---
id: "lib-2026-06-25-the-shift-to-agentic-ai-evidence-from-codex"
type: "library"
title: "The Shift to Agentic AI: Evidence from Codex"
lang: "EN"
org: "OpenAI"
source_type: "Paper"
date: "2026-06-25"
week: "2026-W26"
url: "https://nosov.ai/library/lib-2026-06-25-the-shift-to-agentic-ai-evidence-from-codex/"
original_url: "https://cdn.openai.com/pdf/5d1e1489-21c0-43e4-9d42-f87efdbf0082/the-shift-to-agentic-ai-evidence-from-codex.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Large-scale usage data from OpenAI's own platform, analysed by OpenAI's economic research team with Columbia, Wharton and Duke economists, but every figure comes from the vendor's logs and two academic co-authors were paid contractors."
summary_basis: "full-text"
original_minutes: "60"
tags: ["agentic ai", "workflow redesign", "parallel agents"]
---

# The Shift to Agentic AI: Evidence from Codex

Tier A (Strong evidence): Large-scale usage data from OpenAI's own platform, analysed by OpenAI's economic research team with Columbia, Wharton and Duke economists, but every figure comes from the vendor's logs and two academic co-authors were paid contractors.

Summary based on the full text.

## In short

A June 2026 paper by Drew Johnston, Aaron Chatterji and colleagues at OpenAI, with David Holtz (Columbia) and Prasanna Tambe (Wharton), using Codex usage logs to measure the move from chat-style AI to agents that act on a user's behalf. It compares three populations: individual subscribers, organizational accounts, and OpenAI's own staff. Weekly active Codex users grew more than fivefold in the first half of 2026, but adoption is very uneven. Inside OpenAI, Codex has almost fully replaced ChatGPT for work; outside, it remains a minority tool whose heavy users behave like managers of parallel workstreams.

## Key numbers

- 99.8%: share of OpenAI workers' output tokens produced in Codex rather than ChatGPT, June 11, 2026
- 17.3%: organizational (Business/Enterprise) users who used Codex in the last 28 days
- 2.1% to 25.6%: individual users sending at least one task estimated at 8+ human hours, Dec 2025 to May 2026
- 28.6%: OpenAI users who ran five or more concurrent agents in the week to June 11, 2026

## What it reports

- Weekly active Codex users grew more than fivefold between January 1 and June 1, 2026.
- Codex share of output tokens across Codex and ChatGPT on June 11, 2026: 99.8% for OpenAI workers, 63.3% for organizational users, 16.5% for individual users.
- Fewer than 1% of active individual users used Codex in the last 28 days, against 17.3% of organizational users.
- In organizational accounts the average engineer produces 26.8% of tokens on Codex; legal users only 1.9%, rising to 17.6% among legal users who have adopted it.
- Users sending at least one task estimated at one hour or more of expert human work: 35.4% in December 2025, 70.2% in May 2026. For tasks of eight hours or more: 2.1% to 25.6%.
- Skill use (shared reusable instructions) rose from 5.4% on March 1 to 26.6% on June 11, 2026; inside OpenAI 96.2% of active users invoke a skill.
- Concurrency: 67.4% of organizational and 63.9% of individual users never run parallel turns; inside OpenAI only 10.7% run a single workflow and 28.6% run five or more agents.
- The median OpenAI employee had Codex running 2.5 hours on June 11, 2026; the 99th percentile ran about 71 agent-hours per day, up nearly 88% since April 7.
- Median monthly output tokens per OpenAI employee from November 2025 to June 2026 rose 13x in legal roles and more than 50x for researchers; at least 10x in every function.

## How they got it

Automated, privacy-preserving classifiers label requests by persona, task type and estimated human completion time, so researchers do not read messages. Most estimates use a 4% sample of users; task-duration estimates use a 0.1% sample of individual accounts that opted into training data. The persona classifier was validated against HR job titles for a small sample of OpenAI employees (over 90% agreement for engineers and sales).

## Limits

- All data comes from OpenAI's own product; the authors include OpenAI staff and two paid contractors. No independent replication.
- Output tokens and estimated task hours measure activity, not productivity, quality or business value.
- OpenAI is an extreme case: unlimited internal usage, high buy-in and staff close to the models. The authors state it is not representative.

## Why it matters for companies

The gap between 99.8% inside OpenAI and 17.3% in client organizations shows that access to the same model does not produce the same use. The authors attribute the difference to context: file and system access, management expectations, skills and review processes. For a traditional company, the planning question shifts from licence counts to which work can be delegated, who reviews it, and how shared instructions get written and reused. Roles built on supervising several agents at once are already common at the frontier and will affect team design and career ladders.

Original: https://cdn.openai.com/pdf/5d1e1489-21c0-43e4-9d42-f87efdbf0082/the-shift-to-agentic-ai-evidence-from-codex.pdf