OpenAI's 'jagged frontier' report: coding agents finish 46% of 2–4 hour research tasks alone

OpenAI has published its second report tracking how far AI agents can accelerate research itself. The headline number: coding agents now complete 46% of research tasks sized at 2–4 hours of human effort without help. The title, 'The jagged frontier', signals the main caveat. Capability is not a smooth curve. Agents excel at some task categories and fail badly at others of similar length.
The framing extends OpenAI's September 'automated research intern' milestone into a measured, task-length-based benchmark. Measuring by the time a task would take a human has become a common way to track agent autonomy. A move toward multi-hour tasks is significant because it describes work that used to require a junior researcher's afternoon.
The release lands in a week dominated by cost-per-task claims. GPT-6.2 Astra and Sol launched with OpenAI saying they cut per-task costs. Airbnb published how it picks coding-agent models on Bedrock by cost per task. AWS's Swami Sivasubramanian also open-sourced BenchProofer, arguing that agents can pass tests and still be wrong on inputs no test checks.
That last point is the skeptic's lens on the 46% figure. How 'completion' is verified determines how much the number means. The current source is a secondary dev.to write-up, so readers should look for OpenAI's own methodology: which task categories lag, and how success was graded.