OpenAI Says Its Coding Agents Now Finish 46% of Medium Research Tasks Alone
OpenAI published an internal report on November 18 claiming its coding agents autonomously complete nearly half of research tasks that would take humans two to four hours. The figure, drawn from the company’s own tracking between March and October 2026, is its most concrete public benchmark yet for AI-driven research acceleration. But the report also documents wide variation across task types, growing reliance on parallel agent fleets, and persistent human oversight — a picture the company itself labels “the jagged frontier.”
The headline number and its limits
OpenAI’s report, “Research acceleration: The jagged frontier,” states that the zero-intervention success rate for tasks estimated at two to four human hours rose from 23% in March to 46% in October. For tasks estimated at four to eight hours, the rate climbed from 13% to 35%. The company cautions that these buckets mix capability gains with researchers’ shifting choices about what to delegate: as agents improve, staff may hand them different tasks within the same duration band, so the rates do not isolate model improvement alone [1].
Sharp differences by task type
Breaking results down by activity using the O*NET AI R&D taxonomy developed by Epoch AI, OpenAI found autonomous success rates from August 1 to October 18 ranged from 79% for monitoring training runs and 72% for answering technical questions, down to 35% for launching training runs, 29% for building evaluation and RL environments, and 25% for writing inference and serving code. The company notes that lower-performing categories could become bottlenecks to overall uplift [1].
Usage surging, costs climbing
At API list prices, the median researcher’s daily agent spend reached $2,400 in October, up 62% from September; the 90th percentile hit $23,000, up 105%. Measured in “agent workdays” (eight hours of agent runtime), the average rose elevenfold from 0.5 to 5.7 per researcher per calendar day. By mid-October, more than two-thirds of researchers ran at least four agents concurrently at their daily peak; over 10% ran sixteen or more. Internal agent spend in September was roughly twice the non-equity compensation bill for researchers, though equity makes up 70–80% of target pay. Agent compute still accounts for only 2% of total research compute [1].
Researchers say agents are critical — but flawed
A survey of 101 research staff (late August to early September) asked what they would trade to keep frontier coding agents for three months. The median answers: 40% of team headcount or 30% of team compute. Researchers also estimated they would need three months to match one month of agent-assisted output using only GitHub Copilot plus ChatGPT. OpenAI flags the survey’s ad-hoc sampling and the known unreliability of self-reports [1].
In follow-up interviews, ten researchers described agents as strong at implementation, debugging, and large-scale scanning, but weak on research taste, code quality, data quality, and writing. Agents “rabbit-holed on unnecessary steps,” produced “AI slop” pull requests, and made invalid analysis assumptions. Workflows now require close human review: “Go make one. I’ll review it and make edits. Then go do the rest,” one post-training researcher said. A pretraining researcher added: “I do think I have more taste in terms of the ideas. But I have no confidence that will continue to be true for any significant period of time” [1].
Ideas still come from humans — mostly
OpenAI asked key contributors to GPT-6 Astra and GPT-6.2 Astra where their most important ideas originated. For both releases, post-training and safety researchers said people generated all the ideas; models helped implement them. Pretraining researchers credited people for GPT-6 Astra ideas. For GPT-6.2 Astra, two pretraining researchers described developing ideas in conversation with Codex; one said Codex identified key insights, another said a Codex suggestion made the main idea work [1].
Safety work growing fastest
Documented researcher workweeks in Safety & Alignment almost doubled from January to September, the largest absolute increase of any research area. OpenAI attributes this to new hires and existing staff shifting focus, with an inflection after the “Hugging Face incident” in July 2026. The company notes the boundary between post-training and safety work is nontrivial and may evolve [1].
Output metrics: activity up, capability trend tracked
Lines of code changed per active contributor and experiments per active experimenter (1.68× the 2025 baseline in the four weeks to October 18) have risen sharply. OpenAI also introduced an internal version of Epoch AI’s Capabilities Index (OECI). Its frontier for OpenAI models has risen 16.4 points per year since August 2025, versus the 14 points per year Epoch observed since late 2024. OpenAI notes its models score higher on OECI than on Epoch’s public ECI, likely due to benchmark mixture differences [1].
Caveats and context
OpenAI stresses that historical human-AI collaboration trends may not predict future fully autonomous progress. The report’s methods — classifiers on Slack posts, pull requests, and internal documents — proxy attention allocation, not formal team assignments. The labor-cost comparison excludes equity. The company reiterates that publishing these metrics is not a statement of intent to pursue full recursive self-improvement (RSI), but a call for industry-wide RSI reporting norms. It cites three Anthropic reports, a Google DeepMind section in the Gemini 4 Argon safety report, and a METR-Anthropic collaboration as other recent contributions to this measurement effort [1].
What it means
The 46% figure is a specific, testable claim about a defined task band inside one lab, measured with disclosed (though evolving) methodology. It reflects real, deployed utility for structured engineering tasks — monitoring, Q&A, some launches — while confirming that open-ended build work remains low-autonomy. The elevenfold rise in agent workdays and the headcount/compute trade-off responses signal that agents have become load-bearing infrastructure for OpenAI’s research velocity. Whether that velocity translates into sustained capability acceleration, and whether the jagged frontier smooths or deepens, is what the company’s own OECI and future reports will test.
Sources
- Research acceleration: The jagged frontier (openai.com)
- OpenAI Researchers Consume $2400 of AI Per Day as Agents ... (finance.biggo.com)
- OpenAI Coding Agents Hit 46% Task Completion Rate in 2026 (emergent.sh)
- OpenAI says agents independently completed 46% of two-to-four-hour ... (mezha.net)
- OpenAI says its agents now finish nearly half of 2 to 4-hour research tasks (thenextweb.com)