Codex Hit 4M Weekly Users — 5× Growth in 3 Months

OpenAI's Codex — the AI agent that writes, fixes, and reviews code on its own — has reached 4 million weekly active users, a fivefold jump in the last three months alone, with monthly growth running at 70%. Arnaud Fournier, CTO of DeployCo (OpenAI's enterprise deployment arm), shared these figures in a new interview and added a striking claim about costs: "the price of intelligence has dropped a hundredfold over the past 18 months."

That cost collapse is real in aggregate — AI has become dramatically cheaper to use since 2024. But Fournier acknowledged a counterintuitive wrinkle: GPT-5.5, OpenAI's latest premium model, actually costs 49 to 92 percent more than its predecessor. The reason is that this model now takes more time to think before answering — it works through a problem more carefully rather than responding instantly. So AI is getting cheaper for most everyday tasks, and more expensive specifically for the cases where you want the deepest, most careful answer. Both things are true at the same time.

Where Codex is actually being used

Fournier provided three enterprise examples that show what AI adoption at scale actually looks like today. BBVA Bank shifted from annual credit-document automation to continuous daily and weekly risk assessments — same task, but now running constantly instead of once a year. John Deere uses AI to reduce the amount of chemical fertilizer applied in fields, with measurable cost savings and environmental benefits. Stadler, a German waste-sorting company with about 650 employees, reports that 85% of its staff use ChatGPT as part of their daily work. These aren't AI experiments or small-scale pilots — these are operational dependencies.

Germany stands out in the growth data: Codex usage grew 720% there since January 2026, placing it first in Europe and in the global top five markets. Fournier attributes part of this to Germany's engineering culture and appetite for tools with measurable, quantifiable outcomes. It's not the country you'd typically think of as an AI-adoption leader, but the numbers don't leave much room for argument.

The ROI question nobody has fully answered yet

When asked what return on investment companies are actually seeing from AI, Fournier was refreshingly honest about the limits of the data. "It's early days," he said — most deployments only started 6 to 12 months ago, which is too short a window to measure comprehensive ROI with confidence. His practical advice: "get a Codex license for $20 and start." Don't wait for someone else's case study to tell you whether it's worth it. Run your own experiment on your own work.

That's genuinely useful guidance, but it also reveals where we are in the adoption curve. Four million weekly users and 70% monthly growth is not a niche experiment — that's mainstream adoption by any standard. But "we're still figuring out ROI" is not the same as "the ROI is obvious." The companies with the clearest returns are the ones who committed early and built real operational dependencies — the BBVAs and John Deeres — not the ones running occasional tests on low-stakes tasks.

Fournier also dismissed concerns about EU regulation as "barely a brake anymore" in practice — a notable statement from someone deploying AI across European enterprises. Whatever friction the EU AI Act was expected to create in enterprise deployment, it apparently isn't showing up in Codex's 720% German growth figures.