OpenAI’s GPT-8 Plan Is Bigger Than It Sounds

Boris Power, OpenAI’s head of applied research, ignited a wildfire of speculation when he estimated that 80% to 90% of the company's research effort is building past GPT-8. This striking remark, made at the September 23 Fellows Forum AI Summit, describes the allocation of research toward future models like GPT-7, GPT-8, and beyond, not an independently verified budget or staffing breakdown. It surfaced as OpenAI’s GPT-6 Astra, its most capable deployed model, was just reaching users.

Power’s statement is a crucial planning horizon indicator, but it does not establish that GPT-8 is nearly complete, currently training, or scheduled for release. Research in this context encompasses a broad spectrum of activities far removed from a finished product. It includes:

• Testing new training methods
• Building evaluation systems
• Improving data pipelines
• Designing safety checks
• Enhancing chip and network reliability
• Running smaller, exploratory experiments

A future model name often signifies a direction of travel for long-term inquiry, rather than a blueprint for an imminent release. The percentage highlights OpenAI’s forward-looking strategy: most long-term research effort points past the current product generation, not that thousands of people are secretly finishing GPT-8 right now. This distinction is vital for accurately interpreting the company’s ambitious roadmap.

Power’s estimate focuses on research effort, not a single, ongoing training run for a future model. When Power suggests 80–90% of research points toward GPT-7, GPT-8, and beyond, he describes a broad spectrum of preparatory work. This includes foundational experiments and engineering that precede any full-scale pretraining.

Research can encompass many activities. Teams might be:

• Testing novel training methods
• Building robust evaluation systems
• Improving data pipelines
• Designing comprehensive safety checks
• Enhancing chip and network reliability

These efforts collectively lay the groundwork for future systems. Some experiments may directly inform later models, while others might never yield a public product. A future model name often functions as a direction of travel, guiding these diverse investigations, rather than labeling a finalized product blueprint.

Therefore, the quote offers insight into OpenAI’s long-term investigative priorities. It tells us more about the company's strategic outlook and what problems it is exploring than about a specific model’s progress, its projected capabilities, or an imminent launch date. This critical distinction separates exploratory research from product development.

Frontier AI labs must plan across overlapping model generations because the foundational requirements for future systems demand long lead times. Critical elements like compute capacity, energy grid connections, networking infrastructure, extensive data pipelines, and robust safety evaluations cannot materialize on demand; they require years of advance preparation.

OpenAI’s ambitious Stargate initiative exemplifies this forward-thinking strategy, targeting over 10 gigawatts of planned AI capacity in the United States. This infrastructure signals the company’s sustained ambition and preparation for future models, though it does not confirm the size or schedule of any specific upcoming release. A flagship site in Abilene, Texas, for example, is already training models like GPT-5.5, illustrating how tightly research roadmaps and physical construction now intertwine.

This technical groundwork is crucial for repeated, high-stakes training runs. Reliable clusters and resilient networking, often redesigned to maintain operation even when parts of the system fail, significantly reduce costly interruptions. Such infrastructure directly supports the efficiency of the research pipeline, ensuring that months-long training runs can complete cleanly. For more details on OpenAI’s broader research efforts, consult the OpenAI Research Index.

OpenAI’s own safety assessment of GPT-6 Astra describes it as the first model to reach the company’s "critical level for cybersecurity capability." This means Astra can assist in finding unknown software flaws and developing exploits for protected systems with reduced human guidance. While the company says Astra is safer overall than its predecessor and not uncontrollable, this assessment shifts the risk conversation beyond simple chatbot errors.

This escalation underscores why safety research must run in parallel with capability development, not as an afterthought. OpenAI's Frontier Governance Framework outlines tracking severe risks like cyberattacks, biological harm, and manipulation. The company's preparedness framework dictates required protections before deployment.

These crucial checks—including evaluations, secure systems, and clear stopping rules—cannot materialize once a powerful model is complete; researchers must design them during development. This integrated approach ensures safeguards scale with advancing capabilities, providing a critical counterbalance to the rapid pace of infrastructure build-out.

Readers can monitor several indicators for progress on this front:

• Published evaluations
• Updates to safety frameworks
• Research releases
• New infrastructure coming online

No verified specifications or release dates for GPT-7 or GPT-8 are public, but the trajectory demands that safety measures lead, not lag, the capabilities they aim to govern.

Did OpenAI confirm that GPT-8 is being trained?

No. Boris Power described research attention aimed at GPT-7, GPT-8, and beyond; that does not confirm an active GPT-8 training run.

What did Boris Power say about OpenAI’s research?

At the Fellows Forum AI Summit, Power estimated that 80% to 90% of OpenAI’s research focused on generations beyond its current model.

Why would OpenAI research future models years early?

Methods, data, evaluations, safety systems, computing hardware, and data-center capacity all take time to develop and coordinate.

Are GPT-7 and GPT-8 release dates public?

No. OpenAI has not published verified launch dates, specifications, or a public roadmap for either model.