A research assistant that produces more experiments can still leave its supervisor with more work. The useful question is what survives review.
OpenAI says it has reached its automated research intern milestone. Its September 6 internal report describes systems completing defined assignments under human direction, including work that could take a skilled researcher several days. This is a company-reported result, not an independent evaluation or a separately announced product you can buy.
What the report establishes
OpenAI reports growing agent use, code contributions and experiments. It still assigns research priorities, interpretation and scaling decisions to people. The company presents March 2028 as its target for a more capable automated researcher, not a milestone already achieved.
The report values mid-August median researcher usage above $600 per day at API prices, with the 90th percentile above $7,000. Those figures are not a disclosure of actual internal cash expense. OpenAI also cautions that activity measures do not translate directly into the overall pace of research. No independent replication was checked for this article.
Separate an assignment from a research program
My reading is that the valuable operating unit is the assignment. A request to reproduce one result has an end condition. A request to improve a research program can keep expanding: new data, new evaluation criteria, more compute and more dependencies. Treating those requests as interchangeable makes both oversight and costing harder.
For a small team, I would begin with an analysis whose answer can be checked without trusting the same agent that produced it. Reproduce a published calculation from supplied data, compare two implementations against a fixed test suite, or explain why a known experiment failed. Keep choosing the next research direction outside that first trial.
Use this delegation worksheet
| Field | Write down before starting | Acceptance evidence |
|---|---|---|
| Question | One falsifiable question and the baseline to compare against. | An answer that addresses the original question, including a negative result. |
| Inputs | Approved files, versions and permitted data sources. | A manifest connecting each result to its inputs. |
| Boundaries | Allowed tools, spend ceiling and actions requiring permission. | A record of tool use and any requested exceptions. |
| Reproduction | Exact command or procedure a reviewer will run. | A repeat run from the recorded starting state. |
| Review | Named reviewer and the expected inspection time. | Accepted, rejected or unresolved, with a reason. |
Measure the work that survives
Count all attempts in the denominator. If ten assignments produce six accepted results, reporting only those six conceals the cost of the other four. Keep rejected outputs and retries in the operating record even when they never reach a pull request or slide deck.
I would track reviewer minutes per accepted result alongside inference and compute consumption. A lower token bill can be a poor trade if verification takes longer. A larger bill can be reasonable if a reproducible answer removes a genuine bottleneck. Neither conclusion follows from usage alone.
Our AI agent cost-control guide covers budget boundaries. Use the same accounting discipline here, but make the unit a reviewed assignment rather than an agent session.
Keep a failed experiment useful
A failed run should return its inputs, logs, attempted changes and the reason the question remains unresolved. Do not reward an assistant for producing a positive-looking conclusion at any cost. A clean rejection of a hypothesis can be more useful than another plausible chart.
Before the trial, write down what must remain unchanged: the evaluation data, success threshold and baseline configuration. If the agent proposes changing one, record a new experiment instead of silently moving the target. This makes comparisons interpretable when several attempts happen in parallel.
The test I would run this week
Choose three comparable, low-consequence assignments. Estimate how long each would take without an agent, then record actual preparation, intervention and review time during the trial. Label the estimates as estimates. Ask another person to reproduce at least one accepted result without reading the assistant’s narrative first.
Use the Codex and Claude Code comparison to think about workflow fit, not as evidence that either tool will reproduce OpenAI’s internal results. This article does not include a hands-on benchmark.
My take: buy evidence, not impressive activity
I would expand access only after the team can explain which assignments became easier and where oversight remained expensive. More autonomous execution is a sensible reward for reliable, reviewable work. It should not be the starting assumption.
Read OpenAI’s report and methods appendix before reusing its figures. Which recurring assignment would you delegate first, and what evidence would make you accept the answer?
