OpenAI throws its core resources behind a fully autonomous AI researcher

OpenAI is changing its research focus, and its main bet is no longer the next chatbot but an autonomous AI researcher. The company wants a system that can work through complex problems for weeks with minimal human involvement and eventually grow into a full-fledged "laboratory in a data center."

A new priority

According to OpenAI chief scientist Jakub Pachocki, the project is becoming a shared goal inside the company, one that reasoning models, agent systems and interpretability tools are meant to converge on. The nearest milestone is to show an "autonomous AI intern" by September 2026: a system that can be delegated a limited set of research tasks lasting several days. The next stage is set for 2028, a fully automated multi-agent system able to take on problems too large or too tangled for a single team of humans.

For OpenAI this is not just an attractive research direction. The company no longer sets the pace of the market alone, as it did in the days of the early GPT models, and it now competes with Anthropic and Google DeepMind not only on model quality but also on what the product stack around those models turns into. The bet on an AI researcher looks like an attempt to move beyond the chat and coding assistant and build the next layer of automation: a system that does not just answer a request but carries out long work on its own, from framing the task to an intermediate result.

"Essentially, you'll have a whole research lab in a data center,"

Pachocki says.

What is already ready

OpenAI believes an early prototype of this approach is already visible in Codex. This agent tool can write and run code, analyze documents, build charts, compile digests and handle long technical chains that people used to work through by hand. Pachocki says outright that the company wants to take what has worked in programming and carry that mode of work over to a broader class of tasks. The logic is simple: if an agent can autonomously close out substantial chunks of engineering work, the same approach can be extended to science and applied research.

OpenAI expects the future system to offer not one "magic" feature but a set of research skills that can be chained together:

• parse large bodies of text, code and notes
• propose hypotheses and ways to test them
• split a large problem into subtasks and run them in parallel
• look for solutions in mathematics, physics, biology, chemistry and government-policy scenarios

According to OpenAI, the technical foundation rests on two developments. First, general-purpose models have become better at holding a goal and context for longer: the move from GPT-3 to GPT-4 already showed a noticeable gain in "stamina" on a single task. Second, reasoning models, which work through a problem step by step and can back up after a mistake, are better suited to long autonomous cycles. On top of that, models are trained specifically on hard math and coding problems so they learn to hold long context and manage several subtasks at once.

Where the weak points are

The most obvious question is not ambition but reliability. Independent researchers acknowledge that the idea of an autonomous AI scientist looks like a logical extension of the success of programming agents, but they warn that once a system has to perform not one step but a long chain of actions, the probability of error at each stage begins to accumulate. In tests by the Allen Institute for AI, strong models already showed a good level on scientific tasks but still made mistakes regularly. Even if recent versions have improved, the problem does not go away: producing a strong idea is one thing, reliably carrying a multi-day study through to a correct result is another.

There is also a harsher layer of risk. Pachocki says OpenAI is betting on chain-of-thought monitoring: the model leaves a kind of working notes as it solves a problem, while other models or researchers watch for dangerous or simply incorrect behavior. But he himself admits that full control does not exist yet. An autonomous system may misunderstand an instruction, step outside the bounds of its task, become the target of an attack or get access to tools that are too powerful. For that reason, in his view, truly powerful versions of such agents will have to be kept in isolated sandbox environments and restricted until oversight mechanisms become much more reliable.

What this means

If OpenAI delivers even part of this plan, the next leap in AI will come not from chattier chatbots but from systems that can be delegated pieces of real intellectual work lasting days and weeks. But this is exactly where the main test will take place: the market readily believes in fast progress by agents, while science and safety will demand not a demo effect but reproducible results under human control.