Codex joins ChatGPT to become a "workbot"

Last night's OpenAI live broadcast at 11:30 PM did not release a new model, but it highlighted Codex's more prominent position.

In this live session titled "Intelligence at Work," OpenAI focused on showcasing the latest updates to Codex aimed at enterprise and knowledge work. Among the most noteworthy points are the following four:

1. Codex will be more deeply integrated into ChatGPT. In the coming weeks, OpenAI will connect Codex capabilities to ChatGPT, allowing users to invoke it directly within a familiar interface to complete tasks.

2. Six new plugins tailored for specific job roles have been added to Codex, covering data analysis, creative production, sales, product design, public stock investing, and investment banking.

3. Introduction of Sites and Annotations features. The former enables Codex to produce interactive web pages, dashboards, or lightweight apps directly from work results, shareable via links with teams; the latter allows users to select and modify content within documents, charts, slides, and websites generated by Codex, similar to annotations, so AI can continue processing.

4. OpenAI also disclosed key data: Codex's weekly active users have exceeded 5 million, with about one-fifth coming from non-developer knowledge workers. Codex is expanding from a developer-centric tool into a broader enterprise role.

AI White-Collar Workbench

One of the most important signals from this broadcast is that Codex has been integrated into the ChatGPT ecosystem.

Previously, Codex had already been accessible in preview form on the ChatGPT mobile app, allowing users to view task progress, approve commands, and take over or adjust ongoing Codex tasks on their phones. During the broadcast, OpenAI stated that in the future, Codex features will be more broadly incorporated into the ChatGPT app, so users won't need to open Codex separately—they can directly let Codex handle more complex tasks within the familiar ChatGPT interface.

However, this does not mean Codex and ChatGPT will merge into a super app. The new change mainly reflects an expanded role for Codex, not a product merger.

Historically, Codex was primarily aimed at developers and engineering teams. Integrating Codex into ChatGPT can be understood as OpenAI leveraging the larger ChatGPT platform to push Codex into a wider range of work scenarios.

At the same time, OpenAI introduced six plugins designed for specific professional contexts, covering data analysis, creative production, sales, product design, public stock investing, and investment banking. These pack the applications, skills, and workflows relevant to various roles, making them closer to real office scenarios.

This approach is familiar; it’s reminiscent of Anthropic’s launch of ten Claude Agents focused on financial services just a month ago.

However, Claude’s set is more vertical, mainly centered on the finance industry, whereas OpenAI’s approach is more horizontal, expanding Codex from programming into multiple white-collar roles.

Among these, the sales plugin can connect with tools like Salesforce, HubSpot, Slack, Outreach, Clay, helping salespeople with customer research, lead follow-up, email generation, and meeting prep.

The data analysis plugin integrates with Snowflake, Databricks Genie, Hex, Tableau, and other tools, targeting internal enterprise data queries, report generation, KPI explanations, and analysis presentation.

Creative production and product design plugins connect with Figma, Canva, Shutterstock, Picsart, and more, enabling Codex to participate in design assets, visual creation, and product prototyping workflows—not just “idea generation.”

Financial plugins are even more direct. The public stock investing and investment banking plugins connect to Moody’s, FactSet, LSEG, S&P, PitchBook, Hebbia, and other data and financial tools, supporting research, valuation, market analysis, investment banking materials, and investment research.

Putting these two sets of information together, OpenAI is enabling Codex to carry tools, data, and role-specific workflows directly into ChatGPT.

Previously, ChatGPT handled conversations, and Codex executed tasks. Now, these lines are converging. OpenAI aims to let users connect questions, data, files, tools, and work results within a single interface. Sales teams can organize customer leads, analysts can run data and generate charts, design teams can produce assets and modify pages, and finance professionals can assist with research and modeling.

In other words, OpenAI is redefining the boundaries of Codex—from an AI coding tool to a broader AI white-collar workbench.

Delivering Usable Results

Besides role-specific plugins, OpenAI also introduced two new features aimed at work delivery: Sites and Annotations.

