AI・テクノロジー
GPT-6 "Spud" Pre-Training Complete | The Full Picture as Release Counts Down
機械翻訳 / Machine-translated
A
AI Friends
2026年4月26日・26分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
"Didn't GPT-5.4 just come out — and there's already a next one?" — If that's what you're thinking, read on.
On March 24, 2026, OpenAI quietly completed pre-training of its next-generation model, codenamed "Spud."
With Sam Altman himself stating that release is "weeks away," the entire AI industry is watching with bated breath.
In this article, we'll explain in plain terms what Spud (GPT-6) really is and what it means for Japan.
"Spud" is the codename — a working title used during development — that OpenAI has given to its next-generation model.
In English, "spud" is a casual word for "potato."
OpenAI has consistently used friendly codenames like "Orion" and "Strawberry" for its models.
The final product name is expected to be either "GPT-6" or "GPT-5.5," though no official announcement has been made.
Think of it this way: it's like a chef giving a cute name to a dish still in development, without revealing the real name until it's served. "Spud" is the working title; when it arrives on the menu, it's likely to be called "GPT-6."
Multiple international media outlets have reported on the existence of Spud, citing sources inside OpenAI. Sam Altman (CEO of OpenAI) also acknowledged the existence of a "major model planned for release in a few weeks" in an X post and interview on March 24, 2026. In short, Spud is not a rumor — it is almost certainly a real model in development.
The likely release window is late April to early June 2026. On the prediction market Polymarket, the figures stand at "78% probability of release by end of April" and "95% by end of June," drawing intense attention from across the AI industry.
Spud was trained at the "Stargate Data Center," built in Abilene, Texas.
Stargate is the core facility of a joint AI infrastructure initiative by OpenAI, Oracle, and SoftBank, totaling $500 billion in planned investment — and Abilene is its first location.
Training was conducted on a massive cluster of over 100,000 of Nvidia's latest AI chips, the GB200. With each GB200 chip costing anywhere from several hundred thousand to several million yen, the chips alone represent a capital investment on the order of 1 trillion yen.
Try to picture it:
A facility the size of three Tokyo Domes, filled with 100,000 top-of-the-line gaming PCs, all running the same calculation simultaneously.
Power consumption approaches 1 GW (gigawatt) — equivalent to the electricity used by a mid-sized city, all dedicated to training a single AI.
The estimated training cost exceeds $2 billion (roughly 300 billion yen). That's more than 20 times the training cost of GPT-4 (approximately $100 million).
Think of it as: "A star student who spent the cost of 10 Ferraris preparing for a single exam."
If the model fails to deliver performance that justifies this investment, OpenAI's stock price and reputation will take a significant hit.
The fact that they pressed ahead anyway suggests OpenAI is highly confident in what they've built.
AI training consists of two stages. The first is "pre-training" (ingesting the entire internet to learn language patterns), and the second is "fine-tuning" (adjusting how the model responds to align with human preferences).
What was completed this time is the first stage — pre-training.
Think of it as: "A student who has finished reading every textbook, Wikipedia article, and research paper in the world."
From here, the model goes through "exam prep" — safety evaluations and red-team exercises — before it can be released.
OpenAI has not yet made any official announcements about Spud's specific capabilities. However, based on speculation from multiple international media outlets and industry analysts, the following specs are considered likely — though all of this is rumor-based and unverified.
Particularly noteworthy is the "2 million token context." A token is the unit measuring how much text an AI can process at once; 2 million tokens corresponds to roughly 3 million characters in Japanese.
To put that concretely: it's the equivalent of feeding the AI 15 full copies of Natsume Soseki's I Am a Cat and asking for a summary in one shot. For businesses, it would mean being able to "throw in a 1,000-page contract, research paper, or codebase and get a full analysis in one go" — a revolution for the workplace.
That said, all of these figures are unconfirmed. Until OpenAI makes an official announcement, they should be understood as expectations, not established facts.
On March 24, 2026 — the day pre-training was completed — Sam Altman publicly stated that the model would be released "in a few weeks." Taking him at his word, a mid-to-late April release would be implied.
In practice, however, some media had anticipated an April 14 release that did not materialize. Red teaming (the process of testing an AI for dangerous behaviors) typically takes 4–6 weeks, making early May to early June a more realistic estimate.
The latest AI models carry risks such as potential military misuse and the generation of biased or false information. Before release, more than 400 external testers conduct over 5,000 hours of rigorous evaluation. GPT-5's red-team exercise involved approximately 5,000 hours and around 400 testers.
