On September 6, OpenAI published an unusually candid post, *Research acceleration: The view inside OpenAI*. No new products, no pricing — just first-party data on three months of collaboration between OpenAI researchers and coding agents. Before anything else: all numbers below measure inputs, not outputs.
Headline numbers
- By mid-August, the research organization used 3.1 agent-workdays of effort for every workday of human labor. This is an organization-wide total ratio, not "3.1 agents assigned per researcher." At an 8-hour workday, agents collectively burn roughly 25 hours per 8 human hours (3.1 × 8 = 24.8).
- The growth curve climbed from 0.48x on May 3 to 3.14x on August 15 — the final data point is 3.1415. Pi. Either a coincidence or a deliberately chosen cutoff date.
- The median researcher spent over $600 per day on agents (at API list prices); the top 10% exceeded $7,000 per day.
- Lines changed per person in Q3 2026 were 7.02x pre-2025 levels; output tokens grew roughly 124x since November 2025.
- Model lineup: Codex, GPT-6 Astra, plus GPT-5.6 Sol as a classifier.
A self-undermining caveat
OpenAI concedes that over half of the "successful" tasks in the 4–8 hour research tier required at least one human intervention to complete. And "success" was self-judged by GPT-5.6 Sol, a scorer validated on only 25 tasks. n = 25.
Milestones — and an admission
OpenAI states the "AI research intern" tier was reached in September 2026, with a full AI researcher targeted for March 2028. The same post admits research progress will "likely not keep pace with these metrics' growth" — inputs grew 124x, but outputs did not grow 124x.
Reactions
Simon Willison published a lengthy commentary the same day, opening with "Today is OpenAI's RSI day" — recursive self-improvement, using AI to accelerate research that improves AI. He also noted an unexplained spending spike in late July, guessing it reflected early internal access to GPT-6 Astra. The same day, Pachocki published *An Alien Mind*; the two pieces read well side by side — one tallies the bill, one reflects.
In a 196-point Hacker News discussion, the top-voted critique was plain: usage ≠ impact; burning more tokens doesn't mean better papers. Another repeatedly upvoted comment pointed at the circular structure — OpenAI says it uses agents to supervise agents against loss of control, while this post says agents are accelerating the creation of the next generation of agents. The supervisor descends from the supervised; that structure deserves more scrutiny than any single number.
This post pairs with OpenAI's March internal oversight report (discussed here on September 7: AI doing the work, AI standing guard). Two sides of the same coin: how capable, expensive, and human-dependent the workers actually are. Together they form the full picture.
> Sources: openai.com/index/research-acceleration-view-inside-openai (2026-09-06); simonwillison.net commentary 2026-09-06; HN 49587217. All figures cross-checked against the original post and Wayback archives.