The consumer AI landscape has shifted significantly as 2025 comes to an end. A small number of products now dominate daily usage, and major labs are focusing more intensely on consumer experiences. Despite the presence of many competitors, ChatGPT maintains a powerful hold on its audience. For most of the year, less than 10 percent of its users even visited another major model provider. This level of retention suggests that product nuances and user habits are as important as the underlying technology.

When you open Gemini, it has a pop up that says it has Nano Banana. It asks if you would like to do something with it. There is a little pain where you have to pick something, and you might not know what to do. These product nuances are what actually make people take the first step.

While large companies have a strong lead, the market still offers significant opportunities for startups. New methods for creation, such as templates and multimodality, are beginning to reshape user behavior. The quality of AI models has finally reached a threshold that allows for the development of truly scalable applications. This shift marks a transition from early experimentation to a more mature phase of product building, making 2026 a highly anticipated year for consumer AI growth.

The consumer AI market in 2025 shows signs of becoming a winner-take-all space. Most people do not use multiple AI assistants at the same time. Data shows that only 9% of consumers pay for more than one service among ChatGPT, Gemini, Claude, and Cursor. This suggests that once a user settles on a primary tool, they rarely explore other options. For most of the year, fewer than 10% of ChatGPT users even visited a competitor like Gemini.

ChatGPT maintains a massive lead with nearly 900 million weekly active users. However, the growth rates tell a different story. Gemini is growing its desktop user base by 155% year over year, while ChatGPT is growing at a much slower pace of 23%. This acceleration for Gemini is impressive given its already significant scale.

ChatGPT is currently in the lead by far at 800 to 900 million weekly active users. Geminis add an estimated 35% of their scale on web and about 40% on mobile, and everyone else significantly trails this.

Smaller players are finding their own paths by focusing on specific niches. Anthropic is starting to specialize in the hyper-technical user segment. While new models like Nano Banana have gone viral, the core competition remains a battle between the established giants and those carving out deep vertical expertise.

The past year saw a surge in viral image and video models from the major AI companies. OpenAI made waves with GPT-4o and the Sora video app. Google responded with the VO series and the Nano Banana image models. These visual tools captured significant consumer attention and pushed the boundaries of what these models can create.

For OpenAI, it was the ChatGPT 4.o image, the Ghibli moment. And then Sora too. And then for Google, it is VO3 and VO 3.1. And then Nano Banana and Nano Banana Pro in image mod, which went insanely viral.

Justine notes a clear difference in how these companies deliver products to users. OpenAI focuses on a unified experience by building features like shopping and research tasks directly into the ChatGPT interface. Google often chooses to launch standalone websites. This strategy allows Google to create custom interfaces that suit specific tasks better than a standard chat window.

OpenAI tended to keep more things in the ChatGPT interface. The exception there is obviously Sora as a standalone video app, whereas Google tended to launch more things as standalone products. This basically allowed for a more custom interface for each type of product, not just the chat entry and exit.

Aesthetics used to define image models, but recent progress has shifted toward realism and complex reasoning. While Midjourney still holds a unique aesthetic edge, newer models now handle intricate details that make videos and images feel authentic. For instance, cars in the background of a video now move in the correct direction instead of morphing unnaturally. These models can also process multiple image and text inputs to build cohesive designs.

There is one benchmark left that the reasoning image models have not cracked. I tested GPT Image 1.5 yesterday. They sometimes struggle with both reasoning and multi step reasoning. So what I have been testing is you upload a picture of a Monopoly board and you say remove the names of all the properties and replace them with names of AI labs and startups.

Justine notes that while GPT is the closest to solving these multi-step tasks, issues like overlapping text or missing key players remain. Despite these hurdles, new features like character consistency allow users to maintain the same person or style across multiple images. This opens the door for effective storyboarding. Anish points out that the integration of search within image models is a massive, underhyped advancement. This allows for historical accuracy and precise product photography by pulling real-world data directly into the generation process.

