TL;DR: OpenAI’s new paper “The Shift to Agentic AI: Evidence from Codex” documents an inflection point: 97.9% of OpenAI employees now use Codex, active users grew 5× in H1 2026, and non-developer usage exploded — 137× for individuals, 189× for organizations since August 2025. Co-authored with academics from Columbia, Wharton, and Duke, it argues agentic AI has crossed from curiosity into enterprise-wide transformation.
Introduction
On June 25, 2026, OpenAI published “The Shift to Agentic AI: Evidence from Codex” (Source: arXiv — 2606.26959) — a data-rich study of how its Codex platform is reshaping work. Co-authored with researchers from Columbia, Wharton, and Duke, the paper uses internal telemetry to argue that 2026 marks the year agentic AI reached critical mass. Thomas Claburn’s coverage for The Register captured the headline: nearly every OpenAI employee now relies on agents (Source: The Register — OpenAI says 97.9 percent of its employees are now using agents).
Key Findings
Near-Universal Internal Adoption
OpenAI’s workforce has gone all-in. Codex adoption rose from ~40% in August 2025 to 97.9% by mid-2026 — a near-saturation rate signaling the tool is no longer optional but essential infrastructure.
User Growth Across the Board
Active users grew 5× in H1 2026. More striking is where growth comes from: non-developer usage surged 137× for individuals and 189× for organizations since August 2025, reflecting adoption far beyond engineering into legal, operations, marketing, and HR.
Intensity of Use Is Accelerating
It’s not just more users — they’re using Codex more deeply. The share of users submitting 8+ hour tasks increased nearly 10× since the start of 2026, signaling growing trust in agents for long-running autonomous work. The median legal-role user now generates 13× more monthly output tokens than in November 2025.
What This Means
The data points to a genuine tipping point. When 97.9% of a company’s employees use agents and non-technical roles generate 13× more output than six months prior, agentic AI is no longer an experiment — it’s the default mode of knowledge work. The academic collaboration lends independent credibility to what might otherwise read as a vendor case study. Organizations on the sidelines should treat this as a wake-up call: the shift isn’t coming — it already happened.
FAQ
Is Codex just for developers? No. The most dramatic growth — 137× individuals, 189× organizations — comes from non-developer roles. Legal, marketing, and operations are the fastest-growing segments.
What is an “8+ hour task”? A complex, multi-step objective delegated to Codex that runs autonomously for eight or more hours. The 10× increase shows users trust agents with high-stakes, long-duration work.
How reliable is internal data? The paper uses telemetry — actual token counts, session durations, task metadata — not surveys. External academic co-authors from Columbia, Wharton, and Duke add rigor.
Where can I read the full paper? On arXiv and the OpenAI blog. The PDF is at OpenAI’s CDN.