OpenAI: agents succeeded unaided in 46 per cent of assigned tasks estimated at 2–4 hours
In October, OpenAI’s coding agents succeeded without human help in 46 per cent of assigned research tasks estimated to take 2–4 hours, up from 23 per cent in March, TokenPost reported in its coverage of an OpenAI report. Results differ widely between task types.
On 18 November OpenAI published a report on how coding agents are changing its own AI research. Coding agents are programs that write and run code using a language model. The figures are the company’s internal measurements, and the report does not mention any external verification.
In October the agents succeeded without human help in 46 per cent of assigned research tasks estimated to take 2–4 hours, up from 23 per cent in March, TokenPost reported. For tasks estimated at 4–8 hours, the rate rose from 13 per cent to 35 per cent over the same period. OpenAI writes in the report that researchers may meanwhile have started giving agents different kinds of tasks. The rise may therefore reflect both better model capabilities and researchers’ choices about when and how they use agents.
The report’s title, The jagged frontier, refers to how widely agents’ success varies from task to task. The table shows the agents’ estimated success rate on different tasks without human help.
| Task or task category | Success rate |
|---|---|
| Monitoring runs for issues | 79% |
| Answering technical questions | 72% |
| Communication (category) | 69% |
| Analysis (category) | 58% |
| Building (category) | 43% |
| Building evaluation, reinforcement-learning and agent environments | 29% |
| Code for inference and serving stacks | 25% |
Source: OpenAI, 1 August–18 October 2026, categories by Epoch AI. Figures are estimates.
An average success rate therefore does not show which tasks an agent handles well. OpenAI said the weakest tasks may limit the overall benefit as long as the gaps remain.
At API list prices, researchers’ median spending on agents is about $2,400 a day and the 90th percentile about $23,000. In the September report the figures were $1,500 and $11,000. Even so, internal use – Codex, ChatGPT and the API combined – accounts for only 2 per cent of total research compute.
Researchers interviewed said the agents show poor judgement on research ideas and on the quality of code, data and text. To avoid low-quality results and wasted iterations, researchers have to stay closely involved. The interviewed researchers said the most important ideas for GPT-6 Astra came from people. For GPT-6.2 Astra, pretraining researchers developed ideas together with Codex.
OpenAI put a scenario to 101 research staff in which their team would lose access for three months to coding agents better than GPT-5.2. In the median response, the team would give up 40 per cent of its headcount to regain access. The company cautions that self-assessments may be inaccurate and that respondents were selected ad hoc.
OpenAI said that more lines of code and more experiments do not necessarily mean research is progressing faster. The company believes AI labs should be required to track and publicly report indicators of recursive self-improvement. The term refers to AI accelerating the development of future AI systems.
Read more: How an agent works · Models