OpenAI has found a new job for Codex

When product narratives expand from “writing code” to “getting the job done,” competitive boundaries shift accordingly.

On July 9, OpenAI rolled out a set of interlocking updates on its official website.

The first is GPT-5.6. This model family is split into three tiers: Sol, Terra, and Luna. Sol is the flagship model; Terra balances capability and cost for everyday work; Luna emphasizes speed and price.

The second is ChatGPT Work. An Agent that can call applications and files, work continuously for hours, turn goals into documents, spreadsheets, slides, and websites.

The third appears to be just a client-side adjustment. The Codex desktop app launched in February was folded into the new ChatGPT desktop client. Chat, Work, and Codex become three parallel entry points, and the original ChatGPT desktop app is renamed ChatGPT Classic.

That is to say, beyond model updates, OpenAI has extracted the already-validated capabilities of Codex and extended them into a Work entry point built for general work.

On July 9, in an article titled “Tencent, Alibaba, Byte, redo Office once” (《腾讯、阿里、字节，重做一遍Office》), Hexiafan Finance discussed trends in domestic “Work” products related to this. For example, WorkBuddy added project collaboration; QoderWork launched a professional workspace; TRAE rebranded SOLO to TRAE Work; and Tencent, Alibaba, and Byte have all recently been rebuilding AI office products.

The article’s core point is: coding agents are the first to successfully run the closed loop of “understanding tasks, calling tools, checking results, and delivering files,” so they naturally expand from writing code to documentation, spreadsheets, research, and office work—and ultimately evolve into an Office for the AI era.

And this OpenAI update is exactly a timely case supporting that judgment. The name Codex tells users it will write code; Work tells users it can take over the job.

What’s interesting is that the other side of this update is happening at Microsoft.

Also on July 9, OpenAI announced that GPT-5.6 will become the preferred model for Microsoft 365 Copilot across Word, Excel, PowerPoint, Copilot Chat, and Cowork. Microsoft Copilot—along with Nitin Agrawal, President of Agents Core, and OpenAI API product lead Nikunj Handa—appeared in the announcement at the same time. Handa explicitly said GPT-5.6 will enter Microsoft 365 through the OpenAI API.

That means on the same day, OpenAI both climbed into the engine bay of Office and set up a new cockpit outside the Office doors.

On the surface, this is a model and client update. From a product-narrative perspective, OpenAI has translated parameters and features that used to be understandable only to programmers into work tasks that more people can understand—such as PPTs and reports.

Change to a new name, change the product narrative

OpenAI should feel very confident about this update.

In the product introductions released alongside the update, the title is directly “ChatGPT for your most ambitious work.” “Ambitiou”s-style ambition clearly isn’t enough—so it also adds “most.”

Stepping out of specific parameters and technology, looking at the overall logic of the product, what the new ChatGPT changes first is product division of labor.

On OpenAI’s official site, the three entry points are defined very specifically: Chat handles asking questions, searching, brainstorming, and rapid back-and-forth conversations; Work handles research, analysis, and producing documents, spreadsheets, presentations, reports, or Sites; Codex is responsible for writing code, debugging, running tests and commands, reviewing and editing, and handling code repositories.

The three entry points correspond to three different human-machine relationships.

Chat still uses conversational interaction as its basic unit; Work uses a deliverable task as its basic unit; Codex preserves a professional environment made up of code repositories, terminals, and development tools.

And in the official Work demo, OpenAI even deliberately put this difference into specific work.

Users can ask it to analyze month-end budget variances, turn raw materials into marketing briefs, or prepare for a sales meeting. Users can also observe progress, answer questions, change direction, and approve important actions. OpenAI also lists a longer chain: Work first turns customer research into an event brief, then generates marketing assets from that, adapts them to different markets, and preserves context at every step.

The launch page also includes pitches from several early users.

Virgin Atlantic’s digital product lead Nathan Bolt, when building the company’s five-year plan, had Work analyze competitor airlines using a passenger journey checklist, compare service differences, and then build a dataset for the team to review.

OpenAI says this analysis went from taking weeks to taking hours.

NVIDIA business development manager Will Daney assigned Work the customer registration before the GTC conference, meeting scheduling, and sales preparation. After the event, he had it compile hundreds of meeting notes and customer notes.

According to Daney, the original Excel workflow consumed about 40% of pre-meeting preparation time. Both results come from early test cases selected by OpenAI, not third-party comparative experiments.

Around continuous execution and final delivery, OpenAI has consolidated several capabilities that used to be scattered across Codex and ChatGPT and placed them around Work.

Plugins connect to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRMs, and project management systems. Scheduled Tasks can execute tasks based on time or events and monitor changes. A built-in browser handles accessing websites and online files. Desktop Computer Use can operate local applications after user authorization. Sites lets users turn results into dashboards, project trackers, internal entry points, and interactive reports that can be shared via links.

