Summary

Consumer AI ended 2025 looking winner-take-most: ChatGPT held 800–900 million weekly active users, while only 9% of consumers paid for more than one of ChatGPT, Gemini, Claude, and Cursor. For most of the year, fewer than 10% of ChatGPT users visited another major provider. Olivia also cited Gemini as having added an estimated 35% of its scale on the web and about 40% on mobile, while Claude, Grok, and Perplexity each sat near 8–10%. Anish Acharya’s brand framing was simple: “ChatGPT is like the Kleenex of AI.”
Gemini was the live threat because viral creative models coincided with accelerating growth: desktop users rose 155% year over year versus ChatGPT’s 23%. Gemini reached roughly half of ChatGPT’s mobile scale on Android but only 17% on iOS—“everywhere” yet still “nowhere” in consumer habit. Justine Moore thinks it could get there if it sustains its image-and-video launches, though ChatGPT’s guided templates make the first creation far easier than Gemini’s blank box.
The year’s consumer model breakthrough was image and video models combining realism, reasoning, retrieval, and multiple media. ChatGPT 4.0 image’s Ghibli moment, Sora 2, Veo 3, and Nano Banana showed that accurate details, search-backed logos, consistent characters, and audio combined with video can create viral demand. The next architecture is “anything in to anything out,” potentially merging text intelligence, images, video, and editing into one model.
The labs’ distribution does not automatically produce successful vertical products, creating the panel’s clearest startup opening for 2026. Pulse, Atlas, group chats, Sora, Stitch, Gems, and Opal have not become breakout standalone consumer interfaces; NotebookLM was the notable exception. Bryan Kim’s caveat is that high-frequency assistants will remain hard to displace wherever the product is primarily text in and text out.
Sora 2 proved demand for AI video creation, not yet for an AI-native social network. A small creator cohort generated content for TikTok, Instagram, X, and Reddit, while in-app consumption, remixing, and commenting did not seem as strong as initially; the better analogy was “CapCut,” not TikTok. Bryan’s bull case is that humor could create a new status game through prompting skill and cultural awareness. Anish asked whether exporting still makes TikTok with Sora videos “strictly better.”
The most defensible near-term market may be prosumer and enterprise workflows, where depth of usage can invert traditional consumer economics. ChatGPT enterprise usage was said to be up roughly 8–9x year over year, while Claude and Comet showed the value of persistent workflows and cross-tool context. Usage charges above subscriptions are already producing consumer AI products with more than 100% revenue retention: “Maybe all of AI is actually a power user story.”
Compute remains the strategic constraint: labs must trade training against inference and entertainment traffic against coding intelligence, while focused application companies avoid that internal conflict. Anish said xAI was “probably the only” model company not bottlenecked on compute, “from my understanding,” while first-party-only labs also leave room for multi-model products serving power users. With model quality now sufficient to “build a real, scalable app,” the closing hope was that 2026 becomes a huge year for consumer builders.

Deep dive

1. ChatGPT owns the habit, but Gemini owns the momentum

Olivia Moore’s opening scoreboard made concentration concrete: only 9% of consumers paid for multiple leading AI products, and fewer than 10% of ChatGPT users visited another major provider for most of 2025. ChatGPT had 800–900 million weekly active users. Olivia also cited Gemini as having added an estimated 35% of its scale on the web and about 40% on mobile; Claude, Grok, and Perplexity were around 8–10%.
The snapshot concealed a sharp change in direction. Gemini’s desktop users were growing 155% year over year against ChatGPT’s 23%, while Android put Gemini at roughly 50% of ChatGPT’s mobile scale versus 17% on iOS—evidence that Google distribution works, even if consumer habit lags.
Justine’s answer to whether Gemini could overtake ChatGPT was “yes,” conditional on continued execution. Best-in-class image and video models generate “nearly infinite demand” from professionals and viral trends, pulling users into unfamiliar Google products; Anish framed the counterweight as ChatGPT’s status as “the Kleenex of AI.”

2. Creative models advanced from aesthetics to grounded reasoning

Justine identified the year’s consumer model hits as ChatGPT 4.0 image and its Ghibli moment, Sora 2, Google’s Veo 3 and Veo 3.1, plus Nano Banana and Nano Banana Pro. OpenAI largely kept features inside ChatGPT, whereas Google spread launches across Gemini, AI Studio, Labs, and standalone sites with more specialized interfaces.
Midjourney still stood apart for its aesthetic sensibility, especially without expertise in prompting, but the frontier shifted toward realism and reasoning: background pedestrians and cars moving correctly, multiple images and text resolving into one cohesive design, and infographics replacing the old triumph of merely rendering a letter correctly.
Anish’s underhyped point was accuracy through search. Historically accurate scenes, real product photography, market maps, correct company lists, and logos require retrieval as well as visual generation; similarly, Veo 3’s viral unlock was the non-obvious decision to bring audio together with video.
Limits remain visible in multistep composition. The panel’s benchmark—replace every Monopoly property with AI labs and startups without omissions, duplicates, overlaps, or misplaced names—still challenged GPT Image 1.5. Yet character and style persistence already turns repeated generation into storyboarding: once the model makes something useful, “you want to generate more.”

