2025年末，消费者AI走到哪一步了？

For most of the year, less than 10% of ChatGPT users even visited another one of the big LLM providers.

You open Gemini, and it has a pop-up that says, “We got Nano Banana. Would you like to do something with it?” There’s a little bit of pain where you have to type something.

Yeah. I don’t know what to do.

These are product nuances that I think make people actually take the first step.

The models have gotten to the level of quality where you can build a real, scalable app on top of them. The hope is that 2026 will be a huge year for consumer builders.

Today we’re talking about who won consumer AI in 2025. This was arguably the year that we saw the big model providers—OpenAI and Google, more than anyone else—make a major push of their own into consumer, both in terms of new models they released and in terms of new products, features, and interfaces that target the mainstream user.

You might wonder, why does it matter who is in the lead here? There are some early signs that the general LLM assistant space might be trending toward winner-take-all, or at least winner-take-most. Only 9% of consumers are paying for more than 1 of the group of ChatGPT, Gemini, Claude, and Cursor.

And for most of the year, less than 10% of ChatGPT users even visited another one of the big LLM providers, like Gemini. If we had to call it now, ChatGPT is currently in the lead by far, at 800 million to 900 million weekly active users. Gemini has added an estimated 35% of its scale on the web and about 40% on mobile, and everyone else trails significantly. Claude, Grok, and Perplexity are all at about 8% to 10% of the usage.

But especially in the last 3 to 6 months, things have been changing very quickly. With the launch of new viral models like Nano Banana, Gemini is now growing desktop users 155% year over year, which is actually accelerating even as it reaches more scale. That’s pretty crazy to see. ChatGPT is only growing 23% year over year.

We’re starting to see players like Anthropic almost specialize within consumer, owning different verticals, like the hypertechnical user. So today we’ve brought together the a16z consumer team to recap what we saw this year from the big model companies in consumer and also to predict what might be ahead of us in 2026.

Cool. Well, thank you, Olivia. It’s been a super fun year. If we wind the timeline back to last January, maybe we should start with what we saw launch—products, what worked, and what didn’t. Justine, tell us what you saw this year. OpenAI, Google—what are you paying attention to? What have you changed your mind on?

Those two in particular had a ton of consumer launches, like Olivia mentioned. From a model perspective, I would argue that their most viral models this year, at least among consumers, were in image and video.

For OpenAI, it was the ChatGPT 4.0 image, the Ghibli moment, which is crazy that that was this year.

It’s crazy that this was this year. It feels like it was years ago.

And then Sora, obviously—Sora 2. For Google, it’s Veo, Veo 3, and Veo 3.1, and then Nano Banana and Nano Banana Pro in image models, which went insanely viral, probably comparable to, if not beyond, the Ghibli moment for OpenAI.

I think in terms of the product layer, what we saw was that OpenAI tended to keep more things in the ChatGPT interface. polls, group chats, shopping, research, and tasks all launched inside ChatGPT as the core. The exception there is obviously Sora, as a standalone video app.

Whereas Google tended to launch more things as standalone products. They shipped a lot through Google AI Studio, Google Labs, Gemini, and the plethora of Google services available for launching a product. But they would also ship things as standalone websites that you could go to and visit, which basically allowed for a more custom interface for each type of product—not just the kind of chat entry, chat exit, or image video exit.

Well, Justine, I have a question for you on that. It felt like 18 months ago we were talking about Midjourney, and most of the multimodal models were defined by aesthetics and realism. Is that still true? What changed this year?

There were definitely different styles still. Midjourney, when you talk to people really deep in image and video, still kind of stands apart for this aesthetic sensibility that a lot of the models don’t have if you don’t know how to prompt for it.

But I would say this year in particular, we made a lot more strides on realism and also on reasoning within both image and video. That includes all of the little details that make an image or a video actually seem real. For example, if you have a person walking and talking, the people and the cars in the background, if they’re on a street, should be moving in the correct direction. They shouldn’t be morphing and looking strange.

In image, we were able to have multiple input images and text and reason across all of those uploads to create a cohesive design or something like that, which was not something we saw happening last year for sure.

