OpenAI's 80-90% GPT-7/GPT-8 Research Claim, Explained
An OpenAI exec said 80-90% of research targets GPT-7, GPT-8, and beyond. Here's what that claim does and doesn't mean for future releases.
What did the OpenAI executive actually say?
At the Fellows Forum AI Summit on September 23rd, OpenAI’s head of applied research, Boris Power, estimated that 80 to 90% of the company’s research is focused on GPT-7, GPT-8, and beyond. That’s the full claim: an estimate, offered on stage, by one executive. It is not a leaked internal document, not a budget breakdown, and not a roadmap with dates attached. The number spread fast online because it sounds like confirmation that OpenAI is quietly racing toward superintelligent models while the public is still getting used to GPT-6 Astra. The real meaning is narrower and more interesting than that.
TL;DR
- Boris Power’s estimate came from a single stage appearance, not a published OpenAI research budget, so it should be read as a rough gut-check from a company insider, not an audited fact.
- “Research” is not the same as “training a model right now.” It covers training methods, evaluation systems, data pipelines, safety checks, chip and networking reliability, and small experiments that may never ship as a public product.
- GPT-6 Astra already crossed a notable safety threshold, with OpenAI’s own safety report naming it the first model to reach the company’s critical level for cyber security capability, which raises the stakes for whatever comes next.
- Frontier labs plan years ahead by necessity, because infrastructure like OpenAI’s Stargate program, which has secured more than 10 gigawatts of planned U.S. AI capacity, takes far longer to build than any single model takes to train.
- This is an industry-wide pattern, not an OpenAI quirk. Google said in July it had already started its most ambitious Gemini 4 pre-training run while Gemini 3.5 Pro was still being prepared for release.
- No public specification exists for GPT-7 or GPT-8. There’s no confirmed size, capability list, or launch date for either, so specific feature claims circulating online are speculation, not reporting.
- The more useful signal to track isn’t the model name. It’s infrastructure coming online, updated safety thresholds, and research results, all of which are observable before any “GPT-8” announcement.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
What does “research focused on GPT-7 and GPT-8” actually mean?
The instinct when hearing this claim is to picture one giant training run for GPT-8 already underway in a data center. That’s not what Power described. Research aimed at future generations includes testing new training methods, building better evaluation systems, improving training data, designing safety checks, and making chips and networking more reliable. It also includes smaller experiments that may never become a shipped product at all.
A future model name, in this context, works more like a direction of travel than a finished recipe. Saying research is “focused on GPT-8” doesn’t mean GPT-8 exists in any concrete form. It means the problems researchers are solving today are aimed at capabilities beyond what the current public model, GPT-6 Astra, already does. The clip tells us about OpenAI’s time horizon, not about a release date.
Why does frontier AI research have to start years before a model ships?
Frontier labs can’t afford to wait for one model to launch before starting work on the next, because the physical and organizational pieces take too long to assemble. Hardware procurement, power contracts, data pipelines, safety evaluation frameworks, and training software all have long lead times that don’t compress just because a company wants to move faster.
OpenAI’s Stargate infrastructure program illustrates this directly. The company says it has already secured more than 10 gigawatts of planned AI capacity in the United States, and a flagship site in Texas is training GPT-5.5. That kind of commitment shows research planning and physical construction now move on the same timeline. If most researchers are already working on problems beyond the current generation, that reflects the company betting that capabilities will keep climbing across several product cycles, not that a specific next model is almost finished.
What does GPT-6 Astra tell us about where the risk conversation is heading?
GPT-6 Astra matters here because it’s the clearest evidence of how seriously OpenAI is treating the next jump in capability. OpenAI describes Astra as its most capable broadly deployed model, and its own safety report states that Astra is the first OpenAI model to reach the company’s critical level for cyber security capability. In practical terms, OpenAI says the model can help identify unknown software vulnerabilities and develop exploits for well-protected systems with less human guidance than before. That’s why the company added stricter isolation, encrypted checkpoints, and more monitoring around it.
This doesn’t mean Astra is unsafe or uncontrollable. OpenAI says it’s safer overall than its predecessor. But it does mean the public conversation about model risk has moved past jokes about chatbots giving wrong answers. If GPT-7 and GPT-8 are meant to extend capability further than Astra, the safety infrastructure built for them has to be ready in advance, not retrofitted after the fact.