It is kind of like the VO3 moment when I do not think it was intuitive to people that video would be cracked necessarily by bringing audio together with video in the same place. And that ended up being the thing that made AI video go viral.

The combination of different modalities, like audio with video or search with images, often leads to the biggest breakthroughs. This convergence is what moves AI output from looking like a strange simulation to something that feels truly real.

Productivity apps are currently dominating the mobile market. ChatGPT and Google apps fill the top ranks of the App Store. Bryan points out that users visit ChatGPT about 25 times a week. This high frequency allows the platform to move beyond a simple chat interface toward becoming an everything app. By ingesting data like emails and schedules, the system can provide proactive nudges and summaries that help manage a user's life.

The everything app was always this myth in the western world. OpenAI is trying to move in that direction where it is ingesting enough people are going there enough to start giving really useful proactive nudges.

New features like connectors represent a shift toward a more integrated workspace. These tools allow the AI to access calendars and document stores to analyze long-term data. A user could ask the model to read months of memos and summarize the most interesting points. Although current execution can be unreliable, the potential for the prosumer category is massive. These users run their lives on calendars and need tools that can synthesize information across their entire digital footprint.

Olivia highlights the Perplexity Comet browser as a standout tool of the year. While she uses ChatGPT and Claude as her main assistants, the Comet browser excels through its agentic model and customizable workflows. These workflows allow users to automate tasks repeatedly at set times or specific triggers on web pages. Interestingly, the launch of Comet saw higher sustained traffic than the ChatGPT browser launch, even though ChatGPT has much more distribution.

I think they really executed on it in a first class way in terms of both the agentic model within the browser, but also perhaps more importantly, all of the workflows that you can set up that allow you to basically run the same task over and over.

Perplexity also expanded its capabilities by launching an email assistant and acquiring several agentic startups. This high shipping velocity signals a broad ambition that rivals much larger technology labs. Looking ahead, there is significant potential for the company to double down on dedicated prosumer interfaces that cater to power users.

The market for image and video models is driven by a constant demand for the best technology. Users in fields like marketing and entertainment are often willing to switch platforms to access the newest capabilities. This creates viral trends that draw people into new ecosystems. While Google has a massive distribution advantage through Android and integrations in Workspace, ChatGPT remains the dominant brand. It has become the household name for AI, much like Kleenex is for tissues.

The average person is still just using one AI product. And ChatGPT is like the Kleenex of AI. It is the brand that has become. And so I think that Gemini still has a pretty big hurdle to overcome just in terms of that.

A key difference between these platforms lies in their product design. Justine observes that a blank prompt box can be intimidating for new users. More engaging interfaces use a style similar to TikTok. They show trending themes or suggested templates. Bryan explains that these nuances help users take the first step in creating something. Once a user starts, features like character consistency keep them engaged.

When you open Gemini, it is a blank screen with a pop up and a little pane where you have to type something. You go into other tools and it has a very TikTok like style of here are trending themes that you might want to generate. These are product nuances that I think make people actually take the first step to generate.

There is a clear supply chain of product ideas in the AI world. Smaller application layer companies often popularize new interface formats first. Larger players like OpenAI then integrate these successful templates into their own platforms. This dynamic mirrors the relationship between Snap and Meta. In that scenario, one company innovates on product and the other leverages its massive distribution.

Bryan is currently bearish on the potential for OpenAI to succeed with social features. He uses a framework called inception theory to understand the core motivation behind products. In his view, ChatGPT ultimately addresses a desire to be better, more productive, and more efficient. In contrast, social platforms like TikTok or Instagram address different emotional needs, such as the desire to be entertained or to feel seen and connected.

I look at products based on what I call inception theory. You go like three to four layers down to figure out what the one liner is. For ChatGPT, it is basically help me be better, help me get that information, help me be more productive. For TikTok, it is entertain me. I want my clown to entertain me.

