tends to be that the business model that truly fits and will fund it going forward kind of comes after it lags. Right. You even saw this with Google. You mentioned Google earlier and on that acquired podcast episode they talked about that where it took them. They did not have the business model fully figured out at the beginning and at the inception of uh, search.

Speaker B: Yeah. They thought enterprise, white label, industrial, how they were going to monetize. Right?

Speaker A: That's right. Yeah. Yeah.

Speaker B: And m. I've been thinking about this as this product market, business model fit. Right. Like at the end of the day you call it four fits with channel. I I'll just, just lump channel and business model into one. That how you go to market and how you sell your product has to fit with your market. Right. And um, the tricky thing is is that today the way we buy software is through some sort of well established predictable cost that and that's how the customer wants to buy and there's a mismatch with the cost to serve. Right now in every single exploding market there has been customer subsidy.

Speaker A: Right.

Speaker B: It has either been subsidy by VC$, subsidy by consumers, subsidy by government. There are all kinds of different ways to subsidize early behavior that I think are common. I think the analogy to use is not SaaS, uh, for these companies, but actually like Rideshare or delivery where for a long time there were complicated gross margin dynamics, many unprofit, but actually not as unprofitable as people said. It was just some. Again some were really unprofitable and some were profitable. And it took time to evolve the business model to get it to a point where it started to make sense. I can't get into tons of details about this from my time there at Instacart, but there were things that unlocked new monetization that eventually made the model work. We optimized pricing, we optimized efficiency, we optimize business model and details and we added on new layers of monetization. Um, so I don't think it's as people have always yelled about how everybody's going out of business in every wave.

Speaker A: Ravi Gupta, the former COO of instacred, has talked about this on his podcast. Uh, where like when he, when he got in, uh, for every order they were like losing $10 or something like that.

Speaker B: 12.

Speaker A: Yeah, something like that. Yeah, yeah. And, and so, um, yeah, like I said, this, this is a natural part of the, it's a natural part of the cycle. Now the thing is, is uh, that there tends to be something that forces companies to start to innovate and, and focus on costs. And that's the question is when, when does that forcing function come in? Because to a certain degree right now we're in an environment where there are a lot of dollars to fund this at the moment. And it doesn't feel like a lot of companies have been forced to focus yet. But that forcing function will come eventually. The well runs dry eventually and all these things. And a lot of times it's caused by some other macro market shift and people get more conservative and the dollars start to dry up and then it forces innovation. So this game is you have these models, you fund this uh, net negatively to get all of this distribution. And that gives you the opportunity to sequence to a better business model, sequence to something more defensible before the constraints come in. And because if you don't make that shift right before then, then you tend to be a casualty in, in.

Speaker B: Yeah. And look, we are seeing unprecedented scale of revenue growth for some of these companies in terms of speed to a hundred million, speed to 200 million, et

Speaker A: cetera, which is putting pressure on it, which is creating that pressure in a shorter period of time for sure.

Speaker B: And, but I think, you know, some people look at that and they're like, oh, well, you're selling something that costs two dollars for a dollar. But I don't think knows or cares about the cost, if that makes sense. They're using tools because they create meaningful perceived value for them. M. And there will be elasticity on that end. Right. And I would never bet against consumer desire in the long run. Yeah, the system has a way of, of solving the problem. There will be external forcing functions. But the fact is, is like people are adopting AI technologies faster than any other technology in the history of the world.

Speaker A: Yeah.

Speaker B: And I think that creates a ton of opportunity and I think it is smart for companies not to be over optimizing around their cost structures right now. But I do think at the end of the day, gross margins are really important. Uh, you have to think about them dynamically now in a way that you didn't in the past. Can't just like, sum it up and be like, here's our total cost, here's our total revenue, subtract. You have to think about it in a much more nuanced and segmented basis and start to design your business models around that. And, um, I think there's a lot of opportunity for the companies that innovate on that. The most business model innovation might be the thing that separates the winners.

Speaker A: Yeah, we haven't touched on this yet, but I do think it's important to touch on, which is, you know, everybody at this stage, they're taking bets on the long term. Right. And what to focus on and what's going to create durable value. And so, you know, one of those bets that has been floating around is it's okay to do this because costs are just going to come down 10x. But I, but I think we both read there was an article by, uh, a founder of TextQL called, his name is Ethan Ding, who actually kind of showed some data that this isn't necessarily true. And his whole hypothesis was, yes, the cost of maybe an individual token from the old models has been decreasing exponentially or been decreasing 10x. But if you look at the cost for the newest model, the best performing model, uh, at any given moment in time, that has relatively stayed the same. Now, I haven't seen that chart updated with GPT5. I know cost was a big part of that, that launch, so it'd be interesting to take a look at it. But he was saying consumers don't want the old models, they want the newest models. And on top of that, their use cases are becoming more and more complicated, meaning they're consuming more tokens. And so when you multiply those two factors together, the actual cost of running these use cases doesn't seem to be decreasing 10x. And to balance my position, hey, this is a natural part of the cycle. I also look at it and say, hey, if you're just sit there kind of betting that these costs for your product and your use case are going to come down 10x over time, that also doesn't feel right. Right. And, and as a result, there's something in the middle here of, okay, it's fine that you're playing that strategic game, but there also needs to be some experimentation on, okay, how are we going to align our monetization model over time to support these new dynamics that you're talking about, which is both the cost of serve as well as the massive range of usage that as a result of it.