Yeah, I remember when we were excited about having a letter show up correctly in images. Now we have insane infographics. We can just put up an amazing YouTube video and say, “Give me an image that explains this.”

Yes.

That’s incredibly different. Nano Banana Pro can even generate market maps. I would tell it, “Generate a market map.”

It’s incredible. It either has or will go do the web research within the image model, which is crazy, to get the correct list of companies and then pull their correct logos. That’s insane.

There’s one benchmark left that the reasoning image models have not cracked. I tested GPT Image 1.5 yesterday, and it sometimes struggles with both reasoning and multistep reasoning.

What I’ve been testing is that you upload a picture of a Monopoly board and say, “Remove the names of all the properties and replace them with names of AI labs and startups.” GPT Image 1.5 is actually the closest, but it’s very hard for the models to do all of those steps: remove the names, come up with the new names, put all of the new names in the correct places, make sure there aren’t overlaps, make sure you haven’t mentioned one thing 3 times, and make sure you haven’t left out another big player.

So there’s still some room to go on the image eval. But it’s interesting, especially with the image model from ChatGPT, where you can actually see persistence: it carries a character over into multiple image generations, with the same style.

Yeah, and I thought that was, “Oh, this is actually very interesting. We’re storyboarding.”

Totally. And once it does that, it makes you want to generate more.

For me, it felt like the most underhyped aspect of Nano Banana was the integration with search. It feels like there’s realism, which is physics and other things that make us feel like we’re at the uncanny valley. There’s reasoning, which is applying modifications that are adherent to what the user asked for. But then there’s also accuracy.

A good example of this is product photography. If you say, “Hey, generate a photo of this album cover,” or “Generate a historically accurate photo of this moment in time,” you actually need the search integration. That was nonintuitive, but it’s actually very useful.

Totally. It’s kind of like the Veo 3 moment. I don’t think it was intuitive to people that video would necessarily be cracked by bringing audio together with video in the same place, and that ended up being the thing that made AI video go viral.

Since Veo 3—and now Sora maybe dominates—my social feeds have been full of really realistic-looking AI videos all the time.

One-fifth of my feeds are AI-generated.

Amazing. There have been so many launches this year, and many of them went well, like Veo 3 and Nano Banana. What do you guys think is underhyped, or what products do you think didn’t get enough attention? Bryan?

It’s a good question. I think Pulse is probably still underhyped.

We’re talking about OpenAI and Google, which, to me, fall under the productivity category. If you go to the App Store today, 5 of the top 10 productivity apps are all Google. That’s insane. And ChatGPT is number 1.

We’re talking about a productivity category that helps you do things, and I feel like a lot of people are trying this from different angles: How do I actually ingest your data, your schedule, or your email to make it more helpful and give you more proactive notifications?

A lot of people are working on it. Given the frequency with which people use ChatGPT—which I think is, what, 25 times a week?

Pretty good. Pretty good—3 to 4 times a day.

It feels like it’s in a really good position to actually give you proactive nudges and summaries and help your life in general.

The everything app was always a myth in the Western world. I think OpenAI is trying to move in that direction, where it’s ingesting enough information, and enough people are going there often enough, to start giving really useful proactive nudges. That’s a space I’m excited about.

It’s interesting. But are you a DAU?

I am not a DAU.

Pulse?

Not a Pulse user.

Similarly, I tried Pulse for a while and have largely turned off of it. But I would agree with you that Pulse and a couple of other examples that OpenAI launched this year are new primitives or ideas that feel underhyped, even though the execution is a little off.

The execution—the UX is off. Another example that would be similarly about personal context is their connectors. Now you can connect your calendar, your email, and your documents. You can do this on Claude as well.

And so hypothetically, you could say to ChatGPT, “Read all of my memos over the past 6 months and summarize what's most interesting and least interesting.” I think when that works, it's really exciting. I have found it to be a little bit unreliable so far, but I think as the models get better, they have a real chance to own the prosumer workspace if they get that right.

Prosumer is the perfect category, because we talk about this sometimes, but 99% of people don't run their life on a calendar. We do, but that's what I'm thinking about: the actual average frequency of using ChatGPT. If it's 24 times a week, that's a pretty good place to start. Olivia, I feel like you are the ultimate power user. What are you still using? What's your stack?