Even model naming changes along with the product narrative: the focus is no longer just on parameter scale, but on mapping capability, speed, and cost to different tasks.

On July 9, GPT-5.6 ended a limited preview and began full rollout across ChatGPT, Codex, and the API. Sol is the flagship model; Terra emphasizes the balance of capability and cost for everyday work; Luna emphasizes cost efficiency; and the top-tier ultra mode will default to coordinating multiple Agents working in parallel.

Compared with the past approach of using a single model name to cover multiple tasks, OpenAI defines Sol, Terra, and Luna as long-term capability tiers that can each be iterated. The basis for selection is no longer only “new vs. old,” but also task difficulty, speed, and cost.

From the client setup, Codex is not replaced by Work. After July 9, developers can still set Codex as the default interface for the desktop app, even changing the app icon to the Codex mark. At the same time, only after existing Codex users update their app normally do they get the new ChatGPT desktop client.

The real change is that OpenAI added a general-work entry point beyond Codex and reshaped user expectations by redividing them with the names Chat, Work, and Codex.

That also explains why the old desktop client had to be named ChatGPT Classic. OpenAI’s most successful product form is still the chat box, but in the new desktop client, chat is now only one of the three entry points. The answer is no longer the only product; execution and delivery are put on the same level.

People who used “Codex” the wrong way helped OpenAI invent Work

This update is more like a reorganization driven by the product thinking and needs of AI products.

Codex has not disappeared. The new ChatGPT desktop client provides two optional modes: Work and Codex. Codex is for developers and technical work; Work is responsible for “helping you get the job done.”

In the July 9 launch page, OpenAI clearly states that ChatGPT Work is powered by GPT-5.6. The main differences between Work and Codex are task scenarios, tool environments, and delivery methods—not that each is tied to a specific generation of model.

However, many of the execution styles Work emphasizes already existed in Codex: working around files, calling tools, continuous execution, checking results, and ultimately delivering a finished product. This update is more like abstracting these capabilities out of programming scenarios and repackaging them as a formal entry point for knowledge work.

There are plenty of supporting examples, including customer case studies published by OpenAI.

In May 2026, Madrid private bank Singular Bank’s Juan José Guerrero was already doing something that looks unrelated to programming.

In the past, before client meetings he had to pull positions from multiple systems, manually verify the data, and then piece together a complete investment portfolio view. After building an internal assistant, Singularity, using ChatGPT and Codex, Guerrero could analyze the portfolio in real time during meetings, shifting time back to talking with clients. Singular Bank says that single-session pre-meeting preparation dropped from about 20 minutes to under 1 minute; bankers save 60 to 90 minutes every day. Over a 30-day cycle, the team completed more than 3,500 operations covering 19 workflows.

Why would a tool originally aimed at programmers appear inside a private bank’s internal systems? The answer goes back to Codex’s starting point.

On May 16, 2025, when OpenAI公開 the cloud-based Codex research preview, its boundaries were very clear: it was a cloud software engineering Agent that could write functions, answer questions about code repositories, fix bugs, and submit Pull Requests.

And code just happened to provide the Agent with a rare training arena. On February 2, 2026, when the Codex desktop app launched on macOS, the product was still referred to as an “Agent command center.”

What does that mean?

In simple terms, users can have multiple Agents work in parallel across different threads and work trees, inspect changes, continue asking follow-up questions. At this stage, OpenAI also added Skills and Automations. The former packages explanations, materials, and scripts into reusable workflows; the latter lets Codex categorize questions in the background on a schedule, summarize continuous integration failures, generate release briefs, and run other tasks.

What’s interesting is that at the end of the February launch article, OpenAI left a line: code capabilities “lay the foundation for a broader range of knowledge work tasks.”

At the time, it disclosed that Codex had more than 1 million developer users over the past month. Four months later, the user composition started changing and began reshaping the product definition.

On June 2, OpenAI published a set of more direct data. The numbers showed: Codex had over 5 million weekly active users, more than six times the desktop app launch in February. Non-developers—such as analysts, marketers, operators, designers, researchers, investors, and banking professionals—accounted for about 20% of users, and their growth rate was more than three times that of developers. On July 9, the Work launch page further claimed that more than 1 million people were already using Codex outside software development.

A product aimed at programmers was thus validated early by users as a testbed for general work. In the article “Tencent, Alibaba, Byte, redo Office once,” this phenomenon is already explained.

For example, a similar shift also happened at Anthropic.

Anthropic disclosed on the Claude Cowork product page that the company’s internal non-technical teams—like marketing and data—began bypassing the standard chat interface and instead used Claude Code to handle complex, multi-step tasks; external users showed similar usage. Cowork was created as a result: it preserves the execution capabilities of Claude Code for calling files, tools, and applications, but provides non-technical users with a simpler interface for interaction.

