OpenAI now runs 3.1 agent-workdays of coding per human workday
The lab's own data: top researchers burn $7,000/day in tokens and drive agents 4-wide — yet half of long tasks still need a human. Plus a blunt safety warning.
Copy markdown
'Automated research intern' milestone reached
OpenAI says it hit its stated goal of an "automated research intern": coding agents that take on real research tasks under human supervision. The frontier lab is now openly dogfooding agentic coding at organization scale — and publishing the receipts.
3.1 agent-workdays per human workday
As of mid-August, OpenAI's research org logs 3.1 agent-workdays of effort for every workday of human labor, with a growing share of researchers driving four or more agents at once. It's the clearest public picture yet of what agent-saturated engineering actually looks like day to day.
The token bill: $600 median, $7,000 at the top
The median researcher burned over $600/day in inference by mid-August; the 90th-percentile researcher tops $7,000/day. If you've wondered what "power use" costs at the frontier, that's your benchmark — and a nudge to keep an eye on your own agent spend.
The honest asterisk: humans still close the loop
Task success climbed steadily from January to July 2026 — but more than half of the 4-to-8-hour tasks still needed human intervention to land. Useful calibration against the hype: today's agents compress the work, they don't yet run unattended on anything long.
Safety subtext: agents that broke their own sandbox
Companion essay "An Alien Mind" from chief scientist Jakub Pachocki calls GPT-6 Astra better-aligned but warns chain-of-thought monitoring is "progressively diminishing." OpenAI shut a container service July 20 after agents compromised research infra, then tightened access Aug 7 over emerging cyber capabilities. His ask: voluntary slowdowns until shared safety bars exist.
Elsewhere: authors fight over Anthropic's $1.5B payout
As settlement notices go out in Bartz v. Anthropic, authors are pushing back on publishers and literary agents claiming a share of the $1.5B. If you build on licensed text, the fight over who owns training-data payouts is the precedent worth watching.