OpenAI Codex: 99.8% of Worker Tokens in 2026

What did OpenAI's Codex study find?

OpenAI's Codex generated 99.8% of output tokens used by its own workers across Codex and ChatGPT as of June 11, 2026, according to a research paper published June 25. The paper is titled "The Shift to Agentic AI: Evidence from Codex." It analyzes usage data across three groups: individual users, organizational users, and OpenAI's own workforce.

Codex is an agentic coding and work platform built by OpenAI. It was originally designed for software development but is now used for tasks including drafting documents, analyzing data, and coordinating communication.

How fast is Codex growing?

Codex now has more than 5 million weekly active users, according to Axios. That is up more than six times since OpenAI launched the desktop app in February. The paper also says Codex usage grew more than fivefold in the first half of 2026.

Despite that growth, the paper notes that agentic tooling remains much less broadly used than ChatGPT overall.

Who is using Codex — and how does it differ by group?

The study compares three user populations. Their Codex share of combined Codex-and-ChatGPT output tokens differs sharply:

The paper says OpenAI is an unusually favorable environment for agentic AI. Workers are familiar with frontier models, usage is cheap at the margin, and organizational buy-in is high. "OpenAI usage is therefore not representative of the typical organization today," the paper states.

You might also like

Agility Digit: Live at Amazon, GXO & Toyota

Nvidia Banned Chips Double in Price on China

CAISI Signs AI Reviews With Google, Microsoft

Meta Smart Glasses Switch to Muse Spark AI

How are non-developers using Codex?

This is where the study's most striking numbers appear. Non-developer users rose:

• 137x among individual users since August 2025
• 189x among organizational users since August 2025
• 12x within OpenAI since August 2025

Knowledge workers now make up roughly one-fifth of Codex users and are growing more than three times as fast as developers, per the Axios report. Inside OpenAI, every department — including Legal, Finance, and Recruiting — now uses Codex as its primary AI tool for work.

The paper says non-developers are using Codex for tasks such as documentation, data analysis, and communication, not only code generation. Andrew Hall, a Stanford Graduate School of Business professor, told Axios that he and his students use coding agents like Codex and Claude Code for boilerplate academic tasks, data collection, statistical analysis, and running code to process data.

How does Codex differ from ChatGPT in the study?

The paper separates Codex from ChatGPT for analytical purposes. It calls ChatGPT a conversational AI tool and Codex an agentic AI tool. The authors note the distinction is not absolute — ChatGPT has agentic features like browsing and code execution, and some Codex interactions are conversational.

Still, the paper says Codex use is "strongly oriented toward delegated production." Users ask it to debug, refactor, validate changes, configure applications, draft documents, and analyze data. "These activities are better understood as production than as consultation," the paper says. "Users are asking Codex to do work, not only to provide advice or information."

The paper frames the broader shift this way: "Agentic AI changes the unit of knowledge work from single interactions to delegated, long-horizon tasks."

What tasks is Codex being used for?

Software development remains the core use case. But the paper's task categories extend well beyond coding. Codex is being used for:

• Debugging, refactoring, and validating code changes
• Configuring applications
• Drafting documents
• Analyzing data (including financial analysis)
• Coordinating communication

One chart in the paper breaks down Codex output tokens by inferred OpenAI department. Engineering and coding work remained prominent, but Finance/Biz Ops, Product/Marketing/Ops, and other departments also showed substantial shares of knowledge work and data analysis.

Here's what we know so far: the paper is careful to say OpenAI's internal data is not representative of typical enterprise adoption. It is presented to illustrate the broader point that Codex usage extends beyond code generation — not to claim that most organizations are at the same stage.

What is Codex connecting to?

Codex can connect to email, calendar, documents, spreadsheets, design apps, and messaging apps like Slack and Teams, according to the Axios report. The paper describes agentic systems as tools that can inspect files, execute commands, and create or modify artifacts — distinct from traditional chatbot interfaces.

The Axios report also notes that Anthropic's Claude Code and Cowork were the first agentic tools to attract non-coders at scale. A growing number of power users report feeling mentally fatigued as they try to supervise several fast-moving AI workstreams at once.

The OpenAI research paper was published on June 25, 2026, alongside a post on OpenAI's website. Builders tracking the shift to agentic AI tools or watching how AI content authenticity questions play out across industries will find the usage breakdowns in the full paper worth examining. Those following infrastructure for AI workloads may also note the scale implied by 5 million weekly active users generating tokens at these rates.