OpenAI has completed pretraining on its next major AI model, internally codenamed Spud, and killed its video generation product Sora to free up the GPUs. CEO Sam Altman told employees around March 24 that the company should expect a release in "a few weeks," calling it a model that could "really accelerate the economy."
Whether Spud ships as GPT-5.5 or GPT-6 remains unclear. What is clear: this is not another incremental update bolted onto the same foundation.
A new base, finally
OpenAI President Greg Brockman, speaking on the Big Technology Podcast, laid out what makes Spud different from the string of GPT-5.x releases that have landed over the past several months. "I think of Spud as a new base, as a new pre-train," he said, adding that "maybe two years worth of research" is coming together in this model.
That distinction matters more than it sounds. SemiAnalysis, the semiconductor research firm, had previously flagged that OpenAI's recent models were built on older foundations rather than fresh pretraining runs. GPT-5.4, which launched March 5, was strong, scoring 83% on GDPval and merging Codex capabilities into one model. But it was still iterating on existing architecture.
Spud would be OpenAI's first genuinely new pretrain since GPT-4o, which arrived back in May 2024. If Brockman is right about the "two years of research" claim, that puts the start of this effort somewhere around early 2024, meaning the team was building Spud's foundations while simultaneously shipping five GPT-5 variants in under seven months. Whether that parallel development improved or fragmented the effort is an open question.
What Sora's death paid for
The compute trade-off is the most revealing part of this story. Axios reported that Sora was consuming a significant share of OpenAI's GPU capacity, and the decision to shut it down freed resources for Spud's training run. The Wall Street Journal was blunter: Sora "now looks like an expensive strategic miscalculation."
The collateral damage was real. A reported $1 billion Disney licensing deal, which would have brought over 200 characters to the platform, collapsed. Disney learned about the shutdown less than an hour before it went public, according to Futurism's reporting. No money ever changed hands.
Fidji Simo, OpenAI's CEO of Applications, told staff the company couldn't afford "side quests" while Anthropic was winning enterprise customers. That's a pretty frank admission from an executive whose team renamed itself "AGI Deployment" the same week.
So when does it actually ship?
Altman's "a few weeks" comment came on March 24. It is now April 5. No release date has been announced.
The post-training pipeline (reinforcement learning, alignment, safety testing, red-teaming) typically takes months, not weeks. Altman has a documented pattern of optimistic timelines. When he told The Information the model was ready for release, he may have been describing engineering readiness, not deployment readiness. Those are different things.
Brockman was more measured in his podcast appearance. He described Spud's impact in terms of "improved capabilities" and stressed that any single release is just "an early version of what we have coming." That's the kind of expectation management you do when you're not sure the launch date is locked.
What Brockman did emphasize is a quality he called "big model smell," the subjective sense that a model is genuinely smarter rather than incrementally tuned. "When these models are just actually much smarter, much more capable, they bend to you much more, and you feel it," he said. It's a vague metric, but the kind of thing that tracks with real user experience when it's true.
The broader bet
Spud isn't just a model. It's the engine for a planned desktop "super app" that merges ChatGPT, Codex, and the Atlas browser into a single interface. OpenAI wants users moving between conversation, code generation, web research, and task execution without switching tools.
To make room for all of this, Altman has stepped back from direct safety and security oversight. Safety now reports to Chief Research Officer Mark Chen. Security reports to Brockman. Altman is focused on chips, fundraising, and the $500 billion Stargate data center buildout in Abilene, Texas.
OpenAI is now generating roughly $2 billion per month in revenue and preparing for a potential IPO in late 2026. The Sora shutdown, the Spud bet, the reorg: all of it reads as a company consolidating around its strongest revenue lines before going public. Whether Spud delivers on the "two years of research" promise or just gives the IPO prospectus a better headline is something we'll know in weeks. Or months. Altman's timelines have a way of stretching.
