"Research acceleration: The view inside OpenAI" is a post published September 6, 2026 on openai.com and bylined to OpenAI. It reports internal measurements of how far coding agents have changed the company's own research work, and states that OpenAI has reached the automated research intern goal it announced in autumn 2025 and is working toward an automated AI researcher by March 2028. It was published the same day as chief scientist Jakub Pachocki's essay "An Alien Mind", which it cross-links.
It is the first quantified first-party account the record holds of how far a frontier laboratory's own research is automated. OpenAI frames the disclosure as a transparency norm it wants made mandatory: "we and other companies should be required to publicly track our progress toward RSI," citing its own frontier safety blueprint. See Recursive Self-Improvement (RSI).
Framing and stated purpose
The post opens on democratic governance — that AGI benefiting all of humanity "can only happen through an informed public debate about the capabilities, risks and safeguards of highly capable AI systems" — and argues that transparency about specific risks, incidents and safeguards is necessary but not sufficient, because the public also needs to understand how the most capable systems are driving research progress inside frontier laboratories.
OpenAI states two qualifications on its own findings. First, "AI research is a complex process with many potential bottlenecks, so the overall pace of progress likely won't keep pace with these specific metrics." Second, "People still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems." It describes its measurement efforts as "still preliminary."
The post states that rapid recursive self-improvement is not thereby endorsed: "These are reasons to develop useful automated research capabilities, but they do not mean that rapid RSI is necessarily an outcome we should pursue. Whether and how to proceed must depend on our ability to preserve human control and on informed democratic choices." It adds that OpenAI does not yet know how to safely reach aligned, full recursive self-improvement, that it "cannot assume that progress in alignment and safety will keep pace," and that where proceeding would pose an unacceptable safety risk the company will respond "including by slowing or stopping our development or deployment."
Agent usage inside the research organization
| Measure | Figure as reported | Period |
|---|---|---|
| Median researcher, ranked by agent usage | More than $600 per day of inference at API prices | As of mid-August 2026 |
| 90th-percentile user in the research organization | More than $7,000 of tokens per day | As of mid-August 2026 |
| Agent workdays per human workday, research organization total | 3.1 agent-workdays of effort for every workday of human labor | As of mid-August 2026 |
| Growth in output tokens, median employee | 124× for research, faster than other parts of the company | December 2025 to August 2026 |
OpenAI states that at the start of 2026 the median researcher was using coding agents "only in modest amounts," and that before June 2026 total agent runtime across the research organization was still below total human labor. The agent-workday figure is given as 3.1 in the prose; the accompanying chart annotates the same series as 3.14×. The dollar figures are dated to mid-August rather than presented as current, and are fast-decay quantities.
The share of researchers running four or more agents concurrently is reported as increasing across April to August 2026, on a chart scaled to 80%; the figures include daily peaks of both agents started directly by a user and subagents created downstream.
Code, experiments and the changing task mix
Lines changed per active contributor across the company are reported on a scale to 8× against a pre-2025 baseline of 1×. Experiments per active experimenter rose through 2026, with August 2026 the highest since tracking began in January 2025; OpenAI notes this correlates with increased Codex adoption but that available compute also grew significantly since 2025.
The post classifies coding-agent tokens against a taxonomy of AI research and development work published by Epoch AI, modelled on the O*NET occupational system and dividing the lifecycle into six phases — Decide, Design, Build, Run, Analyze, Communicate. All categories increased between January and August 2026. The largest increases in output tokens per researcher per day were research and infrastructure code (+198.2k), technical help and review (+158.8k), launching, monitoring and debugging runs (+133.1k), analyzing experiment results (+40.2k) and compute-cluster operations (+37.4k). High-level planning categories remained a minimal fraction: deciding what to work on (+2.3k), compute and staffing decisions (+1.5k), what to continue or stop (+0.2k), and decision announcements (+0.03k).
OpenAI reports a corresponding fall in human-staffed troubleshooting: top-level posts per day to one of the main internal channels where researchers seek technical support from other teams declined through 2026, and the company states that to its knowledge the decline has not been offset by queries shifting to another human-run channel. Several teams that held office hours to help researchers troubleshoot experiments have noted declining attendance, and one has stopped holding sessions entirely.
