Sam Altman and Bill Peebles on Sora's launch: 70% of users are creating, and this is 'the GPT-3.5 moment for video'

Summary

Sora is roughly ten days old and already defying the conventional wisdom that most social platform users only consume. Bill Peebles, who leads the Sora product team, reports that 70% of users are actively creating content, a figure he describes as vastly higher than any legacy social platform. Approaching 2 million weekly active users, the platform is positioning creation, not passive scrolling, as its core loop.

Sam Altman frames the current product as 'the GPT-3.5 moment for video,' an analogy both he and Peebles lean on heavily. The implication is deliberate: just as GPT-3.5 showed clear promise but required GPT-4 to deliver genuine utility at scale, Sora 2 is the proof-of-concept inflection point, not the ceiling. Peebles notes the jump from Sora 1 to Sora 2 has been compressed relative to the GPT-1 to GPT-3.5 timeline, and both expect progress to accelerate further.

Model Capabilities

The headline technical advancement in Sora 2 is what the team calls Physics IQ, the model's ability to handle complex real-world dynamics including backflips, gymnastics routines, refraction, and fluid interactions. Prior video generation models failed reliably on these prompts. Peebles describes the model as 'hyper steerable,' capable of generating a coherent narrative arc from a minimal prompt while also supporting granular directorial control for advanced users.

A language model layer can augment text prompts, but Peebles is explicit that core physical accuracy must be embedded in the video model itself and cannot be outsourced to a prompt rewriter. Known artifacts still exist. Peebles identifies a tendency toward 'wired speech patterns' and rapid-fire dialogue as Sora 2's equivalent of the GPT em-dash tell. Altman flags doors and occasional object-phase-through errors as areas earmarked for the next iteration.

Product Design and the Cameo Feature

The Cameo feature, which lets users insert themselves or friends into AI-generated videos using a few seconds of reference footage, is driving the outsized creation rate. Peebles frames it as eliminating the friction barrier between consumption and creation. Crucially, the team chose to restrict likeness generation to the Cameo onboarding flow rather than allow open image uploads, a deliberate safety and consent design that has become a central argument in early Hollywood conversations.

Altman made his own Cameo publicly accessible at launch, drawing immediate pushback from Hollywood contacts who called it reckless. By day three, several of them had reversed position, calling it smart. He expects the dynamic to flip within months, predicting that rights holders and celebrities will be more concerned with not appearing frequently enough than with appearing at all.

The feed is AI-generated content only, with no mixed upload capability. Altman calls this 'a subtle but extremely important design decision' in shaping how users relate to the platform. He pushed back against the all-AI feed approach early in development before concluding the team was right.

Monetization and the Creator Economy

There are no ads on Sora today. Altman attributes this simply to the product being ten days old, not to any strategic objection to advertising. Peebles says creator monetization is a 'top priority' and that the team is actively designing a model around Cameos, brand and character licensing, and rights-holder revenue sharing. He commits to updates 'over the coming weeks' but offers no specific structure or timeline.

The Sora API is already generating what Altman describes as the fastest-ramping revenue of any new OpenAI model launch, though he does not give a specific figure. He says he plans to share an intended product roadmap with Peebles, an unusual step for OpenAI, specifically to help third-party developers plan around upcoming capabilities.

Competitive Posture

On the question of competitors replicating the Cameo feature, Altman is relaxed. He points to ChatGPT as precedent, noting that competitors have copied its interface down to deliberate design mistakes without materially eroding OpenAI's position. The stated advantage is the velocity of innovation and the coherence of the full-stack offering, from model training through to consumer product experience.

Altman acknowledges that Sora will never capture every use case for video AI and frames the API as an explicit acknowledgment of that. The strategic logic mirrors the ChatGPT API playbook: monetize infrastructure, let the ecosystem build verticals OpenAI will not pursue, and maintain the consumer product as the flagship.

Compute and Infrastructure

Altman says his day-to-day focus has shifted almost entirely to compute acquisition rather than internal allocation decisions. When pressed on what remains outside his control, he cites power delivery, specifically the difficulty of securing 10 gigawatts of power for delivery within the next year, as the binding constraint, not chip supply or partnership coverage.