OpenAI reports 46% success rate for autonomous research agents on 2-4 hour tasks
·2026.11.18 17:00
Key point
Success rates for 2-4 hour research tasks rose from 23% to 46% between March and October 2026, with daily agent costs reaching a median of $2,400 per researcher.
Details
- OpenAI's internal data shows coding agents autonomously complete about half of 2-4 hour research tasks, with estimated success rates rising from 23% to 46% between March and October 2026.
- The "jagged frontier" of agent capabilities is evident in task-specific success rates: 79% for monitoring training runs, 35% for starting them, and 25% for writing inference/serving stack code.
- Daily agent usage costs per researcher reached a median of $2,400 in October 2026 (a 62% increase from September), with the 90th percentile at $23,000.
- Researcher agent workdays increased 11-fold to 5.7 per person per day, with over two-thirds of researchers running at least four agents simultaneously.
Research Impact and Metrics
- Safety & Alignment documented research effort (workweeks) increased by 77% from January to September 2026, the largest absolute increase among OpenAI research areas.
- The OpenAI Epoch Capabilities Index (OECI) frontier increased by 16.4 points annually since August 2025, exceeding the 14-point annual trend observed by Epoch since late 2024.
- Research experiments per active experimenter averaged 1.68 times the 2025 baseline during the four weeks ending October 18, 2026.
Workflow and Limitations
- Agents excel at code implementation, debugging, and experiment supervision but lack "taste" in research ideas, code quality, and writing.
- Human intervention remains necessary to prevent low-quality outputs ("AI slop"), inefficient experiments, and incorrect analysis assumptions.
- OpenAI researchers currently have no agentic coding quota, though a temporary harness efficiency improvement in August 2026 initially caused a sharp drop in input token utilization before being gradually reverted.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.