OpenAI Triples Research Speed | The Era of AI Researchers Spending $600 a Day
機械翻訳 / Machine-translated
A
AI Friends
2日前・12分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
What you'll learn from this article
On September 6, 2026, OpenAI made an important announcement: it had achieved the goal it set in the fall of 2025 to "realize an automated research intern by September 2026."
According to the announcement, within OpenAI's research organization, AI agents are now performing 3.1 days' worth of research tasks for every single day a human researcher works. In other words, one AI is effectively doing the work of roughly three people for every one human.
This figure signals that a major shift is taking place in the world of AI research. Researchers are no longer advancing their work alone — they are using multiple AI assistants to accelerate their research.
There is another point worth noting in OpenAI's announcement: as of August 2026, the median researcher is spending more than $600 per day on AI inference.
At roughly 90,000 yen per day (based on a rate of 150 yen to the dollar), researchers are investing that amount every day in AI computing. This "inference cost" refers to the computational expenses incurred when asking an AI a question, having it write code, or having it analyze experimental results. Fees accumulate each time OpenAI's API is used, adding up to more than $600 a day.
In short, for cutting-edge researchers, AI is no longer an "occasionally useful convenience" — it has become an "indispensable daily partner used to its full capacity."
So what exactly is the "automated research intern" that OpenAI claims to have achieved?
According to OpenAI's explanation, it is "a system capable of executing clearly defined research tasks under human direction." For example, tasks that would take a skilled human researcher several days can now be delegated to AI.
One important caveat: this is not "an autonomous scientist that independently selects its own research agenda." It is a supervised system in which a human gives a clear instruction — "please investigate this problem" — and the AI executes it and returns the results.
In other words, AI does not decide on its own what to research; rather, it plays the role of determining how to solve problems that humans have defined. It is similar to the relationship between an intern at a university or company and a senior researcher who assigns tasks and provides direction.
OpenAI's announcement also includes concrete data illustrating changes in the research landscape.
Throughout 2026, the number of experiments per active experimenter has been increasing, with August 2026 reaching a record high since January 2025. This means the number of experiments a single researcher can conduct has grown dramatically.
It was also reported that as of mid-August 2026, the median researcher in OpenAI's organization uses a coding agent every day. A coding agent is an AI that automatically writes and revises code.
For context, let's also look at what is happening outside of OpenAI. At rival company Anthropic, it was reported that as of May 2026, AI — specifically "Claude" — was writing more than 80% of the code that was merged (accepted into the codebase). In other words, AI is responsible for the bulk of software development there.
In this announcement, OpenAI also revealed its next goal: to realize an "automated AI researcher" by March 2028.
What is the difference between an "automated research intern" and an "automated AI researcher"? While an intern requires human direction, a researcher-level system is expected to more autonomously identify problems, form hypotheses, design experiments, and interpret results.
That said, experts hold cautious views as well. There are concerns that AI's ability to generate effective research ideas remains a bottleneck. Even when AI-generated ideas appear convincing at first glance, human researchers often find them to be ineffective when actually put to the test.
Nevertheless, leading AI labs are beginning to automate a large portion of their research and engineering operations. The effective "workforce" of each frontier lab (organizations conducting cutting-edge AI research) is predicted to grow from thousands to tens of thousands, and potentially even hundreds of thousands, within the next one to two years.
OpenAI's announcement carries significant implications for Japan's research community as well.
First, the gap in AI utilization could directly translate into a gap in research competitiveness. Environments where researchers can invest $600 a day in AI inference versus those where they cannot could see research speed differences of more than threefold.
Second, the skillset required of researchers is changing. In addition to traditional domain expertise, the ability to effectively leverage AI agents will become essential. The capacity to judge "what to delegate to AI and what decisions humans should make" will be critical.
Third, the way research budgets are conceived will also change. Beyond personnel costs, AI inference costs will need to be counted as a significant line item in research expenditures. Japanese research institutions and companies will need to adapt to this new cost structure.
The automation of AI research is no longer a story about the future — it is an unfolding present reality. We are moving from an era where researchers use AI as a tool, to one where they collaborate with AI as a partner. We are witnessing that turning point right now.
This article is a cross-post from AI Friends.
