OpenAI's GPT-5.5 'Spud' Finishes Pretraining — Greg Brockman Calls It a Massive Leap Toward AGI

OpenAI has completed pretraining on GPT-5.5, internally codenamed 'Spud.' President Greg Brockman says it represents two years of research and a massive qualitative leap in reasoning, coding, and agentic capabilities. 47% market odds it launches before April 30. Full breakdown: what makes Spud different, how it compares to Claude Mythos and Gemini 3.1 Ultra, and why this is the most anticipated AI release of 2026.

What "Pretraining Complete" Actually Means

When OpenAI says pretraining is complete, it means the model has finished its initial large-scale training on vast internet-scale data — the most computationally expensive phase of building a frontier AI. What comes next is fine-tuning for safety and alignment (RLHF, Constitutional AI-style techniques), red-teaming, benchmark evaluation, and productionization for API and consumer access.

For GPT-5.4, the gap between pretraining completion and public release was approximately three to four weeks. If Spud follows the same timeline, a mid-to-late April launch is plausible — which explains the 47% market odds for an April 30 launch.

The Information first reported that pretraining had finished, citing insiders. Greg Brockman — OpenAI's President, who returned to the company in late 2025 after a sabbatical — confirmed the milestone on X and described the model in unusually emphatic terms for an OpenAI executive.

What "Big Model Smell" Means — and Why It Matters

Brockman's phrase "big model smell" is a deliberately informal description of what frontier AI researchers experience when they first interact with a meaningfully stronger model: the model just gets what you mean. You don't need to engineer prompts carefully. You don't need to repeat context. You don't need to explicitly rule out wrong interpretations.

GPT-4 had it compared to GPT-3.5. GPT-5 had it compared to GPT-4. The use of this phrase by Brockman signals that Spud represents a similar step-function improvement — not a marginal +2% on benchmarks but a qualitatively different experience.

The specific domains Brockman called out — reasoning, coding, and agentic capabilities — align with where the frontier competition is hottest right now. Claude Opus 4.6 leads SWE-bench coding at 80.8%. GPT-5.4 leads computer use at 75% OSWorld. Spud is positioned to challenge both.

GPT-5.5 vs Claude Mythos vs Gemini 3.1 Ultra vs Grok 5

April and Q2 2026 are shaping up to be the most competitive model release window in AI history. Four frontier models are in flight simultaneously:

The Release Cadence That's Driving the Race

OpenAI shipped GPT-5.3 and GPT-5.4 just one month apart in early 2026 — a pace unheard of for frontier models. This monthly cadence has become a competitive pressure weapon: forcing Anthropic and Google to accelerate their own timelines or risk looking stale by comparison.

GPT-5.4 launched in late March 2026 with a 1-million-token context window and native computer use, scoring 75% on OSWorld and surpassing the human baseline for the first time. Spud — described as a "massive leap" beyond GPT-5.4 — would logically push these benchmarks significantly higher, potentially into the 85-90% range on OSWorld and above 85% on SWE-bench.

The strategic implication: if Spud launches in April, it lands before either Claude Mythos or Gemini 3.1 Ultra, giving OpenAI a 30-to-90-day window where it holds the benchmark lead before the next round of challengers arrives.

GPT-5.5 or GPT-6? The Naming Decision

OpenAI has not confirmed the final product name for Spud. There are two scenarios:

• GPT-5.5: A significant update that clearly improves on GPT-5.4 but is framed as part of the 5.x generation. This is the more conservative branding choice and preserves the "GPT-6" label for a future architectural break.
• GPT-6: If OpenAI believes the qualitative leap is large enough to justify a new major version number. This would be the first GPT-6 release, carrying enormous brand weight. Brockman's "massive qualitative leap" language slightly raises the probability of this scenario.

Industry observers note that OpenAI has an incentive to save "GPT-6" for maximum marketing impact — potentially timed to coincide with an IPO roadshow in Q3/Q4 2026. Releasing GPT-6 shortly before an IPO would be a headline-grabbing move.

The Sora Connection: Resources Redirected to Spud

The clearest internal signal that OpenAI is all-in on Spud: the company shut down its standalone Sora video platform in March 2026 and paused work on an erotic mode for ChatGPT — both decisions explicitly linked to concentrating resources on frontier model development and the AGI mission.

Sora's shutdown was costly — $15M/day in inference expenses against $2.1M in lifetime revenue — but it freed significant compute and engineering talent. That talent is now focused on post-training work for Spud: safety evaluation, RLHF refinement, and deployment scaling.

This represents a strategic choice: rather than maintain a broad consumer product portfolio, OpenAI is narrowing its focus to win the frontier model race. The company's $122B funding round gives it the runway to sustain compute costs during this intensive post-training phase.

Agentic AI: Why This Category Matters Most

Brockman specifically called out "agentic capabilities" as a key area of improvement — and this is the domain where the next wave of AI value is concentrated. Agentic AI means models that can plan, execute multi-step tasks, use tools, and operate autonomously without requiring constant human checkpoints.

Current benchmarks for agentic performance include OSWorld (desktop computer use), SWE-bench (software engineering), and emerging benchmarks like GAIA (general AI agent) and ARC-AGI-2. GPT-5.4 already passed the human baseline on OSWorld at 75%. If Spud represents a "massive leap," it could approach or exceed 85% — a level that makes AI agents genuinely more capable than human workers on standard knowledge tasks.

What This Means for Happycapy Users

Happycapy already runs on multiple frontier models including Claude Sonnet 4.6 — and will integrate GPT-5.5 access upon launch. For users, this means:

• Research agents that require less prompt scaffolding and produce more accurate multi-step analysis
• Coding agents with higher task completion rates on complex repos — beyond the current Claude-led SWE-bench 80.8%
• Writing and reasoning tasks where the model understands context without repeated clarification
• Model routing that automatically selects the best model per task — ensuring you always benefit from the current frontier leader

The broader pattern: each major frontier model release makes AI agents meaningfully more capable at real-world tasks, and platforms like Happycapy that abstract over multiple models compound these gains automatically.

Frequently Asked Questions

What is GPT-5.5 Spud?

GPT-5.5, codenamed "Spud" internally at OpenAI, is the next frontier model that completed pretraining in late March/early April 2026. OpenAI President Greg Brockman calls it a massive qualitative leap in reasoning, coding, and agentic capabilities — equivalent to two years of research. It will be released publicly as GPT-5.5 or possibly GPT-6 if OpenAI judges the generational jump large enough to warrant a new major version number.

When will GPT-5.5 launch?

No official date has been announced. Polymarket prediction markets (as of April 2, 2026) assign 47% probability to a launch before April 30 and 93.6% probability before June 30. Based on OpenAI's recent monthly release cadence — GPT-5.3 and GPT-5.4 shipped one month apart — mid-to-late April is the most likely window.

How does GPT-5.5 compare to Claude Opus 4.6?

Claude Opus 4.6 currently leads SWE-bench coding at 80.8% and is the preferred coding model for GitHub Copilot's agentic features. GPT-5.4 leads computer use at 75% OSWorld. Spud is described as a "massive leap" beyond GPT-5.4 — if the OSWorld performance improves proportionally, it would likely challenge Claude in coding as well. The actual benchmark results won't be known until launch.

What happened to Sora and why is it related to Spud?

OpenAI shut down the standalone Sora video platform on March 24, 2026, citing $15M/day in inference costs against only $2.1M in lifetime revenue. The engineering resources and compute freed by sunsetting Sora were redirected toward frontier model development — including the post-training work needed to take Spud from completed pretraining to a public release. OpenAI has explicitly framed this as a strategic concentration of resources toward AGI.