ChatGPT and Codex Merge: A Billion Users Gain a Super Agent

At the "Intelligence at Work" online launch, OpenAI product leadership officially announced that Codex will be embedded into ChatGPT in the coming weeks, creating a unified AI experience for nearly one billion weekly active users.

The rollout introduces three core upgrades:

Agent plugins : customizable plugins that tailor Codex to individual roles and preferred tools.

Annotations : inline annotations that let users directly edit Codex output, instantly refining code or text.

Sites : a single‑URL workflow that generates interactive web apps and shares them across teams.

According to the "The Next Era of Knowledge Work" report, Codex weekly active users have jumped six‑fold since February, surpassing 5 million, with knowledge‑worker usage now accounting for 20 % of the base—growing three times faster than developer usage.

Six role‑specific AI colleagues are being launched, covering data analysis, creative production, sales, product design, public‑market investment, and investment banking. They integrate 62 mainstream enterprise applications and expose more than 110 skills. For example, when product lead Alexander asked the data‑analysis agent to investigate a sudden rise in subscription cancellations, the agent interpreted the semantic layer, generated the necessary SQL query, produced a full report, and then transformed that report into a presentation deck—all within minutes.

In a creative‑production scenario, Alexander started from a product image, prompted the agent to build a mood board, refined lighting and style with remix, generated advertising graphics via imagegen, and exported the result to Canva as reusable assets, collapsing a multi‑day workflow into seconds.

The new Annotations feature solves the "final tweak" problem: users simply select the portion they want to change, describe the modification, and the agent edits only that segment, leaving the rest untouched. An example shows a table region being turned into a chart in seconds, or a navigation bar’s font being updated without affecting surrounding content.

Sites extends this capability to full‑stack output. A single prompt can turn any Codex result—financial forecasts, product plans, or release notes—into an interactive web app with a shareable URL. The generated sites are not static; the embedded agent retains full business context and can continuously maintain the app.

Underlying the Codex engine is GPT‑5.5, which OpenAI claims delivers comparable output quality while consuming only one‑third of the tokens used by previous models. A slide presented at the event showed a curve where intelligence (measured by model performance) rises as token usage drops, illustrating the cost efficiency of GPT‑5.5.

In the broader AI landscape, Anthropic’s Claude Code and its Cowork product have already targeted knowledge workers. Codex’s three upgrades map closely to Cowork’s selling points, but OpenAI’s advantage lies in distribution: Codex’s capabilities are baked into ChatGPT, making them instantly available to all users, whereas Cowork requires a separate desktop download.

OpenAI’s strategy positions AI agents as the default capability of ChatGPT, lowering the barrier to AI‑augmented work and aiming to shift the economic value of AI from a small set of companies to the broader workforce.