OpenAI says its research organization now runs 3.1 agent-workdays for every human workday. It's a striking number, and OpenAI's September 6 research update is careful about what it means: a lot more work is running through coding agents. It doesn't say research got 3.1 times faster, and you shouldn't read it that way either.
If you're deciding how much to spend on agents, that gap is the whole question. Buying more execution capacity is easy. Knowing which of the extra work was worth doing is the hard part.
What OpenAI actually measured
The update describes rising agent use and more experiments. According to The Rundown's summary, OpenAI puts experiments per researcher "at an all-time high." Humans still set the priorities and make the judgment calls, and more than half of the successful tasks estimated at four to eight hours needed at least one human intervention. OpenAI also warns that activity measures don't translate directly into research progress, and notes that its available compute grew at the same time.
Those caveats belong right next to the headline. Agent runtime is an input. A successful experiment, a useful negative result or a validated improvement is an output. How one turns into the other is exactly what you need to find out, and the size of the first number can't tell you.
Measure the whole job
OpenAI uses Epoch AI's research-work taxonomy, which splits research into decisions, design, implementation, execution, analysis and communication. It's a better lens than counting conversations with a model, because work can speed up at one stage and pile up at the next.
Say your team can now launch twenty experiments in the time it used to launch five. If one researcher still has to inspect every result, interpretation becomes the bottleneck. If the extra runs test poorly chosen ideas, finishing them proves nothing. If they kill a dead end earlier, the benefit is real even though nothing new ships.
For a small team, start with one recurring job and track four things: time to an accepted result, review time, corrections and execution cost. Compare similar work across several runs, and keep the failed attempts in the record. Otherwise a workflow that looks efficient may simply have moved its expensive steps out of view.
A customer number to compare
In a separate September 8 case study published by OpenAI, 1Password reports a 20.9% productivity improvement in its Codex cohort and a 10.9% drop in median pull-request cycle time. Different company, different measures, so you can't line it up against agent runtime inside a research lab. It's also vendor-published, not an independent trial.
What it gets right is the kind of outcome it names. A faster review-and-merge cycle is much closer to delivered work than token consumption. Even so, ask how productivity was defined, how the cohort was picked and whether quality held up.
Our read: agents are expanding how much work one person can attempt. The next management problem is choosing and checking that work. Budget for the whole loop, including the person who decides whether each result deserves to survive.
