OpenAI Reveals Internal Research-Acceleration Data: 7x Code Changes, Safety Effort Up 77.2%
On November 19, OpenAI researcher daveholtz shared a thread of internal "research acceleration measurement" data spanning R&D output, agent task performance, usage costs, staffing priorities, and governance judgment—an overall picture of AI deeply permeating the R&D process while driving cost and governance pressures.
Confirmed
- Output time series: daily code changes per active contributor have climbed to roughly 7x the pre-2025 average (a steep rise since late 2025); daily experiments per experimenter are up about 1.67x versus the 2025 baseline; daveholtz also cited OECI rising 16.4 points year over year.
- Agent task performance: when stratified by estimated human completion time, success rates for long tasks completed by AI without human intervention have risen overall—for example, 4–8 hour tasks went from about 12% in March to about 35% in October, though the figures are affected by both capability gains and shifts in task mix; monitoring tasks reached a 79% success rate.
- Usage cost: as of October, the median researcher spends $2,400 per day at API list prices on agentic tools (Codex, ChatGPT Work, Dots), with the 90th percentile higher still; token costs have reached 2x personnel costs.
- Researcher valuation: internal surveys and interviews show the median answers for keeping frontier coding agents for three months were a willingness to give up 30% of team compute or 40% of team headcount; qualitative interviews also flagged concerns such as research taste.
- Staffing: from January to September 2026, safety and alignment headcount time grew 77.2%, the largest increase of any research area, versus pre-training +39.9% and post-training +14.6%.
Not Yet Confirmed
- All figures come from daveholtz's thread; OpenAI's original report, methodology, and full dataset have not been released.
Why It Matters
- These numbers offer a rare quantification of AI's real boost to R&D efficiency inside a frontier lab: multiplicative output gains coexist with heavy token spending, showing that "AI accelerating AI R&D" has entered an operational phase with substantial costs.
- daveholtz's closing judgment: as AI gets better at improving itself, competitive pressure and governance challenges will grow; OpenAI sees keeping humans in the loop as the key task and plans to keep refining its metrics methodology and expanding reported domains.
2026-11-19 ~ 2026-11-19 · 9 related posts
Primary sources
- OpenAI: researcher code output now 7x pre-2025 average, OECI frontier rising 16.4 points/year — daveholtz ·
- OpenAI: median researcher spends $2,400/day on agents, token costs now 2x labor costs — daveholtz ·
- OpenAI: as AI improves itself, governance challenges grow — humanity must stay in the loop — daveholtz ·
- [source] OpenAI: median researcher spends $2,400/day on agents, token costs now 2x labor costs — daveholtz · 2026-11-19
- OpenAI data: zero-intervention agent success hits 79% on run monitoring but drops for algorithm implementation — daveholtz · 2026-11-19
- OpenAI researchers would trade 40% of team headcount to keep frontier coding agents — daveholtz · 2026-11-19
- [source] OpenAI: researcher code output now 7x pre-2025 average, OECI frontier rising 16.4 points/year — daveholtz · 2026-11-19
- OpenAI data shows human time on safety and alignment up 77.2% in 2026, fastest-growing research area — daveholtz · 2026-11-19
- [source] OpenAI: as AI improves itself, governance challenges grow — humanity must stay in the loop — daveholtz · 2026-11-19
- OpenAI data: 90th-percentile researcher now spends $23,000 on agentic AI tools — kliu128 · 2026-11-19
- OpenAI survey: researchers would trade 40% of headcount to keep 2026 coding agents — kliu128 · 2026-11-19
- OpenAI internal data shows smooth gains in autonomous research task length — kliu128 · 2026-11-19