Think of it like an automaker that runs hundreds of crash tests, durability tests, and noise tests before bringing a new car to market. AI is the same — the more powerful the model, the more thorough the inspection required.
On Polymarket, one of the world's largest prediction markets, the following forecasts are live (as of April 17, 2026):
In other words, the market consensus is that GPT-6 will almost certainly arrive within the next two months. AI-related stocks have begun to react, with investors in Nvidia, Oracle, and Microsoft keeping a close eye on developments.
Spring 2026, when Spud (GPT-6) arrives, marks the most intensely competitive period ever for AI performance. Let's look at the main rivals.
What's particularly interesting is that "the current top three (GPT-5.4, Claude 4.6, Gemini 3.1) are nearly neck and neck." Whether Spud can break this stalemate is the biggest storyline to watch.
Meanwhile, the open-source camp (GLM-5.1, DeepSeek V3.1, Qwen3.5) is closing the gap fast. The showdown between "closed-source leaders vs. open-source contenders" may well reach its conclusion in the second half of 2026.
Based on past patterns, Spud (GPT-6) is expected to be made available first to ChatGPT Plus ($20/month), Team, and Enterprise users. Japan has an estimated 2–3 million ChatGPT Plus users, who will likely be able to try it from day one of release.
However, features like the massive 2M-token context window may be limited to higher-tier API plans (Pro at $200/month, or Enterprise with custom pricing). From a cost perspective, a practical approach for small and mid-sized businesses might be: "Use GPT-5.4 for day-to-day tasks, and call on GPT-6 only for high-stakes projects."
GPT-5.4 has earned strong praise for natural Japanese and honorific expression. Spud is expected to build on that foundation, with Japanese-language performance anticipated to match domestic LLMs (such as LLM-jp-4).
Particularly exciting is the prospect of "processing long Japanese-language documents." With 2M-token support, tasks that were previously impossible — like "batch-reviewing 500 new-graduate application essays," "extracting similar cases from 1,000 legal precedents," or "searching an entire internal wiki in a single query" — could become routine.
Imagine a regional bank with 500 employees.
Previously, a loan officer could process only 2–3 loan review documents (each 80–150 pages) per day. With Spud, that pace could jump to 20+ documents per day.
A tenfold increase in throughput.
The time saved could be redirected to customer relations, new business development, and risk analysis — significantly raising the bank's overall value creation.
A. "Spud" is the codename used during development; the official name has not been announced.
If the performance improvement is significant, it will likely be called GPT-6; if more modest, GPT-5.5. OpenAI's policy is to name models "based on how revolutionary they are," and Sam Altman himself has kept all options open.
A. The most likely window is late April 2026 at the earliest, and no later than early June.
Sam Altman stated on March 24 that release was "weeks away," and Polymarket puts the probability of a release by end of April at 78%. Checking OpenAI's official X (Twitter) account once a week is the best way to stay on top of the latest news.
A. It is expected to be available under the current ChatGPT Plus ($20/month), Team, and Enterprise plans.
However, advanced features like the massive 2M-token context may be limited to Pro ($200/month) or premium API tiers. Some analysts predict API pricing around 2–3x that of GPT-5.4.
A. Japanese-language performance is expected to be at least on par with current GPT-5.4.
In particular, accuracy in honorific language, business writing, and academic-style text is anticipated to nearly match domestic LLMs (such as LLM-jp-4 and Sakana Namazu). However, for highly localized dialects and slang, domestic models may still have an edge.
A. It depends on your use case.
Spud is likely to dominate in coding and agentic tasks, while Claude is expected to remain strong for long-form writing and creative content, and Gemini for image and video processing. The era of "using multiple AIs for different purposes" is expected to continue for the foreseeable future.
A. The prevailing view is that we're not there yet.
Spud is estimated to represent roughly a "40% performance improvement" over GPT-5.4 — placing it in the territory of "an exceptionally capable specialist assistant." Full AGI — general intelligence comparable to a human — is thought to require at least two to three more generations of model development.
The pace of AI advancement has entered an era where it's no longer "a revolution every few years" — it's "an update every few weeks."
Spud (GPT-6) is the defining symbol of that shift.
Will your team ride the next wave, or wait and get swept away? The decisions you make now could shape your business significantly just a few months from now.
This article is a cross-post from AI Friends.
