OpenAI cancels Sora and a $1B Disney deal to prioritize GPT-5.5 codename Spud, with Polymarket pricing a 78% chance of an April launch.

OpenAI Axed Sora and Bet It All on GPT-5.5 ‘Spud’

by
💡 Key Takeaways

  • OpenAI killed its Sora video model and terminated a ~$1B Disney licensing deal to redirect all compute toward GPT-5.5, codenamed ‘Spud’
  • Polymarket gives 78% odds for an April 30 launch; the model merges coding, reasoning, and speed capabilities into a single unified architecture
  • OpenAI renamed its product division to ‘AGI Deployment’ — the first time AGI has appeared in any major tech company’s official org chart

One billion dollars. That was the price tag on OpenAI’s licensing deal with Disney. In April 2026, OpenAI walked away from it — all for a model codenamed GPT-5.5 ‘Spud.’

The Sora video generation model is dead, the Disney partnership is terminated, and every freed compute cycle has been redirected toward a single objective. Polymarket’s prediction market currently prices a 78% probability of launch by April 30.

OpenAI has renamed its entire product division to “AGI Deployment” — the first time the term artificial general intelligence has appeared in the official org chart of any major technology company. This is the story of what Spud is, when it arrives, and why OpenAI is betting everything on it.

Sora Is Dead: How OpenAI Cleared the Decks for Spud

OpenAI officially shut down Sora, its AI-powered video generation model. At the same time, the company terminated its licensing agreement with Disney, a deal valued at approximately $1 billion. Neither decision was made in isolation.

The logic is straightforward: strip away non-core initiatives and consolidate every available GPU hour behind Spud. In a competitive landscape where Anthropic and Google are accelerating their frontier model releases, OpenAI chose concentration over diversification.

The takeaway here is that Sora was not a throwaway side project. It was a revenue-generating platform with a billion-dollar partner. Walking away from that kind of deal signals a level of conviction about Spud’s commercial potential that goes well beyond typical model upgrades.

Key point: Killing Sora and terminating a $1B Disney deal is not a cost-cutting move. It is a resource concentration strategy — OpenAI is cannibalizing its own products to feed Spud.
GPT-5.5 Spud unified AI model concept visualization (AI-generated)
Three energy streams — coding, reasoning, and speed — converging into a unified core, visualizing Spud’s integrated architecture (AI-generated) — Generated by OpenAI DALL-E 3

Codename ‘Spud’: Pretraining Is Done, the Clock Is Ticking

On March 24, 2026, The Information confirmed that Spud’s pretraining phase was complete. GPT-5.4 had launched just 19 days earlier on March 5 — meaning OpenAI finished training its successor before the predecessor had even settled into production.

Sam Altman told employees that Spud is a “very strong model” capable of accelerating the economy. Greg Brockman, OpenAI’s president, characterized Spud as the culmination of “two years of research” with a distinctly larger-scale feel. Those are not phrases executives deploy lightly at a company that ships frontier models quarterly.

The critical distinction, however, is that pretraining completion does not equal launch readiness. Post-training — which includes safety alignment via RLHF (Reinforcement Learning from Human Feedback), red-teaming, and deployment testing — typically takes 4 to 12 weeks. GPT-5.4’s post-training cycle ran approximately six weeks.

That six-week precedent sets a realistic baseline. From March 24, six weeks places the most likely launch window at early May 2026 — not April, despite what the betting markets imply.

What GPT-5.4 Already Achieved: The Baseline Spud Builds On

To understand what Spud represents, it helps to measure the distance from GPT-5.4, which launched on March 5 and already marked a significant leap. That model shipped with a one-million-token context window, a 33% reduction in hallucination rates compared to GPT-5.2, and native computer-use capabilities — the ability to autonomously navigate desktops, browsers, and multi-application workflows.

Mario Rodriguez, GitHub’s Chief Product Officer, captured the shift at launch: “Developers don’t just need a model that writes code. They need one that thinks through problems the way they do.” GPT-5.4 moved in that direction. Spud aims to finish the job.

The key here is that Spud is not starting from scratch. It inherits a foundation that already operates computers autonomously and holds a million tokens of context. Whatever improvements Spud delivers will compound on top of an already high baseline.

GPT-5.5 or GPT-6? The Naming Question That Reveals Everything

Whether Spud ships as GPT-5.5 or GPT-6 has not been decided. OpenAI has stated publicly that the final name depends on the magnitude of Spud’s benchmark improvements over GPT-5.4.

If performance gains are incremental — meaningful but evolutionary — the model becomes GPT-5.5. If the leap is generational, it becomes GPT-6. In other words, the name is itself a verdict on the model’s capabilities.

This naming strategy reveals something deeper about OpenAI’s internal uncertainty. Even the company that built the model does not yet know whether it has a strong upgrade or a paradigm shift on its hands. The benchmarks, still being compiled during post-training, will decide.

In simpler terms: It is like applying to colleges before SAT scores arrive. The report card (benchmarks) determines whether you apply to a state university (GPT-5.5) or an Ivy League school (GPT-6).