That's a great question. From all of the larger model companies, actually, I would have to say the thing that I'm still using the most—and was maybe the most impressed by this year—was the Perplexity Comet browser. I was not using Perplexity as my core general LLM assistant; I use ChatGPT and Claude much more. But I think they really executed on it in a first-class way, in terms of both the agentic model within the browser and, perhaps more importantly, all of the workflows that you can set up that allow you to basically run the same task over and over, either at a preset time or when you trigger it on a certain webpage.

To me, that was a really exciting launch. If you look at the data, the spike at launch and the sustained traffic for Comet was actually much higher than for ChatGPT's own browser launch, Atlas, which is kind of crazy given how much more distribution ChatGPT has than Perplexity. I think they also launched an email assistant this year, and they made a couple of acquisitions of really strong agentic startups. What I would love to see from them next year is more of these dedicated prosumer interfaces. I feel like that would be an awesome direction for them to double down in. They do feel like the startup that has the biggest breadth of ambition, alongside the labs and big tech. It's very impressive, just the number of things they've shipped this year.

Yes, definitely. One thing I wanted to ask you, Justine, is Gemini feels like it's having a real moment because of all the image and video models. Do you think it can overtake ChatGPT? Is there truly that much demand for these types of models?

I think yes. What I've seen, basically, is that there is always nearly infinite demand for the best-in-class image or video model, because then you have a mix of tons of different people seeing it and wanting to use it. If you're using it professionally—in marketing, entertainment, storyboarding, or whatever—you always want to be using what's at the forefront of the field. So you're totally fine going somewhere other than ChatGPT and Sora to get access to Veo.

Even if you're an everyday consumer, so many new viral trends are created around new capabilities of the best-in-class image and video models. That ends up driving users into different products that they may have never tried before. You might be downloading the Gemini app or accidentally ending up on Google AI Studio, which I know they're trying to make more for developers, to use Nano Banana Pro, which a lot of users experienced in the past couple of months.

The interesting thing about Gemini to me is that, hypothetically, they benefit from the massive Google distribution advantage. If you look at Android, Gemini is at about 50% of ChatGPT's scale on mobile, whereas on iOS it's about 17%, so clearly something is working there. They launched a little Gemini widget within Chrome recently that encourages you to use it. They're launching it within Google Docs and Gmail and other things.

Yeah, but I think that most average people are still just using one AI product. ChatGPT is like the Kleenex of AI. It is the brand that has become synonymous with the category. I think Gemini still has a pretty big hurdle to overcome just in terms of that, but if they keep doing what they're doing with these amazing viral consumer creative-tool launches and model launches, they could get there next year.

What do you think about this? It's really interesting when you look at Gemini, which is everywhere.

Yeah, but yet nowhere, to some extent, right? When you look at the actual usage, people still think of Kleenex and go to ChatGPT. The interesting thing is also the product sensibility. This morning, I had 2 panes open: OpenAI's image model and Google's Gemini. I basically used the image functionality. When you open Gemini, it's a blank screen with a pop-up that says, “We got Nano Banana. Would you like to do something with it?” It's the little pane where you have to type something. I don't know what to do.

ChatGPT, you go in and it has a very TikTok-like style: “Here are the trending themes that you might want to generate.” You click on “I want a sketch pen,” or whatever, and they just use one of their pictures and create something amazing. Then it says, “Would you like a holiday card? Would you like a...” These are product nuances that I think make people actually take the first step to generate something. Once you have it, you have character consistency, so you keep going.

That's interesting, in that I think OpenAI and ChatGPT have proven that there's deeper product sensibility. This is a funny, maybe slightly non-kosher thing to say: I worked at Snap. When you look at Meta versus Snap, famously, Evan Spiegel was chief product officer at Snap. I wonder if there is a world where the ChatGPT team innovates on the product front again and again, and Google, with distribution, looks at them like, “That's cool. Let's just integrate it and keep going,” and actually plays that game.