There is a significant divide between productivity tools and social connection. While features like group chats might be useful for small groups planning a trip, they still function in a utility-based way. They do not necessarily help people feel seen or deeply connected to others in the way successful social networks do. Productivity and social connection are two different parallels in product direction, and shoving social features into a productivity-first tool often fails to bridge that emotional gap.

New consumer AI platforms like Sora have seen massive success as creator tools but struggle to build native social communities. Much of the content appearing on social feeds is now AI-generated, yet users typically create these videos and then export them to established platforms like TikTok or Instagram. This dynamic suggests that these apps are currently functioning more like creative utilities, similar to CapCut, rather than standalone social ecosystems.

Sora's competition or analogy isn't actually TikTok, it's actually CapCut. It's like a creative tool.

The traditional status game that drives social participation is fundamentally changed by AI. Social apps usually thrive on people publishing sensitive or curated versions of themselves to gain status. When content is known to be AI-generated rather than a real representation of a human, that specific status game is lost. Instead, the new status game revolves around the technical skill of prompting and the creator's cultural awareness.

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. The status game comes then with can you prompt something very cool.

Humor represents a unique opportunity for these platforms. It exists at the intersection of technical prompting ability and cultural relevance. While a tool might generate high-quality video, its ultimate success as a social product depends on keeping creation and consumption in the same place. If the best AI videos are simply exported to TikTok, then TikTok remains the superior social product.

Olivia prefers Claude over ChatGPT as a primary language model because it is opinionated in a way that helps with building AI workflows. Features like artifacts and skills allow users to set up tasks that run over time. However, these tools are often geared toward technical users or engineers. For Claude to become a mainstream consumer product, the platform needs to improve accessibility for the average person.

I think Claude actually launched a lot of really powerful things this year around artifacts and skills where you can set up tasks or workflows to run over time. I do think the reason it hasn't hit the mainstream yet is even the way they built those things is geared towards a technical user.

Accessibility issues are evident in how new features are introduced. Anthropic was one of the first to launch file and slide deck creation, but the feature was hidden deep within settings menus. This lack of discoverability limits the growth of the user base. Data shows that US teens are three times more likely to use Character AI than Claude. While Claude is beloved in tech circles, it has yet to find broad relevance with younger or less technical audiences.

The product design choices from Anthropic are unique and specific. They have made surprising bets on tools like a command line interface and Claude Code. These choices reflect a high minded approach to design and craft. While these features are great for power users, they contribute to the perception that the tool is built for a specific, advanced audience.

Meta has developed powerful models like the SAM3 series that focus on segmenting video, image, and audio. These tools allow a user to describe a person or object in natural language and have the AI track them throughout an entire video. For example, you could ask the system to find a child in a specific shirt and then apply effects like blurring them out even as they move in and out of the frame. While these are currently more for developers than general users, they represent a significant step toward advanced consumer editing tools.

Their strongest models are actually not consumer facing models. It is their SAM3 series. Basically for video, you can upload a video and you can describe in natural language like find the kid in the red T-shirt and it will find and track that person across the entire video.

Instagram recently introduced an AI translation feature that clones a user's voice to translate their video into different languages. This includes re-dubbing with accurate lip-syncing so the creator appears to be a native speaker.

Grok is experiencing the fastest progress in the industry regarding image and video models. Within a few months, they moved from simple image generation to shipping features for text-to-video and audio synchronization. Justine notes that the speed of progress is the steepest slope she has seen. Elon Musk intends for Grok to eventually produce interactive video games and full movies. This focus on viral templates and entertainment suggests a strategy that prioritizes engagement over just standard chatbot functionality.

I think their image and video progress is probably the steepest slope I've seen of any of the companies. They are shipping so fast to launch new features. It was initially just image to video, then they added text to video, then they added audio, then they added lip sync with speech.

