ChatGPT Becomes Codex: Conversation Is No Longer the Product

This week, OpenAI is rebuilding ChatGPT around Codex—an agent that writes, runs, and edits code on its own, without a step-by-step conversational response to a question.

Formally, this is an update to a single product.

Essentially, this is a public admission that a dialogue with a model is worth almost nothing to a business unless it produces a working result.

Throughout the past year, AI success was measured by engagement: message and session counts and time spent in chat. Ben Thompson explains in Stratechery how OpenAI is shifting its flagship product from chat to Codex because chat metrics do not translate into revenue and retention as reliably as the metric “the agent completed the task.”

For an executive paying for subscriptions and integrations, this is a turning point: value comes from results that reach production.

A polished model response does not pay back the budget on its own.

This is where the TTU principle—time to use—comes in: a tool’s value is measured by the time from task to working result. A chatbot responds in seconds, but its answer still has to be manually turned into code, tests, and a deployment—and chat does not measure that time. Agent-based Codex shortens this exact stage: it writes the patch, runs tests, fixes errors in a loop, and opens a PR.

The model did not become smarter during that time. The “wrote the code → works in production” loop

became several human iterations shorter.

The task is assigned and the PR is ready—it looks simple.

A simple result does not mean an easy process: for an agent to safely write to someone else’s repository, run tests, and avoid breaking production, the model needs a disciplined engineering loop around it—review gates, an execution sandbox, and rollback when tests fail. The business does not see this work.

Without it, the agent remains a demo toy and never makes it to production.