The interesting thing there is that ChatGPT Images just launched yesterday, when we're filming this, in ChatGPT. Brand new. They had image models for years, and it took them that long to come up with a separate, relatively basic interface for generating images. I would almost argue that the application-layer companies, like Krea, Hedra, and Higgsfield, popularized that template format, did it first, and did it better. They are ChatGPT's product people.

Exactly. Flawless. Flawless. Well, maybe going in a slightly different direction, BK, I'm very curious for your take on OpenAI's social features, because it does feel like that's something where you really have to get product execution right, but also network design. There are some efforts around Sora 2—we should talk about that. There's also group chats within ChatGPT. You're our social guy, or have been historically. Bullish, bearish? Where's your head at?

Bearish for now.

To me, the reason is 2-fold. Historically, we look at products based on what I call inception theory: you go 3 to 4 layers down to figure out what the 1-liner is, which is, “I want my dad to love me.” When they think of a product, they're like, “That's for me,” as well as for a lot of you. I look at some of these products, like ChatGPT. Ultimately, when you peel the onion 5 times, I think essentially it's, “Help me be better. Help me get that information. Help me be more productive. Help me be more efficient.”

And then when I think about social features—Meta, Instagram, whatever, or even TikTok—the 2 layers of information or emotion that it's trying to address, to me, are, for TikTok, “Entertain me. I want my clown. Entertain me.” And then the other layer is, “I'm lonely. I want to be seen. I want to connect with people.” To me, these are 2 very different parallels in the product direction.

OpenAI's product is incredible. It's magic. It's amazing. But it's ultimately a “see me” or “help me” category, which is essentially why it's number 1 in the productivity category. Now we're trying to take this and shove it into people's lives and say, “Guys, connect. Connect better and actually feel like you're being seen.”

Even the group chat function, which I love, will be so good to plan a trip and actually have that common pain. But I think it still stops at probably a headcount of 2 to 3 people planning something in a “help me” way, versus, “Oh, I feel like I understand Anish so much better because I've sort of done that.” Largely, over time, I think that's the reason for that division. But that is not to say you can't build a separate product that completely addresses that.

I think Sora 2 was the other big social push this year from all the consumer AI companies. It was basically like a TikTok feed, but with all AI-generated video, and you could make cameos of your friends.

The cameos were a very good bet. It was a strong bet.

Yeah, and I think what we've seen in the retention data and how we're seeing it used is that it was massively successful as a creator tool. Now my feed is probably 2/3 AI slop, if not more, and over 50% of it is now Sora, whereas before it was all Veo and some claim. But it has not been as successful as a social app.

People—a small number of creators—are creating a ton of content and then bringing it out to TikTok, Instagram, X, and Reddit, where it's going massively viral. But it doesn't seem like there's as much consumption happening in the app, as much remixing, or as much commenting, especially as there was initially.

You know, in a funny way, the way I think about it is that Sora's competition or analogy isn't actually TikTok. It's actually CapCut.

Hmm. It's a funny way. It's almost like a creative tool.

Yeah, interesting. I think what I was going to say is that it goes back to your earlier point: the kind of motion that drives social apps is both these positive and negative feelings of, “Oh, I'm publishing this thing of myself that's kind of sensitive, or that I want people to think is this or that or this other thing.” That's what drives participation on the app.

Yeah. The status game.

Yeah, it's exactly the status game. And when it's AI-generated content, and people know it's not real—like, a real representation of you as a human being—the status game is lost a little bit.

Absolutely lost. I think the status game then becomes: can you prompt something very cool? But that's a different type of product, and that's why I think it goes viral on Twitter and all these other existing platforms.

My counterpoint, or bull case, for Sora 2 is that I actually think the status game was about humor more than anything else. And humor is the intersection of knowing how to prompt and being culturally aware. So I think that if they iterated on that, that's a direction that nobody has captured before.

Yes, but if you can export those videos, isn't it true that TikTok with Sora videos on it is strictly better than Sora? We talked about it so much: the ultimate social product is where consumption and creation both live together, or where the output of it is not native to other platforms like TikTok or YouTube Shorts.

So what do folks think of the challengers? We're talking about Sora 2. I mean, Meta—it's crazy to talk about Meta as a challenger. I guess in this context they are, but I think Claude, Perplexity, and Grok are the more obvious names for challengers. Olivia, what's your take?