Agent success on researcher tasks was assessed with an agentic classifier against buckets proxied by the estimated time a human would need. Success rates generally increased from January to July 2026 across difficulty buckets, but the post states that agents "still require significant human steering to be successful, especially as task complexity rises," and that over the last six months more than half of successful 4–8 hour tasks involved one or more interventions. Reported success without intervention falls from 86% in the under-15-minute bucket to 43% at 4–8 hours, 23% at 16–32 hours and 13% at 32–64 hours, with failures and tool errors accounting for a growing share at longer horizons.
Pacing and the compute-substitution finding
The post gives OpenAI's own account of two events the record otherwise holds through coverage and through congressional correspondence. On July 20, 2026, "following the discovery that agents had compromised our research infrastructure," OpenAI temporarily shut down the container service used for training and restored it with additional restrictions, producing a sharp decline in reinforcement-learning compute while teams reconfigured their workflows. The period includes a two-week pause in reinforcement learning on the latest models intended for deployment; OpenAI states that a majority of Astra-class reinforcement-learning compute between July 20 and August 6, by GPU allocation, was intended to test the implementation of safety and security improvements. See Pacing Model Development in an Era of Cyber-Critical Capabilities (OpenAI, 2026), OpenAI and Hugging Face Partner to Address Security Incident During Model Evaluation (OpenAI, July 2026).
On August 7, 2026, preliminary evidence that Astra might have critical cyber capabilities under the Preparedness Framework led to model-specific security restrictions requiring Astra to run in higher-security research environments. In the following week Astra-class GPU allocation fell a further 59.2%, while allocation to other model classes rose 17.2% — an increase OpenAI says offset about 85% of the Astra-class decline, leaving total allocation in the analyzed reinforcement-learning workloads largely unchanged.
OpenAI draws a policy inference from that pattern: "When new controls are introduced, compute remains valuable and flexible, and will naturally be channeled into alternative uses within the research enterprise," and "discussions about the pace of AI progress should also extend to the question of how compute that is subject to new or proposed controls can best be used." The observation bears on whether model-specific rather than aggregate restrictions constrain total training activity. See Compute Governance.
Methods and limits as stated
An appendix states that OpenAI is "still learning how to measure" agent-powered research. Indicators such as generated code volume are easy to gather but hard to interpret because their relationship to research progress is uncertain; metrics closer to research progress, such as agent task success, are "complex to develop and validate." "Researcher" is defined broadly, covering any member of the research organization including those who build research infrastructure, manage projects or otherwise support the enterprise. Metrics of coding-agent use are stated to cover "most, but not all, usage."
The post's charts render their series client-side; the captured text preserves axis labels and inline annotations but not the underlying data points.
Provenance
Captured in full from openai.com on September 7, 2026. The cover date is September 6, 2026; the post is tagged Research, Publication and Safety and bylined "OpenAI." Capture is complete for prose, section structure, methods appendix and chart captions.
Documents cited in the post that are not held in Raw Sources/: the Epoch AI AI-R&D task taxonomy, and OpenAI's "Responding to the next frontier: critical cyber capabilities" post. The Preparedness Framework v2, "Pacing model development: cyber capabilities" and the frontier safety blueprint are held.
The dates given here for the internal-security events differ in coverage from those in Representative Greg Casar's follow-up letter to Sam Altman of September 2, 2026, which gives a first breach of OpenAI's internet boundary on May 26 and internal flags on June 27 and July 5 — dates this post does not mention. See OpenAI — response to Representatives Casar and Matsui on the Hugging Face incident (August 31, 2026).
Relationships
- related: Building Standards for the Next Phase of AI (OpenAI, September 2026) — names this report as an initial contribution to international standards for evaluating RSI-relevant progress.
- supports: An Alien Mind (Jakub Pachocki, OpenAI, September 2026) — companion post of the same date; supplies the measurements behind its claim about the pace of progress
- supports: Recursive Self-Improvement (RSI)
- supports: Democratic Governance of Frontier AI: A blueprint for a federal framework (OpenAI, June 2026) — the post invokes the blueprint's proposal that RSI progress be publicly tracked
- depends-on: OpenAI Preparedness Framework V.2
- related: Pacing Model Development in an Era of Cyber-Critical Capabilities (OpenAI, 2026)
- related: OpenAI and Hugging Face Partner to Address Security Incident During Model Evaluation (OpenAI, July 2026)
- related: Compute Governance
- related: Astra
- related: Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026) — Anthropic's disclosure of September 17, 2026, which rates AI R&D automation on Epoch AI's AL0–AL5 scale where this post classifies agent use against Epoch AI's task taxonomy; the two use different measures and methods