3. Product guidance and persistent workflows matter as much as models

Bryan contrasted Gemini’s blank Nano Banana prompt—“I don’t know what to do”—with ChatGPT’s TikTok-like menu of trending styles, one-click transformations, and follow-up ideas. Those “product nuances” get users through the first creation; his Snap-versus-Meta analogy suggested Google could copy successful interaction patterns and combine them with distribution.
Bryan called Pulse underhyped because he floated roughly 25 weekly ChatGPT uses as a basis for proactive summaries and nudges, moving toward the Western “everything app.” The pushback was immediate: Bryan said he was not a Pulse user, Anish said he had largely turned it off, execution felt off, and “99% of people don’t run their life on calendar.”
Connectors for email, calendars, and documents could let ChatGPT or Claude “own the prosumer workspace,” but remained unreliable. Olivia’s stronger example was Perplexity’s Comet browser: repeatable agentic workflows, sustained traffic above ChatGPT Atlas despite weaker distribution, and a credible path toward dedicated prosumer interfaces.
Olivia still preferred Claude for complex general work because it is “opinionated in an interesting way.” Artifacts, Skills, file creation, and Claude Code are powerful but packaged for technical users; three times more U.S. teens had reportedly used Character.AI than had used Claude, illustrating how MCP, Skills, and command-line craft have not translated into mainstream accessibility.

4. AI video found distribution before it found a native social loop

Bryan’s “inception theory” separates the emotional jobs underneath products. ChatGPT ultimately means “help me be better”; TikTok and similar social products address “Entertain me—I want my clown” and “I’m lonely. I want to be seen.” Adding group chats does not automatically move a productivity product from the first category into the other two.
Sora 2’s cameos were a strong bet, but observed behavior made it a creator tool. Bryan’s feed had become roughly two-thirds AI-generated, with more than half of it from Sora, yet concentrated creators exported their work while native consumption, remixing, and commenting did not seem as strong as initially. The cleaner analogy was CapCut.
The disagreement is worth keeping: Bryan argued AI-generated media weakens the personal “status game,” but his bull case was that humor could supply another one through prompting skill and cultural awareness. Anish responded that this was a different product and asked whether exporting videos makes TikTok with Sora videos “strictly better.”
Justine saw Meta’s strongest AI work in SAM 3 segmentation rather than consumer products; Bryan highlighted Instagram AI translations, which clone and translate a creator’s voice into five languages with lip-sync. Grok showed the “steepest slope” in image and video, rapidly adding text-to-video, audio, lip-sync, and 15-second clips. Elon has stated ambitions for interactive games and movies by the end of next year.

5. Enterprise apps and multimodality define the next platform contest

ChatGPT’s enterprise usage was cited as growing roughly 8–9x year over year. Mandatory workplace use could reinforce consumer habit, while its Apps SDK and Apps Directory could make ChatGPT a cross-tool workflow layer—an outcome with direct implications for SaaS vendors, not merely a new consumer distribution channel.
Justine’s architectural prediction was “anything in to anything out”: images, video, text, templates, and reference material entering one system, with edited or newly generated media emerging. From her conversations with the labs, they are trying to combine previously separate text-reasoning and generation efforts into “a mega-model,” with especially large consequences for design.
Olivia’s macro forecast was “more of the same.” Labs will keep improving models and their core assistants, but dozens of attempts at opinionated consumer interfaces have not worked; NotebookLM was perhaps one success among roughly 20 Google experiments. That leaves room for startups to verticalize increasingly capable underlying models.
Bryan’s caveat was that pure text-in/text-out startups will struggle against high-frequency assistants. He still expects the labs to generate common app types themselves, but said products like Opal had arrived “with a whimper” and were one-model efforts. Anish emphasized why founders may have an advantage: promotion systems reward safe extensions of core metrics rather than risky, opinionated products.

6. Power users change both the product stack and consumer economics

The unglamorous constraint is compute: labs divide scarce capacity between training and inference, then between Ghibli-style entertainment and coding intelligence. Anish’s hedged view was that xAI was “probably the only” model company not bottlenecked on compute, “from my understanding”; application startups do not face the same internal allocation decision.
Multi-model products also have a structural opening because labs and big tech remain first-party-model-only. A single model may provide “80% of what you need,” but power users monetize deeply enough that “maybe all of AI is actually a power user story, and everyone else is just traffic.” Usage fees above subscriptions have already enabled more than 100% revenue retention.
Justine recommended Google Labs’ Pameli as a glimpse of agents plus generation: provide a business URL and it pulls product and brand photos, summarizes the brand’s aesthetic and positioning, then produces three campaigns across copy, posts, flyers, and product imagery. Bryan, disclosing a16z’s investment in Krea, preferred using Nano Banana Pro through it because reusable characters, objects, and styles remove repeated reference uploads.
Anish’s daily utility was ElevenReader for converting saved documents into walk-time audio; Olivia chose Gamma for decks, Granola for contextual meeting notes, and Comet for an accessible AI-native workspace. Olivia also recommended Wabi’s constrained app generation and GPT-52 in Codex or Cursor—even for knowledge work—before Bryan closed on the claim that today’s models can support “a real, scalable app.”

Market Overview — December 30, 2026 UK Homebuilders with Christian Olesen from Olesen Value Fund