Simply put, Sites allow Codex to turn task results directly into accessible, interactive, and shareable web pages.

In the past, AI-assisted work often resulted in text, spreadsheets, code snippets, or analysis outputs. According to OpenAI, Codex can now generate a fully interactive website, dashboard, or lightweight app from work results, which can be shared via links with the team.

OpenAI provided several concrete examples.

For instance, users can ask Codex to create a site summarizing an upcoming client review. Codex will generate an interactive webpage containing relevant product updates, unresolved issues, usage trends, and next steps for that client account.

Another example: users can ask Codex to build a scenario planner based on financial models, allowing management to compare different assumptions directly—eliminating the need to flip through multiple tabs in a document.

According to OpenAI, Sites are not static pages. They can be used to track major project progress, guide customer service reps, or serve as a knowledge base for team creative briefs. OpenAI is also working with early partners like Vercel, Wix, Base44, Replit, Lovable, Figma, Webflow, and Emergent to build out the Sites ecosystem.

Currently, Sites is in preview for Business and Enterprise customers (since it’s an enterprise feature).

Annotations, literally “comments,” were previously used by developers within Codex to modify code, Markdown, and websites—simply by pointing to specific parts and telling Codex what to change.

Now, this approach will expand to more content types, such as documents, spreadsheets, and slides.

According to official demos, users can select a navigation bar within a site to have Codex update fonts; highlight a judgment in an investment opinion to ask where it comes from; or mark a chart in a slide to get clearer labels from Codex.

Looking at these two features together, the results generated by Codex are becoming easier to deliver, share, and modify.

Sites produce pages that teams can open, view, and interact with; Annotations enable users to continue providing feedback and making edits directly on the results.

For enterprise workflows, this is closer to real-world processes—many tasks require iterative refinement, producing a version first, then continuously modifying and confirming.

Codex Begins to Scale

Behind these new features, more crucially, are key data points. OpenAI disclosed that Codex now has over 5 million weekly active users. Since the desktop app launched in February, its user base has grown more than sixfold.

This means Codex is no longer just a new tool tested within developer circles. It is becoming one of OpenAI’s most prominent and promising entry points in the enterprise product lineup.

According to OpenAI, developers remain the largest user group for Codex. But knowledge workers now account for about 20% of weekly active users, with a growth rate three times that of developers.

If Codex were solely an AI programming tool, its scope would be limited to engineers, software development, and code repositories. But now, the user structure is changing—more and more knowledge workers outside of developers are using it for work. The six role-specific plugins are a direct response to this trend, productizing user-driven growth needs.

Another figure worth noting: Sam Altman mentioned during the broadcast that OpenAI’s largest token-consuming customer spends about 100 billion tokens per month, with some clients exceeding that. He also said that cost has become one of the most common questions from customers, second only to “how to simplify AI workflows.”

It’s clear that enterprises are taking AI more seriously and integrating it into real workflows. Only with high usage frequency and task complexity do token consumption reach hundreds of billions per month.

Moreover, once AI is embedded into workflows, costs and complexities also scale up. Longer tasks, more tool calls, frequent generation and editing all increase token consumption.

Token usage is already becoming a financial and management issue for companies—and it’s not isolated: Uber exhausted its 2026 AI coding tool budget in four months; Microsoft has begun canceling Claude Code licenses and shifting teams to GitHub Copilot CLI; Amazon recently shut down its internal AI token leaderboard KiroRank, which had once motivated employees to chase higher token counts, even deploying AI agents to perform low-value tasks to climb the ranks.

Today, Codex has become a core lever for OpenAI’s enterprise workflow entry point. Developers remain the main user base, but the fastest-growing segment is now broader knowledge workers.

In IPO narratives, this broadcast also signals a step in catching up with Anthropic. Anthropic has already put “growth, profitability expectations, and going public” on the table. OpenAI’s live session is also answering the market: Codex is growing and penetrating deeper into enterprise workflows.