I love Claude. I talk to Claude all the time. Claude has replaced ChatGPT for me as my general LLM. I think Claude is opinionated in an interesting way. I also love Claude because I'm willing to invest time into building out AI workflows. I think Claude actually launched a lot of really powerful things this year around Artifacts and Skills, where you can essentially set up tasks or workflows to run over time.

I do think the reason it hasn't hit the mainstream yet is that even the way they built those things is geared toward a technical user or an engineer. I think they tried to make Skills as easy as they could to create, and it still was not anywhere near easy enough for the mainstream consumer.

Another example would be that they were actually the first of the big players to launch file creation, slide-deck creation, and editing, and they branded it as file generation and analysis or something. It was a toggle feature within a setting bar of a setting bar or something, so very few people used it. And yet, to me, it's still the best product across all of them at doing that kind of complex work.

So I love Claude, but I think if they want to be a true mainstream consumer product, they need to dumb it down even more in terms of accessibility.

There was that survey you found recently of U.S. teens.

Yeah, I think it was 3 times more U.S. teens have ever used Character.AI than have used Claude.

Yeah. So I think that shows that Claude is a pretty broad thing.

Yeah. Claude is beloved amongst tech people, but outside of tech people, I think they are maybe struggling to pick up relevance.

It is interesting, though: if you look at the aesthetics, the product design, and the craft, 3 things that Anthropic did were MCP, Skills, and the command-line interface, Claude Code. Those are 3 surprising bets, especially Claude Code. I would have said, “Command-line interface, really? Is this the way that people want to interact?”

You were going to talk about taking over air mail and the thinking cap.

Yeah, that too. So, 3 things—where's the thinking cap? But it's sort of very high-minded design. I don't know if it's mass-market, or maybe that's apologetic on their behalf, but I think it is that it's opinionated and it's great.

Yeah, yeah. I do need to hear Justine's take on both Meta and Grok, as I feel like they both had fascinating years.

Yes. Meta hired all those researchers. I think their strongest models are actually not consumer-facing models. It's their SAM 3 series—Segment Anything for video, image, and audio.

Basically, for video, for example, you can upload a video and describe in natural language, “Find the kid in the red T-shirt,” and it will find and track that person across the entire video, even if they're coming in and out of the frame. It will let you apply effects like blurring them out or removing them or whatever. And you can imagine a similar thing with audio, with different stems, and then with images, with different objects in an image.

I think we're going to see next year, hopefully, some incredible consumer products built on top of those models, but today they're more of a playground for developers than they are a consumer-facing product.

Yeah, given the DNA of the company, the one good consumer feature I think they've launched this year with AI is Instagram AI translations. When you're uploading a Reel now, you can opt in to enable translations, and it will clone your voice, translate it into 5 different languages, apply the translation with your voice, and then redub it with lip sync.

Wow.

So it basically makes it seem like you're a native speaker in whatever language. I would love to see more of that stuff come to the Meta products.

Grok had a crazy year with the companions, with all of the LLM progress and the coding progress. I think their image and video progress is probably the steepest slope I've seen of any of the companies. It was probably 6 months ago that they didn't even have image and video models, and they're shipping so fast to launch new features.

Initially, it was just image-to-video; they added text-to-video, they added audio, then they added lip sync with speech, and then they added 15-second videos. They're just not slowing down the speed of progress, and Elon has made a bunch of statements about wanting more interactive, video-game-type content out of Grok and wanting movies out of Grok by the end of next year. So let's hope it continues to go at that pace.

Do you feel like it's a pincer movement where, on one hand, there's a very infrastructural model layer of, “Let's get to the top of the LLM arena charts,” and then the other one is, “Let's go, entertainment”?

I think that's a little bit of a bifurcated move. Right, the entertainment and the smart side.

Absolutely, but entertainment in a way that we're talking about Anthropic and ChatGPT's general population, but you just said Character.AI is way more popular.

Yes. So then how do we think about that? And I think it's a very interesting strategy in my mind. In the image and video app, since pretty early on, Grok has had templates of popular things, like you're standing somewhere and suddenly a rope drops from the ceiling, and you grab onto it and it swings you out of the scene. They have some really good ones that go viral regularly on TikTok and other places.

