OpenAI has put a date on one of its most ambitious research goals: a fully autonomous AI researcher by March 2028. The target was presented as part of a broader reorganization of the company’s internal structure, alongside new details on infrastructure, safety, products, and future models.
CEO Sam Altman and Chief Scientist Jakub Pachocki framed the plan as a shift in how OpenAI describes artificial general intelligence, or AGI. Rather than presenting AGI as a single all-knowing system, the company is emphasizing tools that help people shape what comes next.
A Roadmap Built Around Research
Pachocki laid out a staged plan for AI research capabilities. The first major milestone is an AI system with “research intern” capabilities by September 2026. That system is intended to significantly accelerate human scientific research.
The larger goal is scheduled for March 2028: a fully autonomous AI researcher capable of running independent research projects. In practical terms, OpenAI is describing a move from AI that assists researchers toward AI that can carry out research work on its own.
The source article says Pachocki pointed to continued scaling of deep learning systems as the theoretical foundation for this push. One concept he highlighted was "in-context compute", described as computing capacity at runtime that extends a model’s reasoning process.
That detail matters because the roadmap is not only about bigger models in the abstract. It is also about how much computation can be used while a model is working through a problem, and how that runtime capacity might support more extended reasoning.
AGI as a Product, Not Just a Research Target
Altman described OpenAI’s work as resting on three pillars: research, product, and infrastructure. The company’s argument is that all three are needed if AGI is going to be useful in practice.
The product goal is a personal AGI that can assist people at work and in everyday life. The source names ChatGPT, a proprietary browser called Atlas, and future devices as examples of how OpenAI wants AI to be accessible in more places.
This framing links the autonomous AI researcher to a wider product strategy. Research progress is not being described as separate from consumer and enterprise tools. It is being positioned as part of a pipeline that connects model capability, everyday interfaces, and the compute required to run them.
For readers, the key point is simple: OpenAI is not only talking about a lab system. It is also describing an ecosystem in which advanced AI capabilities are expected to move into products, tools, and devices.
The Safety Architecture OpenAI Described
Pachocki also outlined a five-layer safety model. The layers range from value alignment, which concerns what the AI aims to achieve, to goal alignment, reliability, adversarial robustness, and systemic safety.
That list shows that OpenAI is treating safety as more than a single technical check. The model spans what the system wants, how it pursues goals, whether it behaves reliably, how it holds up under adversarial pressure, and how it affects wider systems.
A central new research area is Chain-of-Thought Faithfulness. According to the source, this method allows parts of a model’s internal reasoning to remain unsupervised. Pachocki said the approach could reduce negative effects of over-monitoring during training and offer more accurate insights into how models think.
The implication is that OpenAI is trying to balance observation with the risk that too much monitoring changes the behavior being observed. The source does not provide results for this approach, but it does identify it as a major part of the company’s safety thinking.
What OpenAI Said About GPT-4.5 and Future Models
Altman and Pachocki also addressed current and upcoming language models. Many users have asked about the future of GPT-4.5, which the source says is often regarded as the best model for writing tasks.
Altman said GPT-4.5 will remain available until a clearly superior model is ready. He also said next year’s updates will bring “much better” models, especially for writing and creative work.
He also pushed back on the idea that OpenAI is “hiding” internal models. According to the source, Altman said there are prototypes and modular components being tested, but no complete internal system significantly more advanced than what has been publicly released.
Pachocki added that multiple ongoing research projects are expected to merge in the coming months into noticeable performance improvements across the product lineup. The source does not specify which products will see those changes first.
The Compute Buildout Behind the Plan
The roadmap is tied to a large infrastructure strategy. Altman said OpenAI has committed to about $1.4 trillion in infrastructure spending, spanning roughly 30 gigawatts of compute capacity to run future generations of AI models.
The partner list named in the source includes AMD, Broadcom, Google, Microsoft, Nvidia, Oracle, and Softbank. OpenAI is also building modular Stargate data centers, with one already under construction in Abilene, Texas.
The company is exploring new semiconductor manufacturing facilities as well. In the long term, Altman said OpenAI wants to essentially create an infrastructure factory capable of producing one gigawatt of computing power each week, while reducing the estimated cost to about $20 billion per gigawatt over five years.
That infrastructure ambition helps explain why the company’s business model matters. Altman reiterated that OpenAI has no plans to introduce an advertising-based business model. Instead, revenue is expected to come from subscriptions, hardware sales, and enterprise solutions.
He also said the company will need hundreds of billions of dollars in annual revenue over the long run to sustain its infrastructure ambitions. That makes the AI researcher roadmap part of a much larger commercial and technical bet.
OpenAI is also pursuing joint safety standards and transparency initiatives with Anthropic, Google Deepmind, and xAI. Existing partnerships already include shared research on neural network interpretability.
Taken together, the presentation described a company trying to align research milestones, product deployment, safety work, and compute expansion on a single timeline. The March 2028 autonomous AI researcher target is the clearest date in that plan, but it depends on much more than model capability alone.