For years, the AI industry has been telling you that the chatbot era was just the warm-up act. Turns out they weren't lying. New research published by OpenAI alongside researchers from Columbia, Duke, and the University of Pennsylvania shows that Codex, OpenAI's agentic coding and work platform, is growing at a pace that should make anyone who works at a keyboard sit down for a minute.
From Chatting to Doing: What the Hell Is Codex?
Here's the thing most people haven't wrapped their heads around yet. ChatGPT talks. Codex works. The distinction sounds small. It isn't.
ChatGPT will answer your question, draft your email, help you debug a line of code. Codex will actually go do the task. It's what the industry calls an 'agent,' a system that takes a goal, figures out the steps, executes them, and comes back with results. You delegate. It delivers. Or at least that's the sales pitch, and according to Axios, the numbers suggest people are buying it.
The researchers sorted Codex users into three buckets: OpenAI's own employees, outside organizations, and individual users. Then they measured how much Codex was being used compared to ChatGPT, counting by tokens, which is the standard unit for measuring how much an AI system is actually processing. The results showed Codex usage accelerating sharply.
The Numbers They're Dangling
The source article from Axios cuts off at a critical moment, right as it was about to drop the actual usage figures. Which is either a technical glitch or a reminder that the internet loves a cliffhanger. But the framing of the research alone tells you something.
When OpenAI commissions a study with three major universities and the headline finding is that Codex usage is 'accelerating,' that is a company signaling to the market, to investors, and to the press that the agentic era isn't a future promise anymore. It's a present-tense business story. OpenAI doesn't fund academic papers to shrug and say 'eh, growth is fine.'
The three-category breakdown of users, internal staff, corporate clients, and individual consumers, is also interesting. That's the structure of a company mapping a technology's spread from early adopters to mass market. They're watching it climb the ladder in real time.
Why This Matters Beyond the Tech Nerd Beat
Axios frames this correctly: AI is moving from chat and web search to delegated work. That sentence is doing a lot of lifting. 'Delegated work' means tasks that previously required a human being to sit down, think, and execute over hours or days.
We've been here before with automation anxiety, and usually the jobs-are-fine crowd wins the short-term argument. But agentic AI is qualitatively different from a loom or a spreadsheet. It doesn't just speed up one step of a process. It can handle entire workflows, from receiving an instruction to producing a finished output, without a human touching it in between. That's not a faster tool. That's a replacement pipeline.
The workers most immediately in the crosshairs are knowledge workers, the coders, analysts, researchers, writers, and junior associates who spend their days breaking down complex problems and producing structured outputs. Which is, coincidentally, precisely what Codex is designed to do.
The Labs Have Been Promising This for Years. So Why Now?
The AI industry has a credibility problem when it comes to agents specifically. Going back years, every major lab has rolled out demos of AI agents that could theoretically do your taxes, book your flights, and manage your calendar. Most of them fell apart the moment they hit real-world complexity. The systems would loop, hallucinate, fail on steps three and four, and leave users cleaning up a bigger mess than they started with.
What's changed, according to the labs at least, is that the underlying models are now capable enough to handle multi-step reasoning reliably enough to be useful. OpenAI has been pushing this case hard. The research collaboration with Columbia, Duke, and UPenn is part of that push, lending academic credibility to what would otherwise be a marketing claim.
Whether Codex has actually crossed the reliability threshold or whether OpenAI is once again getting ahead of reality is the question the truncated Axios data can't fully answer yet. But the fact that the usage numbers are growing fast enough to publish a research paper about them suggests something real is happening, even if we don't know the exact scale.
Who's Actually Using This Thing
The user breakdown in the research is the most telling detail in the whole story. OpenAI's own employees are one category. That's internal dogfooding, standard practice, not especially surprising. But outside organizations as a distinct and apparently substantial category means enterprises are already deploying this at scale, not just kicking the tires.
Individual users rounding out the three categories suggests Codex has also broken out of the pure enterprise software box and started reaching people directly. That's the pattern for every technology that eventually becomes ambient and unavoidable. First the company uses it internally, then the companies that pay for it deploy it, then regular people start using it on their own. We're apparently at step three.
How fast that individual adoption is growing relative to the corporate side would be genuinely important information. Unfortunately, that's precisely where the article goes dark.
The Dingo Take
Look, the AI hype cycle has burned so many people so many times that healthy skepticism is basically mandatory at this point. We've been promised robot assistants since at least 2016 and mostly gotten autocomplete that occasionally makes up court cases. So the instinct to roll your eyes at 'Codex usage is accelerating' is completely understandable.
But here's what's different this time. OpenAI is not announcing a demo. They're publishing usage data, measured in tokens, broken down by user category, co-authored by researchers from three universities. That's a company trying to show receipts, not hype. And the receipts are apparently interesting enough to put in a paper. The question of whether AI agents are actually reliable enough to do consequential work without human supervision is still genuinely open. But the question of whether people are starting to try it at scale seems to be answered.
The most honest thing we can say is this: if Codex and systems like it actually work even half as well as OpenAI claims, the disruption to knowledge work will be substantial, fast, and deeply unequal. The people who own these tools will benefit. The people who were doing the work the tools now do will not. Nobody in Washington is seriously grappling with that. Nobody in the boardrooms funding this technology is losing sleep over it. And the academic papers measuring the acceleration certainly aren't written to slow anything down.
