OpenAI: 3.1 agent-workdays per human workday

OpenAI published its internal numbers on how much of its own research is now done by coding agents, and the headline figure is 3.1 agent-workdays of effort for every human workday. It also says it hit the goal it set last fall: an automated research intern by September 2026. The number worth copying is not the 3.1. It is that OpenAI bothered to measure it at all.

What actually happened

OpenAI defines a "research intern" narrowly: a system that carries out well-defined tasks under human direction, including work that would take a skilled researcher a few days. Not autonomy. Direction plus duration. The company says it is aiming for a fully automated AI researcher by March 2028.

The operating numbers are the interesting part. As of mid-August, the median OpenAI researcher was running more than $600 a day of inference on internal coding agents at API prices. The top decile was past $7,000 a day. That is per person, on tokens alone.

And OpenAI put a caveat on its own metric that most vendors would have quietly omitted: 3.1 agent-workdays is not a 3.1x productivity gain. Agent runtime can be parallel, redundant, unsuccessful, or heavily steered by the human sitting next to it. Chief scientist Jakub Pachocki separately said he expects the current pace to carry into recursive self-improvement — a claim about the future, not a measurement, and worth filing under forecast rather than fact.

Why this matters for your business

Two things transfer directly, and neither of them is "buy more agents."

First, the ratio is a metric you can actually run. Most small companies deploying AI have no denominator. They have a Claude subscription, a Copilot seat, and a feeling. OpenAI's framing — agent effort against human effort, tracked over months — is something you can approximate this quarter from your API dashboard and your team's hours. If the ratio is climbing and cycle time is not, you have a spend problem dressed up as an adoption win.

Second, $600 a day per seat is what unbounded agent use costs. That is a frontier lab optimizing for research speed with no budget constraint, so it is a ceiling, not a target. But it kills the assumption that agent cost is a rounding error next to salary. At $600 a day, five engineers is roughly $780K a year in tokens. Set per-project budget caps and route cheap work to cheap models before that curve finds you. We wrote about the same math when Writer showed a 41% cost drop from harness changes alone.

The honest read on the milestone itself: an intern that needs a defined task and a human reviewing the output is a good intern. It is not a replacement for the person assigning the work.

Sources: OpenAI, Research acceleration: The view inside OpenAI, Help Net Security.