Saturday, November 21, 2026← Front page
Also this week

OpenAI research acceleration report

OpenAI published a follow-up report, "Research acceleration: The jagged frontier," giving an update on how coding agents have accelerated and changed AI development inside the company, based on its own internal measurements.12 OpenAI explains that it published this research because it holds that AI labs ought to be obligated to monitor and openly disclose signs that their systems are improving themselves recursively. A Latent Space roundup characterized the update as calling for labs to release such signals on a regular basis. inference34

OpenAI says the median researcher now uses about $2,400 a day of agentic AI tools at API list prices, and the 90th-percentile researcher about $23,000, up 62% and 105% from $1,500 and $11,000 in September.5 As of September 2026, OpenAI says the API-equivalent value of agent use was about twice researchers' non-equity compensation, yet internal deployment accounts for only 2% of total research compute.67

Zvi Mowshowitz notes that in OpenAI's report, work done by agents in a day exceeds human working days by roughly eight times for pretraining and post-training, and by 5.7 times across research as a whole.8 OpenAI's estimated zero-intervention success rates vary by task type: 69% for Communicate, 58% for Analyze and 43% for Build between August 1 and October 18, 2026, and as low as 25% for writing inference and serving code.910 OpenAI also says Safety and Alignment saw the largest absolute increase in documented research effort, with workweeks up 77% from January to September.11

Mowshowitz appreciated the report but found it more a data dump of metrics than an investigation telling a story, and says he is not sure he learned or updated much.12

What each sentence rests on
  1. Today, we’re providing an update on the state of research acceleration at OpenAI, based on internal measurements of how coding agents have accelerated and changed the AI development cycle. — OpenAI, Research acceleration: The jagged frontier
  2. OpenAI published an update on internal agent use and research task completion — Latent Space, [AINews] Gemini 4 Neon: GDM's Opus/Sol Contender
  3. we believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI) — OpenAI, Research acceleration: The jagged frontier
  4. proposed regular disclosure of indicators for AI research acceleration. — Latent Space, [AINews] Gemini 4 Neon: GDM's Opus/Sol Contender
  5. the median researcher now uses around $2,400 USD per day on agentic AI tools, and the 90th percentile researcher now around $23,000 USD per day, up 62% and 105% respectively from $1,500 and $11,000 in September. — Zvi Mowshowitz, AI #195: Opening Bell
  6. the API-equivalent value of agentic AI use in OpenAI’s research organization was about twice its non-equity compensation — OpenAI, Research acceleration: The jagged frontier
  7. internal deployment, which combines internal Codex, ChatGPT, and API usage, accounts for only 2% of total research compute. — OpenAI, Research acceleration: The jagged frontier
  8. ‘Agent workdays’ outnumber the human workdays by ~8x for pretraining and post-training, and 5.7x overall for research. — Zvi Mowshowitz, AI #195: Opening Bell
  9. estimated success rates without human intervention was higher for Communicate (69%) and Analyze (58%) than for Build (43%). — OpenAI, Research acceleration: The jagged frontier
  10. Quote not reprinted: this page already repeats as much of this article as we allow. — OpenAI, Research acceleration: The jagged frontier
  11. Quote not reprinted: this page already repeats as much of this article as we allow. — OpenAI, Research acceleration: The jagged frontier
  12. it felt like a data dump of metrics rather than an attempt to do an investigation and tell a story. I’m not sure that I learned or updated much. — Zvi Mowshowitz, AI #195: Opening Bell

Every sentence was checked word for word against quotes from the articles it cites. Each footnote shows its quote, unless the page already quotes as much of that article as we allow; then it links the article instead. A model wrote and arranged the sentences, and a second read checked each against its quotes (written 2026-11-20).

The record: who reported it, checked facts, all quoted views, timeline →