Why did OpenAI cancel Sora to build Spud?

OpenAI faced a 'Code Red' since December 2025, driven by the need to close the performance gap with Claude Opus-4.6 and Gemini 3.1. Sora's video generation team and compute resources were redirected entirely toward Spud and a new AI superapp merging ChatGPT, Codex, and Atlas. The pivot signals OpenAI's shift from creative consumer tools to enterprise productivity AI that generates revenue.

AI BusinessApril 4, 2026 · 7 min read

GPT-5.5 'Spud': OpenAI's Next Model Completes Pretraining — What We Know Before Launch

Q: What is GPT-5.5 'Spud'?

GPT-5.5 'Spud' is OpenAI's internal codename for its next major frontier model, expected to be released publicly in Q2 2026. Pretraining was completed around March 24, 2026, at the Stargate facility in Abilene, Texas, using a cluster of over 100,000 GPUs. The model is designed around intent inference — understanding what users mean from incomplete or ambiguous requests — rather than requiring precise prompt engineering.

Q: When will GPT-5.5 be released?

Pretraining completed around March 24, 2026. OpenAI typically needs 4–8 weeks for post-training (RLHF, safety evaluations, red-teaming, and fine-tuning). This puts the expected release window at late April to June 2026 — Q2 2026.

Q: What makes GPT-5.5 different from GPT-5.4?

GPT-5.4 achieved 83% on the GDPval benchmark but was deemed insufficient to decisively beat competitors. GPT-5.5 Spud focuses on three advances: (1) intent inference — understanding ambiguous requests without clarification rounds; (2) stronger agentic reasoning for multi-step tasks; and (3) a novel capability that OpenAI employees describe as 'very different from what we've seen before,' though specifics remain undisclosed.

OpenAI completed pretraining on its next frontier model — codenamed Spud — at the Stargate facility in Abilene, Texas, using 100,000+ GPUs. Expected in Q2 2026, it's designed to infer what you mean without precise prompts. Here's everything confirmed so far and what it means for the AI competitive landscape.

TL;DR

Codename: Spud | Public name: likely GPT-5.5 or GPT-6
Pretraining completed: ~March 24, 2026
Training site: Stargate Abilene, Texas — 100,000+ GPU cluster
Architecture: speculated sparse MoE, 1–10 trillion total parameters
Key advance: intent inference — understands ambiguous requests without clarification
Expected release: Q2 2026 (April–June)
Why it happened: OpenAI's "Code Red" to close gap vs Claude Opus-4.6 and Gemini 3.1

Why OpenAI Canceled Sora to Build This

The backstory matters. In December 2025, OpenAI entered what insiders described as a "Code Red" — the company felt it was falling behind Anthropic (Claude Opus-4.6) and Google (Gemini 3.1) on complex reasoning tasks that enterprise customers care about.

The decision was stark: Sora, OpenAI's flagship video generation product, was effectively canceled in March 2026. The Sora team and its compute allocation — representing billions of dollars in GPU investment — were redirected to Spud and a new AI superapp that merges ChatGPT, Codex, and the Atlas agentic framework into a single platform.

Sam Altman has described Spud as a "very strong model" capable of "accelerating the economy." OpenAI employees have hinted at a capability that is "very different from what we've seen before" — though specifics remain under NDA.

The pivot reveals OpenAI's strategic reality in 2026: creative consumer tools (video, image generation) don't generate the enterprise revenue needed to sustain $122B in funding and a potential IPO at $852B valuation. Productivity-oriented enterprise AI does.

