The creators of ChatGPT are hardly using ChatGPT for work anymore?
In less than a year, OpenAI has shifted its primary AI from a chat interface to AI agents.
By June 2026, Codex was responsible for 99.8% of OpenAI's weekly output tokens.
Just ten months prior, that number was less than 10%.
The turning point occurred around September last year. Codex was connected to more powerful models and gained additional capabilities, taking on increasingly complex tasks.
Employees gradually realized that instead of back-and-forth Q&A in a dialog box, it was more effective to hand over entire swathes of tasks for it to run autonomously.
And this wasn't just a trial by a single engineering team. Across the entire company, from legal and finance to recruiting, every department has positioned it as their primary AI tool.
Today, over 85% of the average OpenAI employee's output tokens are generated by Codex. Heavier users naturally consume more tokens, pushing the company-wide weighted average to 99.8%.
Thus, within its own birthplace, a chatbot has been replaced by its sibling.
https://openai.com/index/how-agents-are-transforming-work/
OpenAI made this clear in their latest blog post:
Agents are rewriting the fundamental unit of knowledge work—from single-turn Q&A exchanges to entire long-cycle tasks that can be 'thrown over the wall.'
While a chatbot handles one short query at a time, an agent can operate independently for minutes to hours, calling tools, interacting with its environment, and iterating until the job is done.
Now, nearly a quarter of Codex requests correspond to tasks that would take a human over an hour to complete.
OpenAI President Greg Brockman shared this report, stating: Agents are being adopted rapidly, accelerating everyone's work.
The chart he included was precisely this steep internal adoption curve.
The Fire Spreads from Engineering Desks to the Legal Office
Engineers were the first to change.
This is unsurprising, as Codex was originally built for coders.
Since December 2025, the average OpenAI engineer has moved the majority of their work to Codex. Today, 99% of an average engineer's output tokens come from Codex, leaving only a sliver for ChatGPT.
But the fire didn't stop at the engineering department.
Around April 2026, departments like legal, finance, and recruiting—which don't touch code at all—collectively crossed over, adopting Codex as their top tool, and switching even faster than engineers.
Today, over 85% of the average lawyer's or recruiter's output tokens at OpenAI also come from Codex.
Usage across departments has climbed rapidly like a spreading fire.
According to OpenAI's own data, by June 2026, median usage in the research department increased 56-fold, customer support 32-fold, engineering 27-fold, and even the slowest-moving legal department grew 13-fold.
When lawyers hand over work to an agent, that image alone is more convincing than any benchmark.
The True Signal Comes from Those Who Didn't Code Before
If you only see engineers favoring Codex, you might miss the most critical signal.
Since August 2025, growth among non-developer users has comprehensively surpassed that of developers: 137-fold for individual users, 189-fold for organizational users, and 12-fold internally at OpenAI.
A tool born from coding is now being used by more and more people who don't understand code at all.
What do they use Codex for?
The finance team used it to process 24,771 K-1 tax forms, totaling 71,637 pages. This anonymized workflow allowed the team to finish two weeks earlier than the previous year.
The public relations team went even further, building an automatic Slack agent for routing: low-risk speaking invitations are handled automatically, while high-risk ones are flagged for human review.
The tasks entrusted to it are also growing heavier.
By May 2026, 80.6% of individual users had requested tasks estimated to take over 30 minutes, 70.2% over 1 hour, and 25.6% directly assigned tasks exceeding 8 hours.
More subtly, over a quarter of the work business roles do with Codex is actually programming. A finance professional is quietly stepping into the engineer's domain.
The walls between job functions are slowly dissolving.
Breaking down the work done with Codex by OpenAI departments by type: 31% of finance tasks are programming, 25% of product marketing tasks, and even 50% of tasks in the non-technical 'Other' department involve writing code. Codex is gradually erasing the walls between roles.
At this stage, Codex is no longer just a programming agent—it has crossed the boundary to become a general workflow agent.
This is what truly sends a chill down the spine.
From Tool to Executor: Codex Changes Its Identity
Supporting all this is a complete transformation in Codex's role.
It is no longer the code-completion plugin. Today, it can take on an entire engineering task chain: implementation, refactoring, debugging, testing, validation—handling it all end-to-end.
Even early versions could autonomously run for over 7 hours per session, iterating on implementations, fixing test failures, and finally delivering a working solution.
This goes beyond helping you write a few lines of code; it's you assigning a complex task, and it running from start to finish on its own.
Even more telling is the scale of parallelism.
By June 2026, the heaviest users at the P99 percentile could have Codex generate over 60 hours worth of agent runs in a single day, distributed across several parallel agents.
Users are no longer satisfied with asking for an answer; they are commanding a squad of agents simultaneously throughout the day.
Daily Codex agent runtime within OpenAI, divided into five tiers from average users to the heaviest users. By June 2026, the heaviest users could generate over 60 hours of agent workload in a single day.
One person, one day, coordinating over 60 hours of work—that's a week's worth for someone else.
Codex is built on GPT-5.5. It can handle longer tasks with fewer tokens.
But what's most surprising is another thing GPT-5.5 has done.
To speed up without compromising latency, OpenAI had it rewrite the heuristic algorithms for load balancing and partitioning.
GPT-5.5 analyzed weeks of real traffic, wrote a custom solution, and boosted token generation speed by over 20%.
Thus, GPT-5.5 became an engine that started optimizing itself.
An NVIDIA engineer who got early access even said losing access to GPT-5.5 felt like an amputation.
Behind all this, the entity humans partner with has quietly changed: from a Q&A chatbot to an agent that can run long tasks independently.
What remains unchanged are the people who issue commands, make judgments, and bear responsibility. What has changed is the default office action: from opening a chatbox to ask a question, to handing over an entire matter for an agent to run.
This report reads more like a preview of a shift in work style.
Going forward, what will truly create a gap is how large a swathe of work you dare to hand over entirely to AI.
References:
https://openai.com/index/introducing-codex/https://openai.com/index/codex-for-every-role-tool-workflow/
https://x.com/gdb/status/2070199649823297653
https://openai.com/index/how-agents-are-transforming-work/https://openai.com/index/harness-engineering/
This article is from the WeChat public account "New Zhiyuan," author: ASI启示录; edited by: Yuan Yu
