OpenAI’s latest report: Codex replaces ChatGPT, experts make ‘Agent teams’ work

Has the agent really changed its job?

OpenAI’s latest report, “The Shift to Agenetic AI: Evidence from Codex,” provides the most direct set of data currently available.

This report was jointly completed by researchers from OpenAI, Columbia Business School, Wharton School, and Duke University Fuqua School of Business.

The report analyzed the real usage data of OpenAI Codex and compared three groups of people: individual account users, organizational account users, and OpenAI internal employees.

It not only demonstrates how OpenAI maximizes the use of Codex internally, but also gives people a feeling:

The way AI is being used is undergoing a silent paradigm shift.

From conversing with AI to having AI do the work, from doing it themselves to managing a group of AI employees, this process is already happening, but most people are not aware that it is already happening.

Codex users grow fivefold in six months, but adoption is uneven
The report shows that in the first half of 2026, the weekly active users of Codex will increase by more than 5 times.

However, there is a significant difference in the intensity of Codex usage among individual users, organizational users, and OpenAI internal employees.

The report data shows that the proportion of active users who have used Codex in the past 28 days is only 0.7% among individual users, 17.3% among organizational users, and 97.9% among OpenAI internal employees.

That is to say, among ordinary individual users, Codex is still an early product; In corporate organizations, Codex has begun to spread; But within OpenAI, it is close to universal adoption.

The report specifically points out that within OpenAI, Codex has largely replaced ChatGPT as the main working AI interface.

Because OpenAI is an exceptionally favorable environment, with employees familiar with cutting-edge models, low usage costs, high organizational support, and widespread training and informal knowledge sharing.

Although OpenAI’s use does not represent the typical enterprise of today, it demonstrates a signal:

What will the use of the agent look like when all adoption barriers are removed.

Agent tasks are becoming increasingly difficult, with more and more outputs
The report also analyzed the complexity of tasks assigned by users to Codex and set a clever question for testing:

How long would it take for an experienced human to complete the same task without AI?

The result is astonishing.

By December 2025, only 35.4% of active individual users had sent at least one task that required at least one hour of manual completion.

By May 2026, this proportion will double to 70.2%.

Even more exaggerated is the proportion of users who have sent tasks that require more than 8 hours of manual completion, which has skyrocketed from 2.1% to 25.6%, an increase of more than 10 times.

At the same time, output is also skyrocketing.

In June 2026, the monthly output tokens generated by ordinary OpenAI legal employees through Codex and ChatGPT increased by 13 times compared to November 2025, and the growth of ordinary researchers exceeded 50 times.

The report also found an interesting phenomenon:

In the era of agents, users are delegating the most complex tasks from the beginning. The user first presents the grandest and most complex requirements, and the subsequent conversation only involves continuous refinement and revision.

This is completely opposite to the traditional interaction mode of “asking simple questions first and then gradually delving deeper”.

Codex is no longer just about writing code
Although Codex was originally designed for software development, its use has long exceeded this scope.

Among all user groups of Codex, the most common are still software related tasks such as engineering operations, code implementation, and code understanding.

But Codex is also widely used for non coding tasks such as document writing, data analysis, research, collaborative communication, etc.

Within OpenAI, Codex is also used for research, planning, communication, recruitment, sales, product development, and data analysis.

The wider the scope of use, the deeper the penetration of agents, which seems to be a self reinforcing cycle.

The report also compared the differences in the use of different functions.

In the organization, engineering operations and maintenance account for the largest share of output tokens for most qualification groups, but knowledge artifact tasks have a higher proportion among management and executives.

Within OpenAI, individual contributors tend to lean towards engineering operations, while management tends to lean towards collaborative tasks.

Different roles use agents to do different things, but everyone is using them.

The deepest users are managing a group of AI employees
If ordinary users are still working with AI, then the cutting-edge users are already managing a group of AI employees.

Within OpenAI, only 10.7% of employees run only one agent at a time, and nearly 30% of employees manage 5 or more concurrent agents at a certain point in time.

The report describes a working method of “human supervision of an Agent team”, while assigning tasks to multiple Agents, monitoring progress, and selectively intervening.

Externally, this parallel workflow is not yet common.

About 67% of organizational users and 64% of individual users do not use concurrent agents at all.

But within OpenAI, this group of heaviest users averages about 71 hours of running Codex per day.

For example, an employee can have three Codex agents fix bugs, write tests, and organize documentation simultaneously. If these three tasks each run for 1 hour, the report will be recorded as a cumulative run of 3 hours.

The report shows that since April 7, 2026, the daily running time of the top 1% of heavy users’ Codex within OpenAI has increased by nearly 88%.

External users are also growing, but the intensity is significantly lower.

Among organizational users, the top 1% of users with high usage intensity experienced a daily increase in runtime of approximately 25%.

Among individual users, the daily running time of the top 1% of heavy users has increased by about 50%.

Users begin to encapsulate workflows and reuse them repeatedly
Completing a task with AI at once and encapsulating a workflow for AI to repeatedly execute are two completely different depths.

Codex systematizes workflows through “skills” and “plugins”, which integrate instructions, software, and external tools into a package that can be repeatedly called and even shared across users and organizations.

On March 1, 2026, only 5.4% of active Codex users had called skills. By June 11th, this proportion had reached 26.6%.

The gap is more pronounced when viewed by different groups:

25.7% of individual users and 30.4% of organizational users invoked skills, while within OpenAI, 96.2% of active users invoked at least one skill.

The report specifically mentions that the growth of custom skills is particularly fast, which are skills written by users or organizations themselves, targeting specific team standards or workflows.

This is a deeper change in the way we work.

Once an organization makes up its mind, AI transformation can be very fast
Many people think that new technologies always become popular from the top, but Codex’s data gives the opposite answer.

From a job perspective, engineers are the earliest and deepest users of Codex.

Among organizational users, 26.8% of tokens generated by ordinary engineers are on Codex, which has increased fivefold since the beginning of the year.

Data and data analytics practitioners followed closely behind, reaching 15.2%.

But in terms of seniority, the adoption of Codex spans the entire hierarchy, from junior employees to executives.

The report suggests that advanced users use Codex not only for writing code, but also for planning, reviewing, and delegating tasks.

Interestingly, within OpenAI, departments that later adopted Codex, such as legal and recruitment, had a faster transformation speed than the engineering departments that adopted it earlier.

In January 2026, the Codex usage rate of these departments was close to zero, reaching about 20% by early April and skyrocketing to 75% within a month.

This indicates that once an organization is determined, transformation can be very fast.

The person who is best at creating AI is also the one who uses AI the most ruthlessly
The core insight of this report is:

The adoption of agents does not depend on how strong the model is, but on how supportive the organizational environment is.

The same Codex presents completely different usage patterns within individuals, organizations, and OpenAI.

The report suggests that determining factors include access permissions to relevant documents and systems, management expectations, employee skill levels, and the presence of supporting review processes.

This is consistent with the diffusion law of all common technologies in history.

Technology itself is just a catalyst, and real change occurs in how organizations redesign their workflows around technology.

The report used a classic analogy:

During the transition from steam power to electricity in the late 19th century, early factories simply replaced steam engines with electric motors, and the layout and workflow of the factories remained largely unchanged.

The true explosion of productivity occurred decades later when factories were redesigned and production processes were completely restructured.

Agents may be going through the same process.

The current usage situation may only be the tip of the iceberg, and the real change is still ahead.