The Stargate Infrastructure Behind Spud

Spud was trained at the Stargate facility in Abilene, Texas — the centerpiece of the $500B Stargate AI infrastructure initiative backed by SoftBank, Oracle, and OpenAI. The Abilene cluster represents a step-change in AI training scale:

Spec	Stargate Abilene	GPT-4 training (est.)
GPU count	100,000+	~25,000
GPU type	NVIDIA H100 / B200 (Blackwell)	NVIDIA A100
Compute (FLOPs)	~5–10× more than GPT-4 cluster	Baseline
Power draw	~500–750 MW (full facility)	~100 MW (est.)
Location	Abilene, Texas	Multiple sites
Ownership	Stargate JV (OpenAI + SoftBank + Oracle)	Microsoft Azure

What's Confirmed vs Speculated

Detail	Status	Source
Pretraining complete (~March 24)	CONFIRMED	Sam Altman / OpenAI staff
Training at Stargate Abilene	CONFIRMED	Multiple media reports
100,000+ GPU cluster	CONFIRMED	Stargate infrastructure reports
Codename 'Spud'	CONFIRMED (leaked)	Multiple insider sources
Public name GPT-5.5	LIKELY — not confirmed	Industry analysts
Sparse MoE architecture	SPECULATED	Architecture analysts
1–10 trillion parameters	SPECULATED	Compute scaling estimates
'Novel capability' unlike prior models	CONFIRMED (vague)	OpenAI employees (anonymous)
Q2 2026 release	LIKELY	Pretraining date + typical post-training schedule
Intent inference as core advance	CONFIRMED (executive statement)	Sam Altman interview

The Intent Inference Advance: What It Actually Means

The most substantively confirmed advance in Spud is what OpenAI calls "intent inference" — the model's ability to understand what a user wants from incomplete or ambiguous requests, without multiple rounds of clarification.

For context: current GPT-5.4 handles well-formed prompts reliably but frequently asks for clarification on ambiguous requests or makes assumptions that require correction. This creates multi-turn loops in agentic workflows that increase latency and cost. Spud is designed to collapse that clarification gap.

Current behavior (GPT-5.4)

User: "Write a report on our Q1 performance" → Model: "Could you clarify which metrics you'd like covered? Revenue, customer growth, operational costs, or all three?"

Target behavior (GPT-5.5 Spud)

User: "Write a report on our Q1 performance" → Model infers context from prior conversation, connected data sources, and document history to generate a complete report covering likely relevant metrics — flagging assumptions made rather than asking upfront.

For enterprise AI agents, this matters enormously. Every clarification round in an automated workflow introduces latency and sometimes requires human intervention. Intent inference could reduce agent failure rates on real-world tasks by 30–50% without any prompt engineering improvement by the user.

The AI Model Race: Where GPT-5.5 Fits in Q2 2026

Model	Company	Expected Q2 2026	Key focus
GPT-5.5 'Spud'	OpenAI	April–June 2026	Intent inference, agentic reasoning
Claude Mythos (Capybara)	Anthropic	April 2026 (leaked)	Cybersecurity, step-change reasoning
Grok 5	xAI	Q2 2026	6T param MoE, multi-agent verification
DeepSeek V4	DeepSeek	Q2 2026	~1T params, 1M context, Huawei chip support
Gemini 3.2	Google DeepMind	Q2–Q3 2026	Multimodal advances, Workspace integration

Frequently Asked Questions

What is GPT-5.5 'Spud'?

OpenAI's next frontier model, pretraining completed March 24, 2026. Trained at Stargate Abilene on 100,000+ GPUs. Designed around intent inference and agentic reasoning. Expected public release Q2 2026.

When will GPT-5.5 be released?

Pretraining finished late March 2026. With 4–8 weeks of post-training, the expected window is late April to June 2026.

Why did OpenAI cancel Sora to build this?

OpenAI entered 'Code Red' in December 2025, feeling behind on complex reasoning vs Claude and Gemini. Sora's team and compute were redirected to Spud and a new AI superapp — the shift from creative consumer tools to enterprise productivity AI.

What makes GPT-5.5 different from GPT-5.4?

The core advance is intent inference — understanding ambiguous requests without clarification rounds. Also: stronger agentic reasoning, and a 'novel capability' employees describe as unlike prior models (specifics undisclosed).

Build Agents That Work — With Any Frontier Model

HappyCapy builds on the best available frontier models — switching automatically as new releases like GPT-5.5 drop. Your workflows improve without rebuilding.

Try HappyCapy Free