IMPLICATOR
Bloomberg reported Tuesday that OpenAI planned to announce 10 million people using Codex and ChatGPT Work, a company-supplied count for its workplace agents. The figure combines Codex, its coding agent, with ChatGPT Work, a newer agent for broader office tasks, and was nearly twice the level earlier in July. The materials reviewed for this article do not provide a product split, an active-user window or an independent audit of the tally.
ChatGPT Work is the part of ChatGPT built for longer research, analysis and finished deliverables, while Codex remains dedicated to software development, according to OpenAI's help-center page. The count arrives less than two weeks after The Implicator's July 10 coverage noted OpenAI's earlier claim that Codex alone had more than 5 million weekly users. OpenAI says ChatGPT itself has more than 900 million weekly users, a much larger base than the agent products now being measured.
Key Takeaways
- OpenAI says Codex and ChatGPT Work reached 10 million combined users, nearly double the level earlier in July.
- The materials reviewed provide no product split, active-user window or independent audit of the tally.
- A June study found Codex use among 17.3% of active organizational users, versus fewer than 1% of active individual users.
- Epoch's pull-request analysis and Andrew Hall's audit show why agent output still needs careful interpretation and expert review.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
The 10 million-user tally
Neither Bloomberg's report nor the OpenAI help pages reviewed for this article separate Codex users from ChatGPT Work users in the combined tally. They also do not say whether the count means weekly active users, monthly active users, people who started one agent task or people who completed reviewed work. OpenAI's help page says Work and Codex draw from the same agentic usage and credit pool and that the amount used depends on task size, complexity, model and surface.
The June data covered Codex only. In a June 25 paper, researchers using OpenAI data reported that weekly active Codex users increased more than fivefold between January 1 and June 1. In the last 28 days of the observation period, 17.3% of organizational users who were active on ChatGPT or Codex used Codex, compared with fewer than 1% of active individual users. The study also warned that OpenAI workers were an unusually favorable environment, with high model familiarity, low marginal cost and broad internal support.
The June 25 Codex paper
Researchers from OpenAI, Columbia, Duke and the University of Pennsylvania wrote the study. In the seven-day window ending June 11, 26.6% of active Codex users invoked at least one skill, the reusable instructions and tool connections that let workers repeat a task. As of June 11, Codex accounted for 99.8% of output tokens generated across Codex and ChatGPT by OpenAI employees, versus 63.3% among organizational users and 16.5% among individual users.
OpenAI provided the usage data. In May, the researchers estimated that 25.6% of individual Codex users delegated at least one task that would take an experienced human eight hours or more. That result came from a 0.1% random sample of individual accounts whose users opted into training. A model assigned the duration; the study did not observe time saved. The researchers benchmarked their classification prompt against Codeforces problems and describe the result as a measure of task complexity rather than a productivity audit.
Get Implicator.ai in your inbox
Strategic AI news from San Francisco. No hype, no "AI will change everything" throat clearing. Just what moved, who won, and why it matters. Daily at 6am PST.
No spam. Unsubscribe anytime.
Epoch analyzes 7,524 Codex pull requests
In a July 7 analysis, Jaeho Lee and Thomas Kwa examined 7,524 merged pull requests from OpenAI's public Codex repository between mid-April 2025 and mid-June 2026. Their main comparison focused on 41 official repository collaborators and excluded automated accounts. In the partial second quarter of 2026, 8.2% of contributor-days were assigned at least 24 hours of estimated unassisted engineering effort, versus 2.0% in the partial second quarter of 2025. Epoch wrote that the shift was "consistent with growing AI uplift within software engineering," while warning that three LLM judges produced the estimates, the result was not causal evidence and the time-saved figure should be treated as an upper bound. Longer or more complex merged contributions were not necessarily more valuable. The analysis did not examine output quality or validate OpenAI's 10 million-user claim.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
Andrew Hall, a Stanford Graduate School of Business professor, gave Axios a separate example of manual review. Hall said he and his students used coding agents including Codex and Claude Code for data collection, statistical analysis and academic work. When he asked Claude Code to update a paper on universal vote by mail, it missed data and miscoded some of what it collected after producing figures, tables and a draft. "It very much needed an expert, Ph.D.-level student to oversee it quite closely," Hall said.
Frequently Asked Questions
What does OpenAI's 10 million-user figure include?
It combines users of Codex, OpenAI's coding agent, and ChatGPT Work, its agent for longer research, analysis and office deliverables. The company did not provide a product-by-product split in the materials reviewed for this article.
Does 10 million mean weekly active users?
That is not clear. Bloomberg's report and the OpenAI help pages reviewed here do not define an activity window or say whether the count covers people who started a task, completed one or used either product during a particular week or month.
How widely was Codex used by outside organizations?
In the last 28 days covered by a June study, 17.3% of organizational users who were active on ChatGPT or Codex used Codex. The comparable share among active individual users was below 1%.
What did Epoch AI measure?
Epoch analyzed 7,524 merged pull requests from OpenAI's public Codex repository. It estimated at least 24 hours of unassisted engineering effort on 8.2% of contributor-days in partial second-quarter 2026 data, versus 2.0% in the partial year-earlier quarter. Epoch called the estimate noncausal and an upper bound.
Why does expert review still matter?
Stanford professor Andrew Hall told Axios that Claude Code missed data and miscoded some collected data while updating an academic paper. A graduate student had to audit the work closely.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
