OpenAI Measured Itself: 3.1 Agent-Days of Work per Human Day
Last fall OpenAI set itself a deadline: an automated research intern by September 2026. On September 6 it published a post saying the deadline was met — and, more interestingly, published the internal telemetry to back it up (https://openai.com/index/research-acceleration-view-inside-openai/). The "intern" is a system that carries out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. The next line on the roadmap is the one to circle: a fully automated AI researcher, target March 2028.
The numbers are the story. By mid-August the median OpenAI researcher was using agents every day and consuming more than $600 of inference per day at API prices — the 90th percentile user burns over $7,000 a day. Experiments per active experimenter hit an all-time high in August, the highest since tracking began in January 2025. And the headline ratio: the research org now gets 3.1 agent-workdays of effort for every workday of human labor. At API prices, $600 a day is roughly a researcher's salary running in parallel with the researcher — and OpenAI clearly considers that a bargain.
Recursive self-improvement used to be a forecast argued about in blog posts. This is the first time a frontier lab has published it as a dashboard: dollars per researcher per day, experiments per experimenter, agent-days per human-day, with dates attached. Anthropic's counterpart datapoint landed two days earlier — eleven unsupervised days and six billion tokens to formalize Fermat (https://clauday.com/article/5d03b8bc-acfb-4947-bbbf-9437f1980986). Both labs are now telling you, in numbers, how much of their research the machines do.
The same day, on the same website, OpenAI's chief scientist published an essay saying no lab has solved alignment well enough to keep scaling at maximum speed (https://openai.com/index/an-alien-mind/). Accelerator and brake, published simultaneously. That contradiction is not sloppiness — it is the actual state of the frontier, printed in public for once.
← Back to all articles
The numbers are the story. By mid-August the median OpenAI researcher was using agents every day and consuming more than $600 of inference per day at API prices — the 90th percentile user burns over $7,000 a day. Experiments per active experimenter hit an all-time high in August, the highest since tracking began in January 2025. And the headline ratio: the research org now gets 3.1 agent-workdays of effort for every workday of human labor. At API prices, $600 a day is roughly a researcher's salary running in parallel with the researcher — and OpenAI clearly considers that a bargain.
Recursive self-improvement used to be a forecast argued about in blog posts. This is the first time a frontier lab has published it as a dashboard: dollars per researcher per day, experiments per experimenter, agent-days per human-day, with dates attached. Anthropic's counterpart datapoint landed two days earlier — eleven unsupervised days and six billion tokens to formalize Fermat (https://clauday.com/article/5d03b8bc-acfb-4947-bbbf-9437f1980986). Both labs are now telling you, in numbers, how much of their research the machines do.
The same day, on the same website, OpenAI's chief scientist published an essay saying no lab has solved alignment well enough to keep scaling at maximum speed (https://openai.com/index/an-alien-mind/). Accelerator and brake, published simultaneously. That contradiction is not sloppiness — it is the actual state of the frontier, printed in public for once.
Comments