Does safety research keep pace with capability research?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
OpenAI’s frontier governance framework and preparedness framework are built around the idea that severe risks (cyberattacks, chemical and biological harm, manipulation, loss of control) need evaluation systems and stopping rules in place before a powerful model is deployed, not after. That’s one real reason a large share of research effort can be aimed at future generations without that being purely a race for better benchmark scores: safety evaluation has to be designed and tested well ahead of any model it’s meant to govern.
The open question is whether that safety work can actually keep pace with how fast the infrastructure side is scaling. OpenAI has been redesigning the internal networking of its largest supercomputers so that thousands of GPUs can keep functioning even when parts of the system fail, and that work is already deployed across major training clusters. The company is also developing its own chips, with future generations already underway. All of this is built for repeated frontier runs, not a single dramatic launch, which raises the practical question of whether governance processes scale at the same rate as the hardware does.
Are AI models starting to help build the next AI models?
OpenAI is pushing its current models into software engineering, scientific research, web browsing, and computer-use tasks. That means newer models can help researchers write code, analyze experimental results, and search through data while the next generation of models is being developed. That is a real feedback loop: stronger models make research teams faster, and faster teams can test more ideas for whatever comes next.
That’s different from recursive self-improvement, the idea that a system upgrades itself without human involvement. The evidence available doesn’t support a claim that AI systems are autonomously building their successors. What it does support is a more mundane but still meaningful dynamic: human researchers increasingly use AI tools as part of how they build the next AI tools, which can compress development timelines even without anything resembling self-improving AI.
Is OpenAI the only lab planning multiple generations ahead?
No. This pattern shows up across the industry. Google said in July that it had already started its most ambitious pre-training run for Gemini 4 while Gemini 3.5 Pro was still being prepared for a broad release. Anthropic, xAI, and major Chinese labs are all simultaneously shipping current models, hiring research staff, and securing large amounts of compute. At the frontier, no major lab plans one launch at a time, because the infrastructure lead times and research timelines simply don’t allow it.
What makes Power’s 80 to 90% estimate notable isn’t that it reveals a secret plan. It’s that it puts a specific number on a pressure that’s structurally present across the entire industry, and it suggests the currently shipping product is a small visible slice of a much larger internal research program.
Is it true that GPT-8 is already finished or near release?
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
No, and this is the most important thing to get right. Nobody outside OpenAI has a verified feature list, size, or launch date for GPT-7 or GPT-8. The company hasn’t published a roadmap for either. Expectations about longer independent task execution, stronger tool use, better computer control, and deeper scientific reasoning are reasonable guesses based on where the field is heading, not confirmed specifications. Power’s quote establishes that OpenAI doesn’t see GPT-6 Astra as a finish line, even after it triggered the company’s highest cyber risk category to date. It does not establish that a more advanced model is finished or imminent.
Frequently Asked Questions
Did an OpenAI executive really say 80-90% of research targets GPT-7 and GPT-8?
Yes. Boris Power, OpenAI’s head of applied research, gave that estimate on stage at the Fellows Forum AI Summit on September 23rd. It was a spoken estimate from one executive, not a published internal report.
Does this mean GPT-8 is currently being trained?
Not necessarily. Research aimed at future generations covers training methods, evaluations, data quality, safety systems, and infrastructure reliability. It doesn’t mean a finished model called GPT-8 is mid-training right now.
What is GPT-6 Astra, and why does it matter here?
GPT-6 Astra is OpenAI’s most capable broadly deployed model as of this reporting, and OpenAI’s own safety documentation says it’s the first model to reach the company’s critical level for cyber security capability. It sets the baseline that future models would be measured against.
What is Stargate and why is it relevant?
Stargate is OpenAI’s infrastructure program, which the company says has secured more than 10 gigawatts of planned AI capacity in the U.S., including a flagship Texas site training GPT-5.5. It shows how far in advance physical infrastructure for future models has to be planned.
Is OpenAI unique in planning multiple model generations ahead?
No. Google has said it started pre-training Gemini 4 while still preparing Gemini 3.5 Pro for release. Working years ahead of the current public model is standard practice across frontier AI labs, not something specific to OpenAI.
