How good are AI coding agents at real research work — not benchmark puzzles, but the actual tasks researchers do every day? OpenAI's post "Research acceleration: The jagged frontier", published on 18 November 2026 as a follow-up to a September post, tries to answer that question with its own internal numbers. The headline is striking: on tasks estimated at 2–4 hours, the share completed with zero human intervention rose from 23% in March 2026 to 46% in October 2026. On longer 4–8 hour tasks, it rose from 13% to 35% over the same period.
The "jagged frontier" in the title is the point. The gains are not uniform: agents are strong at some tasks and weak at others, and the boundary between the two is uneven. Task-level figures for 1 August to 18 October show the shape of that boundary clearly — monitoring training runs succeeded 79% of the time and answering technical questions 72%, while launching training runs managed 35% and writing inference/serving code only 25%.
Key facts
| Item | Details |
|---|---|
| Post | "Research acceleration: The jagged frontier" (OpenAI) |
| Published | 18 November 2026 (follow-up to a September 2026 post) |
| Zero-intervention success, 2–4 hr tasks | 23% → 46%, March → October 2026 |
| Zero-intervention success, 4–8 hr tasks | 13% → 35%, March → October 2026 |
| Task-level, 1 Aug – 18 Oct 2026 | Monitoring training runs 79% · Answering technical questions 72% · Launching training runs 35% · Writing inference/serving code 25% |
| Median researcher daily agent spend | About $2,400 in October 2026 (+62% vs September) |
| Safety workweeks | Grew 77% "since January" |
| Nature of all figures | OpenAI's own internal self-measurements — no independent audit |
What this means
The most useful reading of this post is not the headline doubling but the jaggedness. Agents now handle routine monitoring and question-answering most of the time without help, yet still fail roughly three-quarters of the time when asked to write inference and serving code — the kind of work where a mistake has real consequences. Anyone planning to rely on coding agents should plan around that unevenness: delegate the tasks the frontier has already crossed, and keep humans close for the ones it has not.
The spending figure tells its own story. A median of about $2,400 per researcher per day on agents in October — up 62% on September — shows that inside OpenAI, agent use is not an experiment any more; it is a large and growing running cost. That also frames the 77% growth in safety workweeks since January: as agents do more, the work of keeping them safe grows with them.
OpenAI's own caution deserves the last word here. The company notes that the upward trend mixes two things: agents getting more capable, and researchers changing which tasks they hand to agents. As researchers learn what agents are good at, they delegate more of that work — which lifts success rates even if the agents themselves had not improved at all. The figures are real measurements of OpenAI's own workflows, but they are not a clean measure of agent capability alone, and no independent audit exists.
A note on verification
Every figure in this post comes from OpenAI's own post of 18 November 2026 and describes OpenAI's internal workflows. No independent audit of these numbers exists. OpenAI itself cautions that the trend mixes capability gains with changes in which tasks researchers delegate. We will update this post if independent measurement or a follow-up report appears.
Source: OpenAI, "Research acceleration: The jagged frontier", 18 November 2026. Last checked: 21 November 2026.
FAQs
What is the "jagged frontier"?
It is OpenAI's term for the uneven boundary of what coding agents can do: they succeed most of the time at some tasks and fail most of the time at others, rather than improving evenly across all work.
How much did success rates improve?
According to OpenAI, zero-intervention success on 2–4 hour tasks rose from 23% in March 2026 to 46% in October 2026, and on 4–8 hour tasks from 13% to 35%. These are OpenAI's own internal measurements.
Which tasks do agents handle best and worst?
For 1 August to 18 October 2026, OpenAI reports monitoring training runs at 79% and answering technical questions at 72%, against launching training runs at 35% and writing inference/serving code at 25%.
Are these figures independently verified?
No. All figures are OpenAI's own internal self-measurements with no independent audit. OpenAI itself cautions that the trend mixes capability gains with researchers changing which tasks they delegate.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!