OpenAI Discloses Internal AI-for-Research Data: 3.1 Agent Days per Human Day, $600 Daily Token Spend

Over the Labor Day weekend, OpenAI published two blog posts (including "Research acceleration") with rare first-party data on AI agents inside its research organization: as of mid-August, every 1 human workday is paired with 3.1 agent workdays of output; agents can produce more than three days' worth of research per researcher per day; task complexity delegated to agents keeps rising alongside code and experiment volume. Fortune reads this as the company moving toward recursive self-improvement.

Confirmed

• 3.1 agent workdays per human workday per researcher (m2, m4, m8)
• Median researcher token spend exceeds $600/day ($12k/month); 90th-percentile researcher exceeds $7,000/day ($140k/month) (m6, m8, m9)
• At the start of the year the median researcher did not use coding agents; usage density rose sharply through the year; code output grew 6x (m6, m9)
• Among successful 32-hour-class agent tasks, over 80% still required human intervention (m3)
• The two posts contrast sharply: one showcases acceleration, while the chief scientist urges the industry to slow down (m7)

Unconfirmed

• User scaling01 argues the published $600/$7,000 figures are just API prices, and OpenAI's real internal compute cost could be 5-10x lower; their uncertainty extrapolation includes extreme cases up to $1M per day. This is a personal estimate, not confirmed by OpenAI (m5)

Why it matters

This is the first time a frontier lab has systematically quantified the intensity of AI-for-AI-R&D, directly bearing on the speed and risk of recursive self-improvement: with agent output at roughly 3x a human researcher's volume, research scaling is shifting from headcount to compute (m4). DataLearnerAI cautions that "successful 6-hour agent tasks" do not mean 6 hours of fully autonomous work, with over 80% of long tasks still requiring human intervention (m3); the coexistence of the acceleration narrative with the chief scientist's slowdown plea highlights internal tension at OpenAI over pace versus risk (m7).

2026-09-07 ~ 2026-09-09 · 9 related posts

• OpenAI internal data: over 80% of successful 32-hour agent tasks still needed human intervention — DataLearnerAI · 2026-09-07
• OpenAI says its 90th-percentile researchers burn over $7,000 in tokens per day — SatelliteNetSec · 2026-09-08
• OpenAI says its AI agents produce 3 days of research per day, even as chief scientist Jakub Pachocki urges industry slowdown — jeremyakahn · 2026-09-08
• Median OpenAI researcher now burns $600+/day in inference at API prices — VraserX · 2026-09-08
• [source] OpenAI says research org runs 3.1 agent-workdays per human workday — Cklly2004 · 2026-09-08
• [source] OpenAI: AI agents now do 3.1 days of research per human workday amid safety scares — fortune · 2026-09-09
• OpenAI researchers now burn $600/day in tokens median, top users $7,000+ — downingARK · 2026-09-09
• Analyst claims OpenAI's published figure is just API pricing, real cost may be 5-10x lower — scaling01 · 2026-09-09

1 near-duplicate retellings: downingARK