OpenAI Runs on Codex Now. The Whole Company, Not Just Engineers.

The headline most people read in June was the model number. OpenAI previewed GPT-5.6, a new three-tier line built for agentic, tool-heavy work. That is the smaller story. The bigger one is what happened to the tool that runs on it.

According to Andrew Ambrosino, who leads the Codex desktop app, nearly 100% of OpenAI employees now use Codex every week. Not the engineers. All of them. Marketing, comms, finance, legal. The company crossed 5 million weekly active users on Codex, growing 6x since January. A product named after code became the way an entire frontier lab does its work.

For a founder watching this, the interesting question is not "how good is the model." It is "how did a coding tool become the default interface for a whole company, and could that happen to mine."

What Actually Shipped: GPT-5.6

The model line underneath Codex moved. GPT-5.6 arrived as a three-tier family, each tuned for a different job:

• Sol, the frontier tier. The most capable, for the hardest agentic and reasoning work.
• Terra, the balanced everyday tier. The one most teams will run by default.
• Luna, the fast and affordable tier, built for high-volume, tool-heavy tasks.

Greg Brockman, OpenAI's president, called it simply "a good model." But the rollout came with a catch worth understanding. Per Dan Shipper of Every, a US government directive limited access to the top Sol tier to roughly 20 pre-approved companies, on temporary national-security grounds. It mirrors the June export restrictions on Anthropic's most capable models. The frontier tier is gated. The everyday tiers, Terra and Luna, are not, and they are what most companies will actually build on.

How a CLI Became a Company OS

Codex started as a terminal tool for developers. It should have stayed niche. Engineers live in the command line, finance and legal do not. So why did everyone adopt it?

Because the surface changed. Codex grew from a CLI into a full desktop app with computer use, an in-app browser, and the ability to drive Chrome through an extension. Once an agent can open a browser, read a document, fill a form, and take actions in real apps, the job stops being "writing code" and becomes "doing the work." Three capabilities did most of the lifting:

That last point is why non-engineers came aboard. When Peter Yang, a product leader, switched from Claude Code to Codex, the reason was not raw model quality. It was reach.

The Record & Replay release landed the same way with practitioners. Dan Shipper, who runs Every and uses Codex daily, reacted in three words.

The Story Underneath the Numbers

Strip the product names away and the pattern is the one that matters for your company. A frontier lab did not roll out agents department by department through a transformation program. It shipped one tool good enough that every function adopted it on its own, then watched the work reorganize around it.

Ambrosino's framing for the shift is worth keeping: when implementation gets cheap, the scarce skill becomes taste, knowing what done looks like and curating toward it. His stated goal for Codex is to build the best desktop app that has ever existed, a home base that orchestrates every other tool. OpenAI is dogfooding that bet on itself, in public, as the clearest live example of an agent-native company.

What This Means for Your Company

OpenAI is a special case. It builds the models, so of course it runs on them. But the mechanism that made adoption stick is not special, and it is the part worth copying.

Codex spread because two conditions were met. The tool could take real actions in real systems, and the context it needed, the goals, the institutional knowledge, the linked data, was within reach. Where those two conditions hold, agents move from engineering into every function. Where they do not, you get a pilot in one team and a stalled rollout everywhere else.

That is the work. Not picking the model, the tiers are converging and you will run whatever is good and available. The work is making each function legible enough that an agent can act inside it: clean context, defined goals, connected data, a clear definition of done. Get that right for one function and the same playbook repeats across the next. That is exactly how the rollout runs, function by function, and it is the work nativefirst does on site.