The headline number doubled. Read what it counts

OpenAI's measure is narrow and, to its credit, stated plainly. A task counts as a success only if an agent completed it with no recorded human intervention, and the hours are estimates of the human work the task would take. On that measure, tasks estimated at two to four hours went from 23% success in March to 46% in October. Tasks in the four-to-eight-hour band went from 13% to 35%.

Then OpenAI qualifies its own headline, in writing. As agents became more capable, researchers may also have changed the types of tasks they delegated within each duration band, so the change in success rates may reflect both capability improvements and researchers' changing choices about when and how to use agents. The duration estimates are OpenAI's and the figures are internal estimates. A rise on a possibly shifting denominator is progress of some kind. It is not yet a productivity figure, and the report does not claim it is one.

Monitoring scores 79%. Serving code scores 25%.

The breakdown covering August 1 to October 18 is where the report earns its title. Agents monitoring a training run for problems succeeded 79% of the time. Answering technical questions scored 72%. Launching a training run dropped to 35%. Building evaluation, reinforcement-learning and agent environments reached 29%. Writing the code for inference and serving stacks, the software that delivers a model to users, sat at 25%.

Grouped more broadly, the pattern holds: Communicate at 69%, Analyze at 58%, Build at 43%. Agents score highest where the work is watching, summarizing and answering, and lowest where the work is constructing the systems research depends on. OpenAI names the consequence itself: so long as performance is this uneven, the lower-performing categories and tasks could become bottlenecks to high overall uplift. Anyone quoting the 46% without the 25% is quoting the average and skipping the constraint OpenAI itself flags.

The usage figures, priced at list

Agent workdays per researcher rose from 0.5 to 5.7 a day. By mid-October, more than two-thirds of included researchers averaged a daily peak of at least four agents working concurrently, and more than 10% averaged at least sixteen concurrent agents.

Valued at API list prices, the median researcher's daily agent use reached about $2,400 in October, up 62% from $1,500 in September. The 90th percentile reached about $23,000, up 105%. These are API-equivalent values for usage, not invoices. OpenAI notes that as of September the API-equivalent value of agent use in its research organization was about twice its non-equity pay, while also noting that equity makes up 70 to 80% of target researcher compensation and is excluded from that comparison. Both halves of that sentence matter. The direction is unmistakable even after the discount: usage, valued at list prices OpenAI sets itself, is large and growing fast, and its measured success still depends heavily on the task.

The survey figure needs the same treatment. Asked a hypothetical, what they would trade to restore frontier coding agents if access to agents better than GPT-5.2 were revoked for three months, the median of 101 research staff said 40% of team headcount, or 30% of team compute in a separate question. That is a statement about how valued the tools feel to the people using them. It is not a staffing plan, and OpenAI does not present it as one.

Why OpenAI published this at all

The report frames itself as an indicator for recursive self-improvement, the process of AI speeding up the development of future AI, and argues labs should be required to track and publish such indicators. It also states, in the same introduction, that its reporting should not be read as an intention to pursue full recursive self-improvement. Readers can hold both: the measurements are useful precisely because they are specific, and the framing is also a policy position, a proposal that the industry adopt a reporting standard OpenAI has already built for itself.

One quieter figure may matter more than the agent scores. Documented Safety & Alignment workweeks rose 77% from January to September, the largest absolute increase in any area OpenAI measured, reflecting new hires and existing researchers shifting toward those workstreams, with growth picking up after the July 2026 Hugging Face incident. The report does not say those weeks were spent checking agent output, and it does not attribute the shift to agent use. What it establishes is narrower: while agent use surged, the largest growth in documented human research effort, on OpenAI's own count, was in safety and alignment.

Sources

  1. [1] OpenAI, "Research acceleration: The jagged frontier," November 18, 2026Read source
  2. [2] Superpower Daily, "OpenAI estimates coding agents finish 46% of its two-to-four-hour research tasks," November 18–19, 2026 (secondary summary; figures cross-checked against OpenAI's report read first-hand)