OpenAI Kills Sora, Bets Everything on 'Spud' Model

OpenAI just sent an internal memo that reads like a strategy pivot disguised as a product update. The company is killing Sora, its video generation model, to free up compute for something called "Spud"—and CEO Sam Altman is using language he typically reserves for things he thinks will change the world.

"This can really accelerate the economy," Altman told staff about Spud, which finished pre-training on March 24th. That's not how you describe an incremental model upgrade. That's how you describe something you're preparing to call AGI.

The structural changes tell the same story. For the first time in OpenAI's history, AGI appears on the org chart—not as aspiration, but as product category. Senior exec Fidji Simo's division is now officially called "AGI Deployment." Meanwhile, Altman is stepping back from direct safety oversight to focus on infrastructure: chips, data centers, supply chains at what he calls "unprecedented scale."

What Dies When Sora Dies

The Sora shutdown wasn't clean. According to Wes Roth's reporting, employees were "shocked" at how much compute the video model consumed. That compute is now redirected to Spud and what OpenAI is calling the "super app"—a consolidation of ChatGPT, Codex, Atlas, and everything else into one product.

The casualties include a three-year licensing deal with Disney that involved a $1 billion commitment for AI-generated production content. The deal never officially closed—no funds exchanged—but it was apparently news to Disney when OpenAI pulled out. Disney issued the corporate equivalent of a shrug: a vague statement about respecting OpenAI's right to pursue its strategy.

The Sora team isn't disappearing. Bill Peebles, who led the division, said the research group will pivot to "systems that deeply understand the world by learning to simulate arbitrary environments at high fidelity." Translation: world models for robotics, the same territory Google DeepMind and Nvidia are exploring. Altman confirmed the team will "prioritize longer-term world simulation research, especially as it pertains to robotics."

This is OpenAI saying video generation was a side quest. The main quest is building something that can act in the physical world.

The Timing Isn't Accidental

OpenAI's memo explicitly names Google and Anthropic as competition. Both companies are reportedly racing toward IPOs this year, and multiple sources suggest OpenAI is struggling to keep pace with Anthropic's enterprise momentum. Anthropic ships tools like Claude Codework, Claude Code, and Dispatch (which lets you control AI agents from your phone) while OpenAI has been promising similar capabilities.

The company did acqui-hire the creator of OpenClaude, which might help. But the urgency in this restructuring suggests OpenAI feels behind.

What we don't know about Spud: parameter count, whether it's a reasoning model, whether it's multimodal, what it will be called publicly. What we do know is that Altman doesn't typically use "accelerate the economy" language for standard releases, and the org chart changes suggest the company is preparing to make AGI claims about something.

Meanwhile, In Actual Science

While OpenAI was announcing Spud, something quieter happened that might matter more: Terence Tao published a paper.

Tao—Fields Medal winner, often called the world's greatest living mathematician—included a sentence that reads differently now than it would have two years ago. After experimenting with Google DeepMind's AlphaEvolve, Tao wrote: "I was led to a way to prove the toy model's integral bound by splitting it into two lower bounds. ChatGPT proved the first and I proved the second."

Let that land. The world's preeminent mathematician just co-authored a proof with an AI model. He split the problem in half, handled one part, and had ChatGPT (likely GPT-4 Pro, though Tao doesn't specify) handle the other.

Tao maintains a GitHub wiki tracking AI-solved Erdős problems—notoriously difficult mathematical challenges. In January, GPT-5.2 Pro solved three Erdős problems in one week. Tao predicted in 2025 that AI would become a "trustworthy co-author" by 2026. That timeline appears ahead of schedule.

This isn't hype. Tao has no incentive to overstate AI's contribution—scientists who falsely claim credit typically do so by under-crediting collaborators, not by giving machines co-author status. If anything, acknowledging AI assistance this explicitly carries professional risk. He's doing it anyway because, apparently, it's true.

The Prediction Market

Roth frames this as a vindication moment for people who've been following the exponential curve versus those who've been betting against AI progress. The argument goes: if you've spent two years predicting AI capabilities would stall, plateau, or reveal themselves as overhyped, how's that working out?

It's a fair question, even if it's not quite fair. Skepticism about AI timelines has been reasonable—most technology predictions fail, most hype cycles collapse. The fact that this one hasn't yet doesn't retroactively make skeptics fools.

But it does make this moment interesting. When Altman, Demis Hassabis, and Dario Amodei said in early 2025 that AI would soon advance scientific research, a vocal portion of the tech community dismissed it as investor relations theater. Now Terence Tao is splitting mathematical proofs with ChatGPT.

When those same executives said superhuman coding was imminent, people pointed out that human software engineers still have jobs. True—but also, OpenAI Codex and Anthropic's tools are generating increasingly sophisticated code, and the labor market is already adjusting.

The question isn't whether these predictions were technically perfect. It's whether the directional bet was right.

What Spud Might Actually Mean

OpenAI is consolidating everything into one product, killing projects that don't serve that vision (RIP Sora), and restructuring around AGI as a shipping category rather than a research goal. Altman is using economic acceleration language. The org chart now includes AGI deployment as an actual job function.

Either this is coordinated messaging ahead of a genuinely significant capability jump, or it's the most elaborate setup for disappointment in AI history.

The difference between those possibilities matters enormously—for OpenAI's valuation, for the broader AI race, for whether we're in a genuine capability inflection or an extended hype cycle. But from a reporting standpoint, what's clear is that OpenAI believes it's in the first scenario and is making irrevocable bets accordingly.

You don't kill a billion-dollar Disney deal on a bluff. You don't reorganize your entire company around AGI deployment if you think Spud is GPT-4.5 with better benchmarks. You don't have your CEO step back from safety oversight to focus on infrastructure unless you think you're about to need dramatically more compute for something specific.

Whether Spud delivers on whatever OpenAI thinks it will deliver—that's the part we'll find out in a few weeks, apparently. What's already delivered is Terence Tao publishing mathematical proofs with AI co-authors, which wasn't supposed to happen quite yet.

The curve keeps going up. The question is what happens when everyone—not just mathematicians and developers, but the rest of the economy Altman keeps mentioning—has to navigate what that curve produces next.

—Dev Kapoor covers open source and developer communities for Buzzrag