Yeah, really, really interesting. Well, maybe switching gears from 2025 to 2026, what are some of your predictions for next year? What do you think we'll see—hardware, models, commerce we haven't spoken about yet? What do we think will play out?

I know we're talking about consumer, but one of the things that's been really underrated for me about ChatGPT that we might see more of next year is that they've really made a push into the enterprise, both with the traditional enterprise licenses and then working with specific companies, even training models for them.

And I think when we think about the fact that most consumers only use 1 general LLM product, ChatGPT Enterprise usage—they published a big study—but it's up somewhat like 8 or 9× year over year. And so we're entering a world now where people have to use ChatGPT for their company or as part of their work. That could really translate into consumer usage, or maybe they become the workspace with the connectors and some of the other things that they're investing in, and someone else owns the consumer use cases.

I think, to that end, we have to talk about their push into apps, and I think whether or not that works is going to be the defining question for them next year.

Yeah, and I think we've all discussed the importance of the Apps SDK and the Apps Directory, as they're calling it, and it's going to be a huge new channel for consumers. I think what's less discussed is that it's hyper-relevant to enterprise.

Where ChatGPT shines is where it's able to operate across a number of tools for 1 workflow. And if you think about the number of things you do in your business day-to-day that operate across many tools, it's most of those things. So I think that will have very interesting implications for the SaaS ecosystem, and it's a part of the app store we're not talking about as much.

Yeah, maybe less of a prediction, but thinking through 2025 and talking about all the big moves from the big labs and from the startup point, I think one of the biggest trends we've seen is app generation. And I think there's a real world where we see the big labs, with the distribution and the frequency of usage of people coming in, start to say, “Look, maybe there is a common type of product and apps that we could actually help you generate within the confines of the big lab products.”

Yeah. I think that’s one of the interesting things which, again, going back to the supply chain of ideas and research, maybe that’s one thing. Again, nothing groundbreaking, but as we know, Studio Ghibli broke the internet. My cousin, who knows nothing about tech, sent me a Studio Ghibli photo. Well, let’s not send this to your cousin then.

And I think that goes to show that templates matter, that style matters.

Yeah. And I think about video, and it’s pretty freaking good.

Yeah. And it’s possible that we’re already at a point where it’s not necessarily just about the capabilities of the models of the big labs, but the stylistic things, the template. Think of TikTok. The core capability is largely still the same: music, trend, dance, go. Except the trend and format keep changing and keep it extremely fresh.

So, I feel like there’s a real world where the repurposer, or a team, or what have you, can start thinking about ways to really build video-first products, entities, live models. I think the cost will go down enough for people to try it out, and I’m excited to see that.

Yeah, I think what I’m most excited about is, along those lines, basically everything becoming multimodal. I call it anything in to anything out, which is basically—initially, especially with these image and video models, you put in a text prompt and get an image or a video out. You couldn’t really do much with it. And now we’ve started to see this with image-edit models, like Nano Banana, FLUX, and the new OpenAI model, where you can put an image in now and get another image out. You can put an image in with a text prompt and a direction, or put an image in with a template and another reference image, and get another image out.

What happens when you can put a video in and get images out that are related to the next iteration of the video? Or you can put a video in and a text prompt about what you want to edit and get the edited video out? From my conversations with the labs, a lot of them are trying to combine all of these largely separate efforts they’ve had across text reasoning and intelligence, the LLM space, and image and video into—what if we can merge those all into a mega-model that can take a lot of different forms of content and produce much more? I think it’s also going to have huge implications for design, because if you think about it, a lot of design is combining images with text, with video, and with different elements in interesting ways.

Yeah. I guess if I think about a macro-level prediction, I think it’s actually going to be more of the same. When we talk about what all of the labs have launched for consumers, they’ve done a great job with models, and they’ve done a great job with incremental things that improve the core experience of using something like ChatGPT or Gemini. In my opinion, we’ve gone through dozens of things that they’ve launched or tried as new consumer products or new consumer interfaces: group chat, Pulse, Atlas, Sora. Google has had a long tail, like Stitch, Gems, Opal—tons.