Enterprise adoption of ChatGPT is growing quickly, with usage increasing nearly nine times year over year. This growth is important because it might change how people use AI in their personal lives. If a company requires employees to use an AI tool for work, those habits often stick. The move into apps and the new SDK are part of this strategy. These tools allow AI to handle tasks across many different platforms at once. This is very useful for business workflows that usually require several separate tools.

We are also moving toward a future of total multimodality. This means you can put in any type of content and get any type of content out. Instead of just typing text to get an image, a user could put in a video to get edited images or a new video clip. Large research labs are working to combine text, image, and video models into one large system.

From my conversations with the labs, a lot of them are trying to basically combine these largely separate efforts they've had across text reasoning, image, and video into a mega model that can take different forms of content and produce much more.

Large labs are good at making models, but they often struggle to build successful consumer products. Many recent experiments from Google and OpenAI have not worked well. This happens because these companies are no longer set up to build unique or opinionated software. Career incentives for employees at big firms favor safe choices. Anish notes that product managers often focus on small improvements to avoid risks. Building a bold new product is dangerous for a career because it might fail or cause legal issues. This creates a big opportunity for startups to build the creative, opinionated products that big labs avoid.

Building opinionated products is a very risky way to manage your career because they are probably not going to work, they are probably going to have a bunch of implications for legal and compliance, and the CEO might yell at you.

Large AI labs face a constant struggle with limited compute resources. They must choose between using their power to train new models or to support current users through inference. This tension is particularly visible when a lab considers releasing a viral product. If an app becomes too popular, it might consume the compute needed to finish their next major model. Startups focusing on the app layer avoid this problem because they do not have to manage that same compute tension.

The labs have this inherent tension between a limited amount of compute. They either spend it on training models or they spend it on inference. If they release a model that goes super viral, it may slow down the next big model they are trying to push forward.

AI is increasingly becoming a story about power users. While casual users provide traffic, power users drive depth of value and monetization. This shift has led to a phenomenon rarely seen in consumer software before AI. Some companies are reporting revenue retention over 100 percent. This happens because users are happy to pay for more tokens or higher usage tiers beyond their initial subscription. Before AI, such high retention in the consumer market seemed impossible.

If you told me pre-AI we would see a consumer company with 100 percent plus retention, I would have said that does not make any sense.

Trying many different products is the best way to understand AI capabilities and form unique opinions. One interesting but under-hyped tool is Pmelie from Google Labs. This agent allows you to input a business URL to analyze a brand's aesthetic and target audience. It then generates full ad campaigns, including Instagram posts, flyers, and product photos that align with the brand's identity.

It shows the future of what happens if we combine agents with generation models that have deep understanding of context that an image model or a video model normally wouldn't have.

For creative work, Krea provides an efficient interface that allows users to save specific characters or styles as tags. This saves time because you do not have to repeatedly drag image references into the prompt. For consuming written content on the go, the Eleven Labs Reader is a powerful tool. It converts PDFs and articles into high-quality audio, which is helpful as reading habits shift toward audio consumption.

Several tools can also improve daily workflows. Gamma can instantly transform text or documents into slide decks with flexible formatting. Granola is an AI-native note-taking app that gains context from previous meetings to improve its summaries over time. For those interested in the web, the Comet browser offers an AI-native workspace that is very accessible for beginners.

Coding tools are also expanding into general knowledge work. While products like Cursor were built for programming, even non-technical people are using them to write essays and organize information. These tools often exceed the expectations of users who have a narrow view of what AI can do.

I am hearing increasingly about people doing knowledge work and writing essays in Cursor instead of just writing code. It is worth trying even for non-technical people. It is just amazing.

Investment activity in consumer companies remains high because technological models have reached a sufficient level of quality. This progress allows builders to create truly scalable applications for the first time. Wabi serves as a successful example of this trend. These advancements suggest that 2026 will be a major year for the growth of consumer technology and new applications.

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. The hope is 2026 will be a huge year for consumer builders.