OpenAI Says Agents Now Cover 3.1 Workdays Per Researcher Workday

OpenAI’s research organization now runs 3.1 agent-workdays of effort for every workday of human labor, the company said in a September 8, 2026 essay that lays out how it intends to convert frontier model capability into revenue. The figure, measured against a standard eight-hour workday as of mid-August 2026, anchors a broader argument by chief financial officer Sarah Friar that more capable models, a user base of more than one billion weekly actives, and a full-stack compute strategy reinforce one another.

The essay, titled “The Work Now Within Reach,” is Friar’s latest in a series on OpenAI’s business and compute economics. OpenAI appointed Friar as CFO in June 2024; she previously served as CEO of Nextdoor and as CFO of Square.

Usage and Revenue

OpenAI said its products reach more than one billion weekly active users and 2.5 million businesses, and that each research advance improves ChatGPT, ChatGPT Work, Codex, and applications built on its API. The company said free access supported by advertising helps people discover where AI is useful, while subscriptions and usage-based offerings let customers spend more as they find more value.

Citing its study of individual ChatGPT plans, OpenAI said daily message volume was roughly 50% higher six months after signup than in the first month, and that users had tried roughly twice as many distinct tasks. The underlying analysis draws on a 0.1% sample of accounts created between October 15, 2025 and May 1, 2026, with activity tracked through May 31, 2026; it measures messages and task breadth relative to each user’s first 28 days and sorts messages into 53 capability categories.

The essay also points to five customer deployments. Boston Children’s Hospital reported more than 40 diagnoses in previously unresolved rare disease cases through AI-assisted research. Cars24 said its agents handle more than one million conversation minutes a month across car buying and selling. Circles reported 65% autonomous resolution of customer service interactions across supported workflows, and Balyasny Asset Management said its Central Bank Speech Analyst cut macroeconomic scenario analysis from two days to about 30 minutes. Replit offers a Free Mode that lets people plan software without consuming their usage allowance. Each figure comes from the customer’s own account as presented by OpenAI.

Agents Inside the Research Organization

The 3.1 agent-workdays figure comes from an OpenAI research update published September 6, 2026, which reported that before June 2026 total agent runtime across the research organization was still below total human labor. That update said the median researcher was using more than $600 per day of inference at API prices by mid-August, and the 90th percentile user more than $7,000 of tokens per day.

The same update said coding-agent success rates rose from January to July 2026 across several task-difficulty buckets, while noting that longer tasks still required human steering: over the prior six months, more than half of successful tasks estimated at four to eight hours of human time involved at least one intervention. In the essay, OpenAI said people still set research priorities and judge results, and that teams are using agents to resolve infrastructure problems that once required specialist support.

Model Capability and Compute Costs

Friar’s essay frames GPT-6 Astra as the current capability driver, describing it as state-of-the-art in computer use, browsing, software engineering, cybersecurity, science, and professional work. In its release post, OpenAI said Astra scores 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench, and that it meets the Critical threshold in cybersecurity under the company’s Preparedness Framework. Those are OpenAI’s own reported results. The company said Astra began rolling out to a limited set of organizations on September 3, 2026, with general availability across ChatGPT paid tiers, the API, Microsoft Azure, and AWS Bedrock in the following days.

On serving economics, the essay reiterates two figures from OpenAI’s July 29, 2026 engineering post on GPT-5.6: production serving software improvements that reduced end-to-end serving costs by 20%, and speculative-decoding work that increased token-generation efficiency by more than 15%. Both gains were achieved with GPT-5.6 Sol operating inside Codex, including autonomously rewriting production kernels, the engineering post said.

The essay also recaps first results for Jalapeño, OpenAI’s custom inference chip, first published August 25, 2026. In InferenceX tests across GPT-OSS 120B, DeepSeek R1, and Kimi K2.5, OpenAI said the chip delivered 1.5 to 1.9 times as much peak token throughput per watt as the commercial systems tested, using rated chip power to normalize the comparison, with end-to-end latency 1.7 to 3.6 times lower. OpenAI said it plans to begin deploying Jalapeño in its compute infrastructure by the end of the year alongside accelerators from NVIDIA, AMD, and other partners, and that second- and third-generation versions are in development.

Friar wrote that OpenAI judges each investment by the demand it can serve, how quickly it becomes productive, and whether the returns justify the capital committed, and that revenue from growing adoption funds further research and infrastructure.