OpenAI’s internal estimates show which research tasks agents were more likely to finish without recorded human intervention, and what the rise in experiments can tell us.
The one-minute summary
- The job matters. OpenAI’s midpoint estimates for success without recorded intervention were 79% for monitoring runs, 35% for launching training, reinforcement-learning or evaluation runs, and 25% for writing inference and serving code.
- Estimates rose for multi-hour tasks. From March to October 2026, the midpoint rose from 23% to 46% for tasks the analysis rated at two to four hours of human work. The mix of assignments may also have changed.
- Researchers logged more experiments per active experimenter. The four-week average ending October 18 was 1.68 times the 2025 baseline. The share attributable to agents remains an open question.
OpenAI, Research acceleration: The jagged frontier, November 18, 2026
Agents had more success monitoring runs than launching them
A researcher monitoring a training run might want to know when a metric first drifted, which log entry came just before it, and whether the same pattern appeared yesterday. Those questions give an agent evidence it can point to. Launching a run introduces more choices. A configuration has to match the experiment, the environment has to support it, and a mistake can spend compute before anyone sees a result.
OpenAI’s November report found a gap of about 44 percentage points between the midpoint estimates for monitoring and launching runs. The launch category includes training, reinforcement-learning and evaluation runs. The chart covers selected tasks in OpenAI’s research organization from August 1 through October 18, 2026. OpenAI’s task breakdown
Chart values with approximate 95% confidence intervals
- Monitoring runs for issues had an estimated success rate of 78.7% (67.2%–88.6%).
- Launching training, reinforcement-learning or evaluation runs had an estimated success rate of 34.9% (26.5%–43.6%).
- Writing inference and serving code had an estimated success rate of 25.0% (10.9%–40.2%).
OpenAI used a GPT-6 Astra agent to assess a sample of interactive coding-agent logs. The analysis excludes known automation patterns and gives each observed researcher-month equal weight. To calculate the midpoint, OpenAI assumes that half of the tasks with unclear outcomes and no recorded intervention succeeded. OpenAI’s task-scoring methods
The distinction matters when a team assigns a whole project. Imagine asking an agent to start a run, watch it and summarize the result. Good monitoring matters after the run starts, but the workflow can stall on a bad launch configuration. An agent that does useful work at one stage can still need close supervision at the stage that determines when the work finishes.
These estimates describe jobs researchers chose to delegate in their own environment. Researchers might hand off familiar monitoring jobs and reserve the hardest ones for themselves. Launch jobs can also vary in scope, tooling and difficulty. For readers deciding what to delegate, the task gap is a reason to test individual steps on their own work.
Unassisted success rose on multi-hour tasks
Kingy’s September coverage looked at OpenAI’s first report on its “automated research intern.” In the November analysis, the estimated midpoint rose from 23% to 46% between March and October for tasks rated at two to four hours of human work. For tasks rated at four to eight hours, it rose from 13% to 35%. OpenAI used a model to estimate how long a skilled person would take, and says those estimates need more validation. Kingy.ai’s September coverage · OpenAI’s November duration analysis
Agents may be getting better at the work, and researchers may be getting better at picking and framing assignments. A four-hour assignment could mean investigating one well-documented failure or untangling several poorly defined ones. OpenAI says the mix of assignments within each bracket may have shifted. It also used a later employee roster to identify researchers in the March through June data. OpenAI’s duration methods
Independence is one useful measure if you are deciding how closely to supervise an agent. For a team deciding whether to keep paying for a workflow, a second measure is more directly useful. How much human time did the accepted result take? That includes preparing the brief, answering questions, checking the result and repairing mistakes. An intervention can be cheap if it takes one clear instruction to get the job back on track. Reviewing a plausible but faulty result can take far longer.
