OpenAI teases compute-heavy product launches — Sora 2, agents, and personalization on the roadmap

Summary

OpenAI plans to launch compute-intensive products over the coming weeks, including Sora 2, enhanced agents, and deeper personalization. Sam Altman signaled the shift in a weekend post, saying some features will be exclusive to Pro subscribers and new products will carry additional fees. Altman reiterated OpenAI's long-term intent to drive down the cost of intelligence, but the near term involves monetizing what's possible when throwing compute at new ideas.

Consumer personalization remains largely unfulfilled as a stickiness lever. Matthew McConaughey's appearance on Joe Rogan highlighted the gap. He described wanting a private LLM fed only with his books, journals, and personal notes, a system that would answer questions based solely on his own worldview. This resonates because it echoes the original custom GPTs pitch from OpenAI's Dev Day in 2023. Two years later, the GPT store has failed to reach mainstream adoption. The top entries are utility plays: astrology birth chart GPT, Scholar GPT, fitness coaches, humanized AI writers, image generators.

Memory currently remembers factual details like your car model or age but doesn't shape personality. When both users ask ChatGPT the same question, they get the same voice, syntax, and editorial style. OpenAI's new personalization tab offers preset tones: cheerful and adaptive, cynic, robot, blunt, listener, exploratory. Users can add custom instructions. But even with this enabled, the personality doesn't feel meaningfully distinct per user and hasn't driven the lock-in that competitors like Reddit and Twitter have achieved with algorithmic feeds.

Personalization as currently implemented is overhead, not moat. It doesn't create network effects or genuine switching costs. It's a feature that's easy to replicate. True stickiness would require the model to develop a unique personality over time that the user genuinely prefers to alternatives. That's a longer-term bet requiring heavy inference compute per user, which is why Altman's framing around compute-intensive offerings may include this.

Sora 2 will likely launch as a premium feature, potentially at $2 per generation or higher. Sora 1 is now roughly a year old and remains hallucinatory compared to video-generation competitors like Google's Veo 3. OpenAI needs to close that gap. High-quality video generation is compute-intensive. Google offers Veo 3 exclusively to Ultra subscribers ($200 per month). Sora 2 is expected to feature audio, 4K resolution, and possibly Nano for video editing. Features will carry incremental costs.

Google charged $500 per month but capped usage with rate limits, leaving money on the table. OpenAI appears to be learning from this. Rather than a fixed subscription, OpenAI may price per generation, potentially $2 to $50 per Sora 2 video depending on length and resolution. This aligns with Altman's messaging around finding the market clearing price for compute-heavy features.

ChatGPT agents remain in early stages but are coming with real transaction capabilities. Operator, the agentic web-browsing prototype, has been rebranded as Agent and is now baked into ChatGPT. Early use cases show it pulling restaurant menus, checking availability, and building custom meal plans. But it's still largely information-retrieval and doesn't book flights or place orders yet.

OpenAI's July 17 Dev Day explicitly promised agents would handle planning and booking travel itineraries, designing and booking entire dinner parties, or finding specialists and scheduling appointments. None of that has shipped at scale. The likely reason is integration complexity. True transaction agents require partnerships with specific services like DoorDash, Uber, and Expedia that have explicitly opted in. These integrations are expected to roll out alongside a Dev Day push in the coming weeks, oriented toward companies building on top of ChatGPT agent APIs.

AnthropiC is gaining traction in consumer and developer mindshare. It has launched a new brand campaign emphasizing problem-solving and human connection, a shift from its earlier corporate messaging. The imagery is warm, colorful, and explicitly human-centered, contrasting sharply with OpenAI's Super Bowl ad (dark, pointillist). Anthropic is also running campus ads and offering cheap or free student subscriptions, signaling a push to own the younger demographic before they default to ChatGPT.

Claude is becoming personified in San Francisco and Twitter circles, treated as a character or friend, not just a tool. This contrasts with ChatGPT (tool-like) and Gemini (no personality). The personification matters because it creates emotional stickiness independent of feature parity.

Claude's strongest differentiator remains coding. When paired with artifacts and HTML instantiation, Claude Code enables interactive research reports and educational software that neither ChatGPT nor Gemini currently match. If Anthropic can make code generation the default consumer experience rather than chat, it may crack a wedge that OpenAI hasn't.

OpenAI's Dev Day timing and content remain unclear. Altman said over the next few weeks, which could mean a dedicated keynote or a staggered rollout. Key questions include whether Sora 2 ships with transaction agents, how much of the new compute-intensive work hits free versus Pro versus a new tier, whether the model that won the International Mathematical Olympiad ships as a new release, and enhanced voice mode availability and duration.

Greg Brockman published OpenAI's internal study on ChatGPT usage, showing that practical guidance (28%), writing (28%), and information-seeking (21%) dominate, with coding at 4.2%. The breakdown suggests enterprise and professional adoption is driving DAU growth, not college students. This implies OpenAI's roadmap is enterprise-first, making consumer personalization secondary unless it also reduces churn at work.