Yeah. None of those are really working, and I think it’s because it’s not the core competency of these companies anymore to build opinionated standalone consumer UIs. Out of all of those, I think the product that’s working the most is NotebookLM, and that’s one of maybe 20 things that Google has tried or experimented with.

So, I think it’s actually very positive for consumer startups in that the models will keep getting better, which the startups can use, and they’ll keep making ChatGPT better and better. But I don’t necessarily think that ChatGPT verticalizes into all of these other amazing use cases or products, and there’s still room for startups to be building there.

I have a yes-and to that: absolutely. But, however, when the input and the output are text—yeah—ChatGPT and Gemini shine the most. No matter how deep you go, no matter how specific you think your text output is going to be, essentially, given the frequency of use by users of the main big-lab products, I think it’s going to be really hard to stitch that and get that away from that usage if your product is mainly text in, text out.

So, I do think you have to be creative around what the angle is that you can use to steal people away from.

You know, I love that you used the word opinionated, because I think that for labs—certainly for big tech, and perhaps increasingly for labs—the priorities get set in their promo committee always. And if you’re a PM, and it’s always the sort of mid-career PMs—and I’ve been one of these—the incentives are always to get promoted, and the way to get promoted is to build something safe that extends a core metric and a core feature.

So, building opinionated products is a very risky way to manage your career, because they’re probably not going to work. They’re probably going to have a bunch of implications for legal and compliance, and the CEO might yell at you. I just think that they are so structured to do incremental things. The more founders do opinionated things, the more advantage they have.

I think, honestly, the big thing we haven’t discussed here, too, is compute, which is that the labs have this inherent tension: there’s a limited amount of compute, and they either spend it on training models or they spend it on inference. And even with inference, there’s this split between the entertainment Ghibli use cases and the coding-intelligence use cases.

I think xAI is probably the only model company that is not bottlenecked on compute, from my understanding, whereas the others have to make really serious and significant calls, like: if we release Nano Banana and it goes super viral, it may slow down the next big LLM we’re trying to push forward. Startups that focus on the app layer don’t have that problem, because there’s no tension there.

Absolutely. Yeah. We’ve talked about this before. I also think that there are categories in which being multimodal allows you to deliver a better proposition to the customer, and the labs and big tech are always going to be, definitionally, first-party-model-only.

So, I think as all the models get better, perhaps 80% of what you need can be achieved from a single model, but for the power users—and so much of AI is a power-user story—you always said that power users are just power users, and I think that’s true in a pre-AI world. But now the depth of value and the depth of monetization is so much higher that maybe all of AI is actually a power-user story. Everyone else is just traffic. Yes.

Yeah. Which is why we’re also seeing consumer products, for the first time ever, have more than 100% revenue retention.

Yes. And that’s what separates the good from the great from the exceptional in consumer AI. To be clear, how that happens is that they charge for usage, often in addition to a subscription. So, you can use beyond whatever your quota is for the month, given your subscription, and pay more.

An upgrade of the tier, or actually buying tokens or more usage. Yeah. That’s what differentiates it. If you told me, pre-AI, that we saw a consumer company with 100%-plus retention in revenue, I’d be like, “That doesn’t make any sense. That doesn’t compute.” Yeah. No pun intended. Exactly.

Well, guys, okay, maybe let’s start with specific recommendations. After this pod, what are the products people should download, or the features or the models? What should folks be using today?

On the multimodal point, I think one really underhyped product that people should check out—not because they’ll use it every day, but because it shows what is possible when you combine an agent with images and text—is Pameli. This is the Google Labs product where you put in the URL of your business, and it has an agent go to the website, pull all of the product and brand photos, summarize what it thinks your brand’s aesthetic is, what it stands for, and what kind of customers it’s targeting, and then generate 3 different ad campaigns for you.

It will generate not only the text, but also the Instagram posts. It will generate the flyer. It will generate the photo of your product wherever it thinks it should be, based on your customer. Very cool product. It would be hard to become a giant standalone product within Google, I think, but it shows the future of what happens if we combine agents with generation models that have a really deep understanding of context that an image model or video model normally wouldn’t have.

