OpenAI estimates its coding agents succeed without human help on 79% of tasks monitoring training runs, but on 25% of tasks writing code for its inference and serving stacks, OpenAI’s 18 November report shows.
That is a gap of roughly 3.2 times between the best and worst task OpenAI highlights (our arithmetic). OpenAI calls the pattern a jagged frontier.
Business Pill: The Cost of Checking
An evergreen explainer on who pays to check work, and why cheap verification changes what can be automated. It does not discuss OpenAI’s report.
The key insight: OpenAI estimates its agents succeed without help on 79% of training-run monitoring tasks but 25% of tasks writing inference and serving code, a pattern it calls a jagged frontier.
Where the Agents Succeed
From 1 August to 18 October, OpenAI estimates success without human intervention at 69% for Communicate tasks and 58% for Analyze tasks.
Build tasks came in at 43%, 26 points below Communicate (our arithmetic). OpenAI uses Epoch AI’s O*NET AI R&D taxonomy for the categories.
At the task level, answering technical questions reached 72%. Building evaluation, reinforcement-learning and agent environments reached 29%.


Where They Still Fall Short
OpenAI puts launching training runs at 35%, against 79% for monitoring them. It says the weak spots matter for the whole pipeline.
“So long as model performance on AI R&D-related tasks is heterogeneous, lower-performing categories and tasks could become bottlenecks to high overall uplift”, the report says.
In interviews, researchers said agents “struggled to launch experiments efficiently” and wrote “AI slop” pull requests. One said the model “generally can’t one-shot come up with well-posed problems that it cannot solve itself”.
Logged experiments per active experimenter averaged 1.68 times the 2025 baseline in the four weeks to 18 October, OpenAI adds.
What OpenAI Wants Measured Next
OpenAI repeats that it believes “AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI)”.
It says it hopes to add reporting on “Differential acceleration of capabilities vs. safety research” and on leading indicators of research acceleration.
Connecting the Coverage
Our first read of the report covered the cost side: OpenAI’s research agents now cost about twice researchers’ cash pay. The same week, Vals AI said GPT-6.2 Astra formalized a theorem in Lean.
The call for RSI reporting sits beside Paul Christiano’s warnings, Dan Hendrycks’ exit from advising AI firms and the MisAnthropic Statement.
The Structural Read
As we read the task data, the checking work is being automated first, while the building work that creates new systems still needs people.
As we read OpenAI’s policy section, it wants labs to be measured on the same signals it is now publishing, including whether capabilities outpace safety research.
OpenAI, Research acceleration: The jagged frontier, 18 Nov 2026
“So long as model performance on AI R&D-related tasks is heterogeneous, lower-performing categories and tasks could become bottlenecks to high overall uplift”
Three Implications
FOR RESEARCH MANAGERS OpenAI estimates agents finish 79% of run-monitoring tasks alone but 35% of training-run launches.
FOR AI BUYERS OpenAI puts Build tasks at 43% success without human help, against 69% for Communicate tasks.
FOR POLICY READERS OpenAI says labs should be required to track and publicly report indicators of recursive self-improvement.
What Is Not Established
All success rates are OpenAI’s own estimates, with confidence intervals it publishes. OpenAI says researchers may have changed which tasks they hand to agents.
The report does not set thresholds for when RSI indicators should trigger action, or say which body should receive the reports. We did not contact OpenAI.
The Bottom Line
By OpenAI’s own estimates, agents now succeed on most monitoring and answering tasks, while build tasks such as serving code and training environments stay below half.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Every figure, quote and chart value is from OpenAI’s 18 November report ‘Research acceleration: The jagged frontier’, read in full. The duration chart is OpenAI’s own, credited. The 3.2-times spread and the 26-point gap are our arithmetic on OpenAI’s figures. We did not contact OpenAI.
Sources: OpenAI, ‘Research acceleration: The jagged frontier’, research report, 18 Nov 2026 · OpenAI on X, 18 Nov 2026, 23:07 UTC