Domestic vendors moved even more directly.

On June 9, 2026, ByteDance’s AI development tool TRAE announced that “TRAE SOLO” would be officially upgraded to “TRAE Work.” TRAE said the product initially served independent developers, but users later started using it to write product requirements, analyze data, draft marketing plans, organize research reports, and coordinate cross-department projects. The official explanation for the rename is: the work users complete has already gone beyond the original name.

QoderWork, meanwhile, from the very beginning of its product design, tried to migrate the Coding Agent’s capabilities to non-programmers. Qoder says the product is built on its CLI Coding Agent, allowing it to plan tasks independently, operate local files, and call Skills.

On June 2, OpenAI pushed this kind of expansion one step further: it released six job-role plugins at once, covering data analysis, creative production, sales, product design, stock investing, and investment banking—connecting to 62 applications and including 110 Skills. Even in that announcement, it specifically said these plugins “do not require programming.”

A month later, ChatGPT Work appeared.

Co-opetition: provide templates for Office, while bypassing Office

When product narratives expand from “writing code” to “getting the job done,” competitive boundaries also change. OpenAI is no longer just facing programming tools, but existing office entry points like Microsoft 365.

On July 9, another OpenAI announcement turned this update from technology and product into business and competition.

On that same day, OpenAI announced that GPT-5.6 will become the preferred model for Microsoft 365 Copilot in Word, Excel, PowerPoint, Copilot Chat, and Cowork.

Nitin Agrawal at Microsoft said this model will help users produce more complete documents, analysis, and presentation materials within their existing tools. OpenAI’s Nikunj Handa emphasized that Microsoft 365 is where hundreds of millions of people write, analyze, create, and collaborate every day.

This means OpenAI did not give up on reaching users from inside Office.

Microsoft owns Word, Excel, and PowerPoint, and it also has enterprise accounts, files, permissions, audit and distribution systems. GPT-5.6 entering these applications via the API means that whether or not users open ChatGPT, OpenAI models may still participate in the work.

Of course, Microsoft is also not waiting passively.

On June 16, 2026, Microsoft executive vice president Charles Lamanna announced that Copilot Cowork is now officially available globally for commercial use.

Based on the official site introduction, Cowork also emphasizes complex, long-duration, multi-tool tasks: users define the work; it executes end-to-end and returns completed results—not advice or drafts. Microsoft said that during the three-month preview, more than half of the Fortune 500 used Cowork. One team compared nearly 4,000 files between two versions of the products; another sales leader compressed a week of stalled deal reviews into a single morning.

Microsoft also listed “multi-model” as one of Cowork’s five differences.

When Cowork went officially commercial on June 16, it ran Anthropic’s Opus 4.8 and Sonnet 4.6. GPT-5.5 could be selected in the Frontier project, and Microsoft’s own-trained Cowork 1 was planned for release afterwards. 23 days later, Microsoft and OpenAI announced GPT-5.6 will become the preferred model for Microsoft 365 Copilot.

But the Work OpenAI launched the same day isn’t just adding a writing button in Word, nor an analytics sidebar in Excel. From the product logic: the user first describes the goal to Work, and then the Agent calls plugins, local files, a browser, Microsoft 365, or Google Workspace to complete the task.

That is, Office can still be involved, but it doesn’t have to be the starting point.

This kind of relationship—cooperating while preserving independent space—did not start only this July.

On July 22, 2019, Microsoft announced a $1 billion investment in OpenAI. The two planned to jointly build Azure AI supercomputing capabilities, with Microsoft becoming OpenAI’s exclusive cloud services provider and a preferred commercialization partner.

By April 27, 2026, the two adjusted the agreement again: Microsoft would remain OpenAI’s main cloud partner, and the licensing of model and product intellectual property would be extended to 2032, but the license would change from exclusive to non-exclusive. At the same time, OpenAI obtained room to provide its products to customers through other cloud service providers. The keywords repeatedly used in the two companies’ announcements had shifted from “exclusive” in the early days to adding “flexibility.”

Of course, at least in the short term, their interests still align. A stronger GPT-5.6 improves Microsoft 365 Copilot’s capabilities, while Microsoft 365 distributes OpenAI models to a large number of enterprise users.

As the capability boundaries of AI in real work become clearer, OpenAI and Microsoft are moving from a clearly defined model supply and application distribution relationship into a competitive landscape where models, Agents, and work entry points overlap. If users keep starting work from Word, Excel, and PowerPoint, GPT-5.6 is inside Office; if users gradually get used to first telling the Agent what they want to accomplish, Work sits outside Office.

Now OpenAI has left room for both outcomes.