Q&A with OpenAI President Greg Brockman about OpenAI's research direction, how far it can push Codex, closing Sora, betting on text vs. world models, and more
it is going to go to AGI. We have line of sight to much better models coming this year.”
Context & Ripple Effects
OpenAI has framed its models as steps toward AGI since its earlier AGI-oriented model roadmap, while later leadership interviews debated whether progress would come primarily from scaling, post-training, and reasoning improvements. Brockman’s comments put Codex and the research portfolio at the center of that continuing strategy.
The timing also matters because reporting in March described OpenAI’s race to strengthen Codex against Claude Code. Closing Sora signals a narrower allocation of attention around the model and product paths OpenAI believes can advance fastest.
First-order effects
- OpenAI’s stated closure of Sora concentrates its near-term research and product focus away from that project and toward Codex, text-centric models, and the capabilities Brockman identifies as nearer-term progress.
- Codex becomes a more consequential execution test for OpenAI’s research direction: its ability to improve and monetize determines whether that prioritization translates into a stronger developer-facing position.
Second-order effects
- Rivals in AI coding tools, particularly those competing for developers evaluating Codex, face a more focused OpenAI rather than a company spreading comparable attention across Sora.
- Teams and customers seeking OpenAI video-generation advances may need to reassess product plans, while demand and internal investment shift toward coding and general-purpose model workflows.
Third-order effects
- If OpenAI sustains this trade-off, frontier AI competition may increasingly be decided by which labs turn general models into high-frequency work products, rather than by maintaining every multimodal research bet.
- The move also highlights a persistent strategic uncertainty: text and code may offer a faster route to useful reasoning systems, but reducing investment in world-model or video work could leave room for competitors pursuing that capability path.
The trend: This is one data point in AI industrialization: frontier labs are pruning expensive research branches to concentrate compute, talent, and distribution on products that can compound model capability through real-world use.