Startup products, though. Do you have a favorite startup product in creative tools?

I think we’re investors in Krea, so this is biased, but I think they’ve really done an exceptional job of being the best place to use every model, or every quality model, across every modality, and also building more of the interface on top of these models.

I now prefer to use Nano Banana Pro on Krea because Krea allows you to save elements, which are essentially characters, styles, or objects that you can tag to reprompt, versus having to drag the same image reference into Nano Banana over and over again.

It’s a good one. I suppose it falls under the startup category. Again, shilling companies, but the one that I use the most is actually ElevenReader. And the reason is, we’ve seen an explosion in podcasts, and there’s a reason for that, right? People are a lot more on the go. The reading capability of us reading, I think, is going down over time.

And so, let's not fight the reality. Let's embrace it. Let's find written material, translate it into listening, and do that. I used to be a power user of tools like Pocket. I didn't have time to read everything that I wanted to read, and it's a saving behavior, right? You're going around and saving all the things you eventually want to consume.

But I think what I do now is similar: I go get all the things I want to read, and I either PDF them or put them in ElevenReader. Once in a while, when I'm on a walk and I have 3 or 4 minutes, at 1.5x speed or 2x speed, I listen to one of these and get the gist of it. I think that's been a good way to use a little bit of time as a sort of semi-normal person.

Well, first of all, I love this question because I am strongly opinionated that by far the best way to get up to speed on AI is just to try a ton of products, and you get opinionated really quickly. I'm actually on Twitter for the whole month of December, publishing one new consumer product a day for people to check out.

So that's one way. I'll name 3 others that I think are especially relevant or interesting that people can plug into their workflows. One would be Gamma for slide deck generation. You can go from text prompt to slide deck, or from document to slide deck. I use it for everything.

Also, the slides have flexible sizes, so you're no longer editing every little pixel in your Google Slides to get it to fit into one, which is great. Another is Granola for note-taking. You might not have any meetings over the holidays, but in the new year, it just gets better and better the more meetings you have on it because it has the context of what you talked about before.

And then, lastly, I'm still going to plug the Comet browser. If you want to try an AI-native workspace, I think that's one of the most accessible ones to start with. I mean, for me, I've spent my whole year obsessed with coding and AI code. It's just been so tremendously fun. By the way, Bryan would take the other side of your argument that the big labs or big tech will win at app generation.

I think they just lack the focus. Products like Opal have been released with a whimper, and they're one model only. I didn't say they would win it. I think that we will see them doing it.

Yes, yes. I think that's true, but I think for the pure consumer side, of course, Wabi is really fun and really capable, and I think they're creating the right sort of constraints on app generation so that you can get a really satisfying, functional result. I think so far there's been a lot of overpromising in app generation, which has discouraged the early users.

I also think if you haven't tried GPT-52 in Codex or in Cursor, it's worth trying. Even for nontechnical people, it's just amazing. I think almost being technical is sort of a constraint because you have a pre-existing idea for what these models can do, and they can do a lot more. I'm hearing increasingly about people doing knowledge work and writing essays in Cursor instead of just writing code.

Wow. Just one thing I'm going to do at year-end, just to plug in a popular trend I see on TikTok: there are people who say, “What is the most unhinged thing I said this year?” And it actually does a review of all the things that you said.

But I think, similarly, it'll be a good thing. I'm going to do this at year-end: “Tell me how to live a better life next year. Give me actual, unvarnished opinions and some direction.” I think it'll be helpful.

I love that idea.

I'm going for a worse life next year. [Laughter.]

Fantastic. Let's go full degen, guys. Any closing thoughts?

That was a good one. I mean, the obvious one is that we are very actively investing in consumer companies, and I genuinely—I think a lot of people say this—I genuinely believe that the models have gotten to the level of quality that you can build a real, scalable app on top of them. Wabi is a great example of this.

And so the hope is 2026 will be a huge year for consumer builders, not just consumers being consumers of a product.

Yeah.

Yes. Well, thank you all for a super-fun year in consumer and AI. We'll be back with more next year, and Merry Christmas, guys. This is a wrap.

Yeah. Happy holidays.

Happy holidays.

[Laughter.] Happy holidays.