OpenAI published a report on November 18 titled "Research acceleration: The jagged frontier", and its headline number is that coding agents inside the company now finish 46% of delegated research tasks that would take a human 2–4 hours, with no human stepping in. That's up from about 23% in March. For tasks a human would need 4–8 hours to do, the rate rose from about 13% to 35% over the same period. These figures come from coverage by DEV Community, GPTunneL, DigitalToday and CellCog, and all four agree on them. We couldn't load OpenAI's own page for this run, so every number here is as those outlets report it.
This is the second time OpenAI has measured its own staff this way; the first report came out in September. It matters because it's the most detailed public data yet on how much frontier-lab research is being handed to AI, and OpenAI uses it to call on every lab to track and publish signs of recursive self-improvement.
The headline numbers
| Metric (OpenAI internal data, as reported) | Earlier | Latest |
|---|---|---|
| Unassisted success, 2–4 hour tasks | ~23% (March) | ~46% (October) |
| Unassisted success, 4–8 hour tasks | ~13% (March) | ~35% (October) |
| Median researcher's daily agent spend (API list price) | ~$1,500 (Sept) | ~$2,400 (Oct) |
| 90th-percentile researcher's daily spend | ~$11,000 (Sept) | ~$23,000 (Oct) |
| Agent workdays per researcher per day | 0.5 | 5.7 |
The duration buckets measure how long a human would take, not how long the agent ran, GPTunneL notes. One "agent workday" is eight hours of agent runtime, according to DEV. The spend figures are accounting values at API list prices, not cash OpenAI paid out; all four outlets make that point. CellCog, which built its numbers from charts OpenAI researcher Kevin Liu posted on X, puts the October median just under $2,500 rather than $2,400. It also says its values were read off the charts by eye.
DEV also reports that by mid-October, more than two-thirds of researchers hit a daily peak of four or more agents running at once.
Why "jagged"
The success rate swings widely by task type. According to DEV and GPTunneL, which cover August 1 to October 18:
- Monitoring training runs: 79%
- Answering technical questions: 72%
- Launching training runs: 35%
- Building evaluation, RL or agent environments: 29%
- Writing inference and serving code: 25%
By broad category, communication tasks succeeded 69% of the time, analysis 58% and building systems 43%. Put simply, agents are reliable at watching and reporting, and much weaker at building new infrastructure. GPTunneL says OpenAI's researcher interviews named recurring problems: weak hypotheses, redundant exploration, flawed analytical premises and brittle code edits. A common workaround is to have an agent produce one "golden sample", check it, then batch the rest.
What researchers say they'd give up
In a survey of 101 research staff, OpenAI asked what people would trade to avoid being limited to coding agents no better than GPT-5.2 for three months. The median answer was 40% of their team's headcount or 30% of its compute. All four outlets report that result.
Safety and alignment working time grew 77% from January to September 2026, according to DEV and GPTunneL. CellCog quotes OpenAI saying effort there was "growing faster than in any other individual research area."
Caveats worth keeping in mind
- Self-measured. Every figure comes from OpenAI's internal telemetry and surveys, with no independent audit (DEV, CellCog).
- Mixed causes. People also changed what they delegate over the period, so the gains combine better models with better task choice. The report doesn't separate the two (DEV, GPTunneL).
- Not end-to-end autonomy. In GPTunneL's reading, the 46% applies to delegated operational work such as monitoring, eval upkeep and refactoring. It is not a rate for whole research projects, and OpenAI says it hasn't achieved closed-loop autonomous model development.
- Small samples. None of the coverage we read gives the number of tasks behind the success rates. The survey is self-reported.
- Rounding. The reported percentage increases (62% for the median, 105% for the 90th percentile) don't exactly match the rounded dollar figures. Don't recompute them from the table.
What to watch
OpenAI's ask is the part with the widest reach: it wants AI labs to track and publicly report indicators of recursive self-improvement. DEV says OpenAI frames that as transparency, not as a plan to pursue full self-improvement. Watch whether Anthropic, Google DeepMind or others publish comparable telemetry, and whether outside evaluators get access to check numbers like these.
For engineering teams, the task breakdown is a practical guide. Monitoring training runs and answering technical questions are where agents already succeed unattended more than seven times in ten. New infrastructure and serving code still need a human checking each change.
Sources
- DEV Community — OpenAI's Coding Agents Now Finish 46% of 2–4 Hour Research Tasks Alone — But the Frontier Is Jagged dev.to
- GPTunneL — OpenAI Agents Complete 46% of 2–4 Hour Research Tasks: Study gptunnel.ru
- DigitalToday — OpenAI researchers' AI agent use surges, study shows digitaltoday.co.kr
- CellCog — OpenAI's Research Report: Top Researchers Spend $23K a Day cellcog.ai