OpenAI's Next Model "Spud" Finished Training — What We Know Before the Release

OpenAI's next major model has a codename: Spud. It finished pretraining on approximately March 24, 2026. It has not been released. What it will be called when it ships — GPT-5.5, GPT-6, or something else entirely — has not been confirmed. What employees say about it internally has leaked. What it actually does, we do not yet know.

Here is everything confirmed, everything leaked, and everything that is genuinely unknown.

What Is Confirmed About Spud

Pretraining is complete. As of approximately March 24, 2026, Spud finished its pretraining phase. This is the most computationally expensive part of training a large language model — training on raw data at scale before any fine-tuning or alignment work begins. Completing pretraining is a major milestone but not a release event. Safety evaluation, red-teaming, supervised fine-tuning, and RLHF alignment all follow pretraining before a model is ready for public deployment.

It is a high-priority project for OpenAI. OpenAI discontinued the standalone Sora video generation app in early 2026 and reallocated resources to Spud completion. Shutting down a visible consumer product to accelerate a model is not a routine decision. It signals that Spud was behind schedule, or that OpenAI considers Spud completion more critical than Sora's consumer video market position, or both.

Sam Altman described it as a "very strong model." In internal communications, Altman characterized Spud as a model that could "really accelerate the economy." This framing — economy-level impact rather than benchmark comparisons — is the first concrete hint about the kind of capability being developed. Altman has not said this about GPT-5.2 or GPT-5.4.

Employees say it contains something "very different." Multiple OpenAI employees have used the phrase "very different from what we've seen before" when discussing Spud. The exact capability being referenced has not been specified publicly. This language is notable because it was not used for GPT-5.1 through GPT-5.4 — incremental improvements were described as improvements, not as category differences.

What Is the Naming: GPT-5.5 or GPT-6?

OpenAI has not confirmed the public-facing name. The internal codename "Spud" is meaningless outside the company. The current model family progression is: GPT-5 (August 2025) → GPT-5.2 → GPT-5.4 (current recommended API model).

If Spud is a substantial architecture change — something genuinely different rather than an incremental capability improvement — it would be named GPT-6. If it is a significant improvement but within the same architectural generation as GPT-5, it would be named GPT-5.5 or follow the existing 5.x pattern.

The "very different" language from employees is the strongest signal toward GPT-6. Historically, OpenAI uses major version increments (GPT-3 → GPT-4 → GPT-5) for models that represent architectural generation changes, not just scale increases. GPT-4 to GPT-5 introduced native multimodality and the thinking/non-thinking routing system. If Spud adds a capability that is categorically new — not just better at existing tasks — GPT-6 branding is likely.

The Terence Tao connection appearing in coverage is worth noting: Fields Medal mathematician Terence Tao was apparently involved in evaluating or advising on Spud's mathematical reasoning capabilities. Tao has previously commented on GPT-4 and GPT-5's mathematical limitations publicly. His involvement suggests math and formal reasoning is one area where Spud represents a significant leap.

What "Very Different" Probably Means

Speculative, but informed by the surrounding evidence:

Autonomous agent capability at scale. GPT-5.4 can act as an agent within bounded contexts. The jump to genuine long-horizon autonomy — running research tasks, managing files, executing code across sessions without human confirmation loops — is the most commonly cited capability gap between current models and AGI-adjacent systems. If Spud closes this gap meaningfully, it fits the "very different" description and also explains Altman's "accelerate the economy" framing.

Scientific reasoning and formal proof. Terence Tao's reported involvement, combined with the economic framing, suggests Spud may have meaningful advances in formal mathematical reasoning — the ability to generate and verify proofs rather than just pattern-match to existing solutions. This would be categorically different from GPT-5.x.

Real-time learning or retrieval integration. Current GPT models have training cutoffs and require external retrieval tools for up-to-date information. A model that integrates real-time learning natively — without the retrieval/generation split — would be architecturally different from what exists now.

None of these are confirmed. The honest answer is that the "very different" capability is unknown until OpenAI announces it.

When Will Spud Be Released?

Pretraining completed March 24. The standard timeline between pretraining completion and public release for OpenAI models has been roughly 3-6 months for major models (GPT-5 took approximately 5 months from estimated pretraining completion to August 2025 launch). For minor updates the timeline is shorter.

If Spud follows the same cadence, the release window is June–September 2026. Q2 2026 (by June 30) is possible if safety evaluations go smoothly. Q3 2026 is more likely for a model described as categorically different — those take longer to red-team.

The $122 billion funding round that closed the same week as Spud's pretraining completion is not a coincidence in timing. OpenAI raised at $852 billion on the narrative that the next generation model is done training and will ship in months. Investors closed at that valuation knowing Spud exists and is real. The IPO preparation (expected 2026-2027) creates additional pressure to ship a flagship model before filing — you do not go public with a 2025 vintage model as your flagship.

What Changes for Developers When Spud Ships

Every major OpenAI model release resets the practical ceiling for what developers can build. GPT-4's function calling enabled the current generation of AI agents. GPT-5's native multimodality enabled vision-native applications. Whatever Spud's "very different" capability is, it will unlock applications that are currently impossible or impractical to build.

For developers: the actionable move is to watch the safety card and system card that OpenAI publishes at release. The system card details specific capability gains, context window, modalities, and API specifications. The API is typically available within 24-48 hours of the consumer ChatGPT release for models in the GPT-5.x series — GPT-6 may follow the same pattern or have a separate developer API release event.

The current recommended API model is GPT-5.4. Applications built on GPT-5.4 will continue functioning when Spud releases — OpenAI maintains backward API compatibility for at least 12 months after a new model ships. There is no urgency to update before Spud is available, but plan for a migration evaluation window when the release happens.

How Spud Fits the OpenAI Roadmap

OpenAI's stated position is that it expects to reach AGI — AI that outperforms humans on most economically valuable tasks — within the current decade. The roadmap as understood externally involves o-series reasoning models (o3, o4) for specialized reasoning tasks and GPT-series models for general capability. Spud appears to be GPT-series.

The Amazon investment's AGI clause (AGI or IPO by 2028) creates a financial context for Spud's capabilities. Amazon's internal AGI evaluation criteria are not public, but the existence of the clause means Amazon has a financial opinion about when OpenAI reaches AGI. Spud completing pretraining the same week the Amazon clause was signed suggests Amazon was aware of Spud's capabilities and structured the clause around a realistic timeline.

Whether Spud meets anyone's AGI definition is unknown. What is certain is that it is the most significant model OpenAI has trained since GPT-5, it is done training, and it will ship in 2026.

Key Takeaways

• Spud is OpenAI's next major model — pretraining completed approximately March 24, 2026
• Not yet released — safety evaluation, fine-tuning, and red-teaming follow pretraining; estimated Q2-Q3 2026 public release
• OpenAI employees call it "very different" — this language was not used for GPT-5.1 through 5.4 incremental updates
• Sam Altman framing: could "really accelerate the economy" — suggests agentic or scientific reasoning capability, not benchmark improvement
• Naming TBD: GPT-5.5 or GPT-6 depending on whether the capability shift is architectural — "very different" language favors GPT-6
• Sora discontinued to reallocate resources to Spud completion — OpenAI killed a consumer product to prioritize this model
• Developer timing: watch for the system card at release; API access typically 24-48 hours after ChatGPT consumer launch; GPT-5.4 API remains supported