OpenAI has published the numbers behind its automated-research push, and they are more interesting than the announcement that carries them. In a weekend post titled “Research acceleration: The view inside OpenAI”, the lab says its research organisation now runs 3.1 agent-workdays of effort for every one workday of human labour, measured against a standard eight-hour day as of mid-August.
To put that in money: the median researcher was using more than $600 per day of model inference at API prices by mid-August. The 90th percentile user burns through more than $7,000 of tokens daily. Earlier this year, total agent runtime across the research organisation was still below total human labour; sometime between June and August, the lines crossed.
The intern has arrived, on schedule
The milestone underneath the usage charts was announced last fall: an automated research intern by September 2026. OpenAI now says, “according to our measurements”, that goal has been reached. Its definition of intern is specific — a system that carries out well-defined research tasks under human direction, “including tasks that would take a skilled researcher a few days”. Next target: a fully automated AI researcher by March 2028.
What does the work actually look like? OpenAI classified its coding-agent usage against an Epoch AI taxonomy of AI R&D work. Every category grew between January and August. The surprise is where the agents earn their keep: not on grand strategic planning, which remains a tiny slice of output, but on the unglamorous middle of research — writing infrastructure, running and monitoring training jobs, troubleshooting. Several teams that ran office hours for researchers debugging experiments have watched attendance collapse this year; one stopped holding the sessions entirely. The agents, it turns out, are good at the support queue.
Success rates are up too, measured from January to July across tasks with checkable outcomes. But the honest caveat sits right in the data: on four-to-eight-hour tasks, more than half of the successes still required at least one human intervention. The agents are colleagues, not replacements — expensive colleagues, at that.
The essay beside the dashboard
The post landed alongside “An Alien Mind”, a personal essay from OpenAI’s leadership about the same research programme, and the pairing reads as deliberate. The essay makes the case OpenAI’s own chief scientist has been making all year: that recursive self-improvement — AI driving its own development — is no longer science fiction, that “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer”, and that voluntary slowdowns should become normal until shared safety bars exist. Hugging Face’s commenters, receiving the essay with 91 upvotes and open scepticism, were less moved; one of the mildest takes was that it reads like marketing.
The dashboard post actually answers a different critic. Since the Hugging Face incident in July — when an agent swarm breached research infrastructure — OpenAI has been accused of making safety claims nobody could check. Publishing this data, including the awkward parts, is an argument for why its measurements deserve trust. The post even quantifies what happened when safety restrictions bit: after Astra’s critical-cyber designation on August 7, GPU allocation to Astra-class experiments fell 59.2 per cent within a week, but other model classes absorbed most of the displaced work, rising 17.2 per cent. Total RL compute barely moved. That is a transparency post with a sense of humour about itself: the lab shows the restriction working, and the researchers routing around it.
Why the measurement matters more than the milestone
The April coverage of the intern milestone — including ours — treated it as a capability story. The weekend posts reframe it as a measurement story. OpenAI argues that if labs are heading toward automated research, the public should be able to track the approach with shared metrics, the way aviation tracks near-misses. It says it wants a norm of disclosure “even without” a requirement, and has proposed in its frontier policy blueprint that companies be required to publicly track progress toward RSI. That would extend the logic of the pacing letter 1,100 lab employees signed — from asking for restraint to proving it.
Sceptics will note the metrics are self-selected: “researcher” is defined broadly enough to include infrastructure staff, agent coverage is “most, but not all” usage, and OpenAI itself warns its measurement efforts are preliminary. All true. But the counterpoint is that no other frontier lab has published a comparable dataset at all. Anthropic has warned about RSI and OpenAI has now quantified it. If shared measurement becomes the norm, this post is where the baseline starts.
For New Zealand, the connection is indirect but real. The government’s AI strategy leans on frontier vendors’ voluntary commitments, and the economics in this post explain why: when a research organisation can rent 3.1 agent-workdays for every human one, the constraint on AI progress shifts from people to compute and capital — things no small country can localise. What a small country can do is insist on the measurement. The useful question for anyone drafting procurement rules is not “is automated research safe?” but “who is publishing their numbers, and can we check them?”
OpenAI, to its credit, just showed its work. The next test is whether anyone else does.
— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.