Where researchers still shape the work
OpenAI asked contributors to the GPT-6 Astra and GPT-6.2 Astra releases where some of their most important ideas came from. A GPT-6.2 Astra pretraining researcher said they originated an idea and that it began working well after Codex suggested a change. The posttraining and safety-monitoring contributors said that, as far as they knew, people originated the ideas they named and models helped implement them. These were accounts of selected ideas from people who worked on the releases. OpenAI’s contributor accounts
The pretraining account shows how difficult it is to assign credit cleanly. The researcher described a contribution that mattered to one idea. Estimating Codex’s share of the release would require a fuller record of who proposed, tested and verified the changes that made it into the model.
A separate interview points to a workflow readers can try. One OpenAI researcher said they started asking agents to produce a single example for review before making the rest. That gives the reviewer a chance to catch a mistaken assumption before it spreads across a batch. OpenAI’s workflow interviews
What the experiment count can tell us
Logged experiments per active experimenter averaged 1.68 times the 2025 baseline in the four weeks ending October 18. OpenAI divided each week’s logged experiments by the number of distinct active experimenters it could identify, then averaged those weekly ratios. On the main platform, OpenAI excluded evaluations and automated runs and capped the number of runs it counted from each account namespace per day. The measure can rise if low-volume researchers stop running experiments, even if the people who remain run experiments at the same pace. OpenAI’s experiment-counting methods
The report also introduces OpenAI ECI, an index built from public and internal benchmark results. It gives a view of how model performance changed on the tests OpenAI included. Estimating how much agents contributed to those changes is a separate research question. OpenAI’s capability index
Suppose a team runs 17 experiments this month after running 10 last month. The extra runs might help it rule out a bad idea sooner or expose a defect that one test would miss. They could also repeat a poorly designed comparison. To judge the gain, you’d want to know what the team learned, what it changed as a result, and how much time and compute it spent reaching that decision.
OpenAI priced interactive agent usage at API-equivalent rates and calculated a 28-day daily average for each eligible researcher. For the window ending October 18, the median was about US$2,400 per day and the 90th percentile was about US$23,000. OpenAI assigned internal models the price of a comparable released model. Estimating the return would require actual operating costs and the value of completed work. OpenAI’s usage methods
Attribution gets harder at the level of a whole lab. Compute availability, staffing, model capabilities and research priorities can all shift during the same period. If an agent helps a researcher get to a decision in days instead of weeks, that matters. Quantifying its contribution across many projects would take evidence about what comparable projects achieve with different levels of agent assistance, under similar conditions.
Attribution also matters for recursive self-improvement, the idea that AI helps develop stronger AI systems that can then contribute to the next round of development. OpenAI’s report offers observations from parts of that process. Establishing a sustained feedback loop would take evidence across successive generations, including how much AI assistance changed the pace and quality of the results. OpenAI’s conclusion and proposed reporting
A practical guide to what to delegate
Start with a task whose result you can check faster than you can produce it. A weekly data-quality report is a good candidate. Given a read-only copy of a dataset, an agent can count duplicate IDs, flag missing fields and point you to the relevant rows. A person decides whether a duplicate is an error, whether a missing value is expected, and whether anyone should change the source.
| How to assign the work | Example |
|---|---|
| Delegate a bounded check | Ask the agent to apply rules you wrote, show counts, cite row IDs and flag cases the rules leave ambiguous. |
| Review the first example before scaling | Ask it to write one validation rule or generate one report on a small sample. Check the interpretation and output before it processes the full batch. |
| Keep a person responsible for consequential decisions | Have someone decide which rules are valid, approve any edits to the source and sign off on the findings people will act on. |
Here’s a first assignment you can adapt.
Use the read-only sample to check for duplicate IDs, missing dates and values outside the allowed categories. For each check, report the count and give up to five example row IDs. Flag ambiguous cases for review. Return one report and wait for feedback before running the checks on the full dataset.
Compare the counts with a spreadsheet formula, a short query or a small sample you’ve already checked. Inspect the rows the agent cites and check what it missed. Time the setup, review and corrections alongside the run. If you repeat the work, compare human minutes per accepted report with your manual baseline, using the same checks and acceptance criteria each time. Expand the assignment after the agent can deliver a report you can verify quickly.