Speaker B: I agree with this thesis. And the reason that we will probably not see the behaviors that people are doing getting cheaper and cheaper is because it's against what we've seen in computing for the past, you know, 50 years. Yeah, we went to the moon with a computer that was the size of a room and cost, you know, $50 million. But like your Apple Watch is like 4,000 times better than that. Now it is obviously still more inexpensive, but our demand for new software, new capabilities and new things always outpaces, you know, is outpacing the speed of the technology. There's sort of, I think of it as a max pain threshold. Like how long am I willing to wait for a keystroke to come back? There's some latency I'm willing to accept. You're just going to add autocorrect Once there's enough speed there or enough cost, you're not going to put a cheaper chip in my computer. We don't still use word processing machines. You know, we use general purpose computers. We have crazy graphics, we have 4K cameras, we're doing podcasts live over the Internet. Like that's going to happen in AI as well. But I do think perceived value will grow as well. And I think that's the piece of the puzzle that that article didn't really get into, which is, yes, cost may plateau or even go up as the number of tokens and the complexity goes up of the tasks. But I think the more I use chap G B T the more they could charge me, right, Because I'm sticky. And the more it's part of my life and the more I can't imagine living without it it. So I think like uber instacart, door dash, et cetera. There's a lot of price elasticity that will come over time as people develop high frequency use cases around these behaviors. And this just goes back to what we always talk about with growth. High frequency use cases have a lot of durability, right?

Speaker A: They do, they do. But so I've, I've been writing, I've been creating an AI growth course. It's our first major rewrite of the growth series in 10 years. I, uh, just wrote the monetization module with that Lauren Motametti, uh, from notion, and 25,000 words on it. It was a huge update. You mentioned perceived value. Uh, it's really important to understand that perceived value can't be incrementally more than the perceived price. It has to be multiple times more to get people over the hump. And there's been a ton of research on this. Historically ah, a lot of traditional SaaS pricing models only captured somewhere between 10, 15% of the actual value created inside of the organization overall. What we're doing with AI is there's this spectrum of pricing models from seep based to usage based to outcome based models. And the mix shift is certainly shifting. Right at the moment everybody's talking about outcome based but actual research by Kyle uh, Poyer shows that there's only a few percent of companies are actually using outcome based uh, models and most are shifting to usage based or some hybrid between C based and usage based as, as a next step. But, but I think the whole point on this is even in the outcome based models is that I still think that people, there's this whole notion that okay, well all, all of a sudden I can now compare this to the price of a human. I'm not sure that's durable over the long term because I think people and how people value software is just fundamentally lower than that. And so maybe over time if you price relative to the human budget and have some data model that, that uh, that you know, helps you create a fundamentally better output than any low cost competitor that might come in, maybe you can maintain that price point, maybe you can maintain it over time. But if, if not then 100% a Ah, lower cost competitor will come in and those prices will come down. But I was even listening to a podcast from uh, the founder of Intercom talking about fin and they priced theirs at 99 cents per outcome even though they or their research showed that for most of their customers it took between 10 and $20 to resolve a ticket. And they debated internally like hey, you know, we could price it at five bucks as an example and that would still be a huge savings over what customers are saying today. They ended up pricing at $0.99 because of this exact reason. They had the discussion in turn and they were just like, they found that people inherently value software differently than um, other things. And so even though that per outcome models that uh, research is showing that you can capture something more like 20 to 30% of the actual value created. It's not like it moves to 80% of the value created and you can charge at that because the perceived value has to be a multiple amount more than the actual price in order to get somebody to convert. And so I don't know like this. It's definitely clear that people have been willing to pay more for AI products. But I, I, but I, but this whole question about it being valued similar to hiring a human and those dollars, I don't think that ends up coming true. I think it's, I think it's much closer somewhere in the middle and much actually closer to software than it is on the other end of the spectrum.

Speaker B: I think that's very likely. I think that's very likely. I agree with you on that.

Speaker A: Yeah. Okay, we're over.

Speaker B: All right.

Speaker A: We got a wrap today.

Speaker B: We're out of time.

Speaker A: That's a wrap for today. Have you been nodding along? As we unpack the messy reality of AI economics and monetization, here's the takeaway. Product teams can't afford to was cycles on slow one off bets anymore. And that is why we built Reforge Build. It's an AI prototyping tool designed for real product teams, not just the greenfield startup. Starting from scratch, Build connects directly to your customer. Feedback and product context helps you spin uh up multiple variations instantly and get real signal from users before you commit resources. In a world where cost, pricing and competition are shifting under our feet, Build is how you de risk product decisions and move with speed. It's in a closed beta now, free. While we're gathering feedback, you can apply@reforged.com build. We hope to see you in the next episode.