In an interview published in April 2026, Jakub Pachocki, OpenAI’s Chief Scientist, made a statement that went almost unnoticed amid the avalanche of predictions about artificial intelligence: inside the company, Codex was already being used for most of the actual programming. He was not talking about auto-completing a few lines of code or occasionally using a chatbot, but about a change significant enough, he said, to have altered the way most people at OpenAI approached programming.

The statement is even more interesting because the interview already belongs to a different stage of the AI race. Pachocki was speaking primarily about GPT-5.2; four months later, OpenAI had reached GPT-5.6. That distance makes the conversation more useful, however: not simply for what the company’s Chief Scientist predicted, but for what was already beginning to happen inside OpenAI before the same changes became visible in products offered to the public.

Codex appears in this discussion not so much as a tool for programmers, but as a prototype for a more general form of AI-agent-assisted work. Asked whether every industry would need to build its own specialized systems—for legal work, finance, healthcare, or other fields—Pachocki argues that the infrastructure built around the model should not remain a limitation for long. Codex, he points out, was already useful for tasks that went beyond programming. His vision is that AI will eventually work directly within the environments where people perform their jobs, connected to tools such as Slack, organizational files and data, and the context it needs to understand what it is supposed to do.

This observation also changes one of the central questions for companies trying to develop an AI strategy. In recent years, many of them have started from the assumption that their competitive advantage would come from adapting or training their own model. Jakob Effron explicitly asks Pachocki whether a company with data, examples, and good evaluation criteria should invest in its own reinforcement learning process—the method through which a model is trained according to the results it achieves.

Pachocki’s answer is more cautious than the industry might expect. Reinforcement learning can be highly effective, he says, but there is an even more efficient way to adapt a model: learning directly from context. Instead of modifying the model for every company or task, you can provide it with the relevant examples, rules, documents, and information precisely when it needs them. Pachocki considers it plausible that this capability could become good enough that many companies would no longer have a compelling reason to reproduce their own complex training processes.

If this evolution proves correct, a company’s advantage will not necessarily lie in owning a proprietary model, but in the quality of the context it can make available to a general-purpose model: internal data, processes, documents, decision histories, and access to applications and systems. In other words, the model could become shared infrastructure, while the difference between organizations would be determined by what the agent knows about each one and what it is allowed to do.

Pachocki also offers an important explanation for the current limitations of AI agents. When a model fails to complete a complex task, the problem is not necessarily that it is “not intelligent enough.” Sometimes it simply does not have access to all the necessary files, infrastructure, tools, or the full context of the problem. From his perspective, the evolution of reasoning models naturally moves from simply generating answers toward using tools and interacting more deeply with the environment in which they are expected to work.

This is where a major difficulty emerges: duration. For a mathematical problem, it is relatively easy to check at the end whether the result is correct. For code, there are tests. But a research project can take weeks or months, and the final outcome does not automatically tell you whether the decisions made during the first few hours were good ones. Pachocki believes the next frontier lies precisely in models’ ability to evaluate intermediate progress—to understand whether a direction is worth pursuing, whether a partial result is promising, or whether an approach should be abandoned before thousands of unnecessary steps are consumed.

This leads to one of the interview’s most ambitious claims. Pachocki says OpenAI is not very far from models that could work autonomously for several days, with less human supervision and probably using more computing resources than today’s typical systems. He defines the difference between an AI “research intern” and a fully autonomous researcher not simply by intelligence, but by how general the assigned task can be and how long the system can continue without human intervention.

Perhaps the most important part of the conversation, however, concerns what happens when these models are used for the very research that produces the next generation of AI. Pachocki says OpenAI had already reached a point where the performance of models available in the short term was beginning to influence the speed of long-term research, because the models were contributing to researchers’ work. He even says that GPT-5.2 Pro had generated some research ideas that he considered small but genuinely impactful. He does not provide sufficiently precise examples for their significance to be independently verified, but the statement is relevant because of who made it: the person leading OpenAI’s scientific research.

This feedback loop is far more important than simply automating programming. If a better model helps researchers build the next model faster, then each improvement can begin accelerating the process that produces the next improvement. It is not yet “recursive self-improvement” in the spectacular sense sometimes invoked in discussions about AGI—humans still set objectives, verify results, and direct the process—but it does represent the beginning of a structural change in how AI research is conducted.

The interview also contains an interesting admission about the competition with Anthropic. Asked about the success of Claude Code, Pachocki acknowledges that programming had been a secondary product priority at OpenAI, even though it had greater importance within the research organization. The enormous success of ChatGPT had focused the company on a particular type of product, while its research teams were already working on capabilities pointing toward a later stage. Anthropic identified the commercial opportunity in coding agents more quickly.

This is significant because it shows how difficult it is to predict the concrete form a technology will ultimately take, even when you are the one building it. Pachocki admits in the same interview that even the success of ChatGPT had not been obvious to him: he had expected generative visual applications, including video, to take off before text-based conversation.

Viewed from August 2026, the interview is therefore more valuable as a document about the transformation of work than as a simple collection of predictions about AGI. At the moment Pachocki was talking about GPT-5.2, OpenAI was already using Codex for most of its internal programming, attempting to extend agents beyond code, and beginning to reorganize research around the idea that models could participate in building their own successors.

The important question is no longer whether AI can write code. Inside OpenAI, that threshold had already been crossed. The question is how much of a project an agent can take over, how long it can work without supervision, and what happens to an organization when the role of humans begins shifting from execution toward setting direction, evaluating results, and making the decisions we are still unwilling to delegate.