OpenAI's agents now complete 46% of 2–4 hour research tasks on their own, up from 23%
OpenAI has measured how coding agents perform in its own research. For tasks that take a human 2–4 hours, agents now finish 46% without any help, up from 23% in March. But success depends on the type of work: agents can monitor model training runs in 79% of cases, yet write inference code (the code that runs the model) in only 25%. The takeaway is to hand routine work to agents and keep core infrastructure with humans.
AI-processed from OpenAI Blog; edited by Hamidun News
OpenAI has published a second report on how coding agents speed up its own AI research. The problem is that everyone talks about "accelerating development," but few show real numbers. The lab measured it: for tasks that take a human 2–4 hours, agents now complete 46% without a single human correction, up from 23% in March.
Imagine telling an AI to analyze the logs of a failed training run or to launch an experiment, and then not touching the results by hand. For 4–8 hour tasks, the success rate rose from 13% to 35%. If you are rolling out agents in your team, this is a guide to what you can already hand them, what you can't yet, and how much it may cost.
Where agents are strong and where they are weak
The speedup is uneven, which is why the report is called "the jagged frontier." Under Epoch AI's O*NET AI R&D taxonomy, from 1 August to 18 October 2026 agents did best at communication (69%) and analysis (58%), and worse at building (43%).
At the level of individual tasks, the spread is even wider:
- monitoring training runs for problems — 79%;
- answering technical questions — 72%;
- launching training runs — 35%;
- building environments for evaluation, reinforcement learning and agents — 29%;
- code for inference and serving stacks (that is, for running a model and serving it to users) — 25%.
OpenAI adds a caveat: growth within each category reflects more than model improvement. Researchers also change over time which tasks they give to agents. While results remain uneven, the weak categories will limit the overall speedup, and humans remain an important part of the research loop.
What it costs and how work is changing
Agent usage is growing almost exponentially. At API list prices, the median researcher spends about $2 400 a day, 62% more than in September ($1 500). At the 90th percentile, meaning the most active 10% of employees, spending is $23 000 a day, 105% more than the $11 000 in September. The growth comes from higher activity, a shift to more expensive models, and changes in the agent harnesses themselves.
The company also counts "agent workdays": every 8 hours of agent work equals one day. The average per researcher was 0.5 such days; it is now 5.7, an 11-fold increase. The figure is highest on the pretraining and post-training teams, at about eight, and about four in scaling and infrastructure. The metric mostly captures interactive work, so it may underestimate automation.
By mid-October, more than two-thirds of researchers were running at least four agents at once at their peak, on average, and more than 10% were running at least sixteen.
In September 2026, the API-equivalent cost of agents was roughly double researchers' cash compensation, excluding equity. Equity typically makes up 70–80% of target compensation. Meanwhile, all internal usage (Codex, ChatGPT and the API) takes only 2% of the compute allocated to research.
In a survey, researchers were asked what they would be willing to give up to keep frontier coding agents for three months. The median answers were 30% of their team's compute or, in a separate question, 40% of their team's headcount.
Safety and open reporting
Researchers' time is shifting toward safety. Documented weekly contributions to Safety & Alignment grew 77% from January to September. The causes are new hires and existing researchers moving into these areas. This is the largest absolute increase among all of OpenAI's major areas. The pace picked up noticeably after the Hugging Face incident in July 2026 and the subsequent work on safety cases for training frontier models. The area includes alignment evaluations, monitorability and reward hacking research. The company itself acknowledges that the line between post-training and safety is nontrivial, and the definition may change.
OpenAI stresses that it publishes these metrics because it believes AI labs should track and publicly disclose indicators of recursive self-improvement (RSI), where AI accelerates the creation of the next AIs. By its own account, the report does not signal an intention to pursue full RSI.
What to do in practice
Don't measure agent adoption by the overall share of "AI-written code." Break work down by task type. Agents take on monitoring, log analysis and technical questions more readily, while launching experiments and infrastructure code are better kept under human control. Plan the agent budget as a separate, fast-growing line item: for the most active users it is already several times the average.
- Nov 18, 2026AWS adds Dynamo to HyperPod: Qwen3-235B delivers 207 tokens/s per GPU instead of 101
- Nov 18, 2026New Jev model answers in 150 ms versus 1.89 s for GPT-6 Astra
- Nov 18, 2026Huang steps down as NVIDIA CEO, company loses more than $1 trillion in market cap
- Nov 18, 2026The Best Coding Agent Controls a Robot in 77% of Tasks, but With Sensors Alone Only 30%
Key Questions Answered
Where agents are strong and where they are weak?
The speedup is uneven, which is why the report is called "the jagged frontier." Under Epoch AI's ONET AI R&D taxonomy, from 1 August to 18 October 2026 agents did best at communication (69%) and analysis (58%), and worse at building (43%).
What it costs and how work is changing?
Agent usage is growing almost exponentially. At API list prices, the median researcher spends about $2 400 a day, 62% more than in September ($1 500). At the 90th percentile, meaning the most active 10% of employees, spending is $23 000 a day, 105% more than the $11 000 in September. The growth comes from higher activity, a shift to more expensive models, and changes in the agent harnesses themselves. The company also counts "agent workdays": every 8 hours of agent work equals one day.
What are the key details on Safety and open reporting?
Researchers' time is shifting toward safety. Documented weekly contributions to Safety & Alignment grew 77% from January to September. The causes are new hires and existing researchers moving into these areas. This is the largest absolute increase among all of OpenAI's major areas. The pace picked up noticeably after the Hugging Face incident in July 2026 and the subsequent work on safety cases for training frontier models. The area includes alignment evaluations, monitorability and reward hacking research.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.