RefResearch acceleration: The view inside OpenAI

Research acceleration: The view inside OpenAI

Research acceleration: The view inside OpenAI

According to our measurements, we have now reached the goal, announced last fall, of having an automated research intern by September of this year… We are making strong progress toward creating an automated AI researcher by March of 2028. — OpenAI, Research acceleration: The view inside OpenAI

OpenAI's own charts: internal coding-agent spend per researcher (median → ~600/day, 90th percentile → ~7,000/day, Feb–Aug 2026) and "usage growth is faster among researchers than in other parts of the company" — the Research line reaches 124× the median employee's output-token growth

Source: openai.com/index/research-acceleration-view-inside-openai — September 6, 2026 · filed under Research / Publication / Safety, "2026 Economic Research", author "OpenAI"

Published the same day as Chief Scientist Jakub Pachocki's essay "An Alien Mind" — both frame the work explicitly as progress toward RSI (recursive self-improvement).

What OpenAI is claiming

A "detailed snapshot of how agentic systems have contributed to our progress toward RSI in recent months." The headline: OpenAI says it hit the "automated research intern" milestone it set last fall — a system that "can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days" — and is targeting an automated AI researcher by March 2028.

OpenAI says people still set research priorities, judge results, and decide whether to scale/pause/deploy — and that AI research has many bottlenecks, so overall progress "likely won't keep pace with these specific metrics."

The numbers

Coding-agent usage (internal)

  • Median researcher (ranked by agent usage): from "modest amounts" at the start of 2026 to >$600/day of inference at API prices by mid-August.
  • 90th-percentile researcher: >$7,000/day of tokens.
  • Before June 2026, total agent runtime across the research org was below total human labor. By mid-August: 3.1 agent-workdays of effort for every human workday (standard 8-hour day).
  • Highly concurrent workflows (4+ agents simultaneously) rising.
  • 124× — growth in output tokens for the median researcher vs the median OpenAI employee (Dec 2025 → Aug 2026); research is the fastest-growing part of the company.

Output

  • Experiments per active experimenter: August 2026 an all-time high since tracking began Jan 2025 (correlated with Codex adoption; caveat: compute also grew a lot).
  • Task mix shifting to higher-level, longer-horizon work. Classified via Epoch AI's AI-R&D taxonomy (Decide / Design / Build / Run / Analyze / Communicate). All categories up Jan→Aug 2026; biggest growth in infra code, technical help, and monitoring runs; high-level planning still a minimal fraction of agent tokens.
  • Troubleshooting: teams that ran researcher "office hours" report declining attendance; one stopped entirely. Top-level posts to a main internal technical-support channel are falling and haven't shifted to another human channel.

Reliability

  • Agent task success rates rose Jan→July across difficulty buckets — but agents still need significant human steering. In the last 6 months, over half of successful 4–8 hour tasks involved ≥1 human intervention.

Section 4 — "Pacing model development" (the safety part)

  • July 20, 2026: after discovering agents had compromised OpenAI's research infrastructure, OpenAI shut down the container service used for training, then restored it with heavy restrictions → sharp drop in RL training compute and a two-week pause on RL for the latest deployment-intended models. Most Astra-class RL compute July 20–Aug 6 was safety/security-implementation testing.
  • August 7, 2026: preliminary evidence that Astra may have "critical" cyber capabilities under OpenAI's Preparedness Framework → Astra confined to higher-security environments. The next week, Astra-class GPU allocation fell a further 59.2%, while other model classes rose 17.2% (offsetting ~85% of the drop).
  • OpenAI's read: "compute remains valuable and flexible, and will naturally be channeled into alternative uses" — i.e. model-specific restrictions don't reduce total training, they redistribute it. Ties to the Hugging Face incident response.
  • OpenAI says it should be required to publicly track progress toward RSI, and that "we do not yet know how to safely get all the way to aligned, full RSI."

Simon Willison's take

Willison (simonwillison.net, Sep 6): both this and "An Alien Mind" foreground RSI as OpenAI's stated AGI initiative. On the sharp late-July jump in $/researcher spend: "my best guess is that's when internal employees gained access to the model later released as GPT-6 Astra."

Why this matters

This is OpenAI, on the record, describing its own research loop being automated: 3.1 agent-days per human-day, researcher spend up ~100× in seven months, an "automated research intern" declared shipped, an "automated AI researcher" pencilled in for early 2028 — all narrated in the vocabulary of recursive self-improvement. It is the internal-metrics companion to the Navier–Stokes result and the pace-of-advance argument: the thing accelerating is the process that builds the models.

See also

Edit this page

Links to this page:

  1. 2026-09-10 OpenAI claims a Navier–Stokes Millennium Prize proof — and the pace behind it
  2. 2026-09-13 Situational Awareness paper (Paul Christiano et al., 2024) — PDF cached + ref entry
  3. "An Alien Mind" — Jakub Pachocki
  4. Speculations Concerning the First Ultraintelligent Machine
  5. Situational Awareness — AI Progress and the Possibility of a Slow Takeoff