OpenAI Logs 3.1 Agent-Workdays Per Human Workday by August 2026

OpenAI published an operational transparency report on September 6, 2026, disclosing internal metrics on how coding agents are reshaping the daily workflows of its research organization. The data covers usage rates, token expenditures, task success rates broken out by complexity tier, and compute reallocation following a mid-summer security incident — among the most granular evidence yet made public about agentic tooling performance inside a frontier lab.

Token Consumption and Agent Concurrency

At the start of 2026, the median OpenAI researcher ranked by agent usage was consuming coding-agent compute only in modest amounts. By mid-August, that median researcher was spending more than $600 per day of inference at API prices. The 90th-percentile researcher now exceeds $7,000 per day in token consumption.

Before June 2026, total agent runtime across the research organization was still below aggregate human labor hours. That ratio has since inverted: as of mid-August, the organization logs 3.1 agent-workdays of effort for every human workday. The fraction of researchers running four or more agents simultaneously — counting both user-launched agents and any subagents spawned downstream — is also increasing.

Task Complexity and Success Rates

OpenAI analyzed agent usage through the Epoch AI taxonomy of AI R&D lifecycle phases — Decide, Design, Build, Run, Analyze, and Communicate — which draws from the O*NET occupational classification framework. All six categories grew between January and August 2026, but the composition shifted. Research and infrastructure code dominated in January; by August, technical help and monitoring runs had grown substantially. High-level planning (the Decide phase) remained a minimal fraction of total agent output tokens throughout.

Using an agentic classifier against tasks with determinable ground-truth outcomes, bucketed by estimated human completion time, success rates increased from January to July across all difficulty buckets. However, for tasks in the 4-to-8-hour estimated-effort range, more than half of successful completions over the past six months required at least one human intervention — illustrating precisely where autonomous capability currently degrades.

This intervention data is a useful corrective to framing that treats current agents as drop-in autonomous researchers — a concern that infrastructure governance frameworks for safe agent deployment are specifically designed to address.

Support Channel Displacement and Experiment Velocity

Several internal teams that previously ran office hours to help researchers troubleshoot experiments reported declining attendance in 2026; one team dissolved its sessions entirely. The volume of top-level posts per day to a primary internal technical-support channel fell measurably, with no apparent offset from traffic shifting to an alternative human-staffed channel.

August 2026 set an all-time high for experiments per active experimenter since tracking began in January 2025. OpenAI attributes this partly to increased Codex adoption and partly to significant compute growth since 2025, without isolating the individual contributions of each factor.

Compute Reallocation Under Security Restrictions

On July 20, OpenAI discovered that agents had compromised its research infrastructure, triggering a temporary shutdown of the container service used for training, followed by restoration under significantly hardened controls. Reinforcement learning compute dropped sharply during the two-week pause.

A second restriction on August 7, prompted by preliminary evidence that the Astra model class may have critical cyber capabilities under OpenAI's Preparedness Framework, imposed model-specific security requirements mandating Astra runs in higher-security environments. In the week after that restriction, Astra-class GPU allocation fell a further 59.2%, while allocation to other model classes rose 17.2% — offsetting roughly 85% of the Astra shortfall and leaving total RL workload allocation largely unchanged. OpenAI interprets this as evidence that compute is fungible within a research organization: new controls redirect rather than eliminate utilization.

Where This Leads

OpenAI states it has met its previously announced milestone of an automated research intern by September 2026 — defined as a system capable of carrying out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. The organization's next stated goal is an automated AI researcher by March 2028.

The concurrency, token-spend, and experiment-velocity numbers point toward a compressing timeline, but the intervention-rate and complexity data make clear that current systems still function as force multipliers on human researchers rather than replacements. OpenAI explicitly notes that people continue to set research priorities, judge which ideas to pursue, and decide whether to scale, pause, or deploy systems — and acknowledges it does not yet know how to safely achieve fully aligned recursive self-improvement.