Three Models Become One: Coding, Reasoning, and Speed Unified

The most consequential technical decision behind Spud is model unification. Previously, OpenAI operated three separate specialized models: GPT-5.3-Codex for coding tasks, GPT-5.4 Thinking for complex reasoning chains, and GPT-5.4 Instant for fast, lightweight responses. Each excelled in its domain, but switching between them required deliberate user intervention.

Spud collapses all three into a single architecture. The user no longer selects which model to invoke. The system routes internally, applying coding expertise, deep reasoning, or rapid response as the task demands.

Beyond routing intelligence, Spud introduces persistent memory across sessions. Users can maintain context over weeks of interaction — a capability that transforms the model from a per-session tool into something closer to a long-term digital collaborator.

GPT-5.5 Spud neural network unified architecture
Spud merges coding, reasoning, and speed into a single neural architecture — no more manual model switching — Photo by JJ Ying on Unsplash

The practical implications for enterprise customers are immediate. Engineering teams that currently maintain separate API integrations for coding assistance (Codex), analytical reasoning (Thinking), and customer-facing chat (Instant) could consolidate into a single endpoint. Reduced prompt engineering overhead, fewer model-switching errors, and persistent context would translate directly into lower operational complexity.

Key point: Spud unifies three specialized models into one architecture and adds persistent memory — turning a collection of AI tools into a single, long-running autonomous agent.

AGI Deployment: Three Letters That Changed an Org Chart

OpenAI has renamed its product division to AGI Deployment — a landmark move. For the first time, the term artificial general intelligence now appears in a major technology company’s official org chart.

The rename is not cosmetic — it signals a strategic pivot from positioning AI as an assistant to deploying autonomous agents that execute multi-step tasks independently. This is a company restructuring its identity around the premise that its models will act, not just respond.

Why this matters: the timing is deliberate. Sora’s termination, the Disney deal cancellation, and the AGI Deployment rename all happened within the same window. Taken together, they form a coherent narrative: OpenAI is restructuring its entire business around the premise that Spud — and the models that follow it — will function as autonomous economic actors, not merely conversational tools.

💡 Related Posts
Novo Nordisk × OpenAI: AI Drug Discovery Arms Race Begins

Polymarket’s 78% Bet: Dissecting the Launch Timeline

Polymarket, the largest prediction market platform, currently assigns a 78% probability to Spud launching by April 30, 2026. The probability rises to 95% or higher for a June 30 deadline. These are not analyst estimates — they are prices set by real capital at risk.

The math behind these odds deserves scrutiny. Pretraining finished on March 24, and post-training (safety alignment, red-teaming, deployment preparation) typically requires 4 to 12 weeks. GPT-5.4’s post-training took approximately six weeks.

At that pace, the most probable launch date falls in early May. Polymarket’s 78% April bet implies a compressed post-training cycle of four weeks or less — an aggressive assumption given the complexity of aligning a unified model.

Spud was trained at OpenAI’s Stargate facility in Abilene, Texas. If the four-week post-training scenario materializes, the model could appear by late April; at the 12-week upper bound, mid-June.

GPT-5.5 Spud training data center facility
A Stargate-class hyperscale data center — Spud’s pretraining was conducted at OpenAI’s Abilene, Texas facility — Photo by Geoffrey Moffett on Unsplash
Key point: March 24 pretraining completion + six-week post-training precedent = early May as the most realistic launch window. Polymarket’s 78% April bet assumes an aggressive four-week timeline.

The Competitive Landscape: Claude Mythos, Gemini 3.1 Pro, and the Race to Ship

Spud enters a three-way race at the frontier. Anthropic’s Claude Mythos — a model that sources describe as potentially the most capable AI system ever built — remains unreleased. Safety concerns have delayed its deployment, giving OpenAI a potential first-mover advantage with a unified model.

Google’s Gemini 3.1 Pro has already staked its claim, recording a 94.3% score on the GPQA Diamond benchmark, one of the most demanding reasoning evaluations in existence. GPT-5.4 countered with its hallucination reduction and computer-use capabilities, carving out a distinct niche in enterprise automation.

OpenAI’s strategy with Spud is clear: rather than competing on individual benchmarks, deliver a unified model that eliminates the friction of choosing between specialized tools. If enterprise customers can replace three separate model integrations with one Spud endpoint that codes, reasons, and responds at speed — while remembering context across weeks — the competitive dynamics shift from benchmark wars to platform consolidation.

When Spud’s actual benchmarks go public, the market will learn two things simultaneously: whether the model earns the GPT-5.5 or GPT-6 name, and whether “AGI Deployment” is a marketing slogan or a statement of intent backed by real capability. The Polymarket odds, the Stargate training timeline, and the naming decision will all converge on the same data point — the first public benchmark release.

💡 Related Posts
1,503 Elo: Anthropic Dethrones OpenAI After 3-Year Reign

References

Stay Ahead of the Curve

Get the latest insights on AI, tech, and markets — delivered in both Korean and English.

Explore More Posts