When the product narrative expands from "writing code" to "getting work done," the competitive landscape shifts accordingly.Article author and source: 36Kr
Header and cover image source | codex
AI-generated summarySource: He Fan Cai Jing
On July 9, OpenAI released a set of interconnected updates on its official website.
The first item is GPT-5.6. This model family is divided into three tiers: Sol, Terra, and Luna. Sol is the flagship model, Terra offers a balance of capability and cost for everyday tasks, and Luna emphasizes speed and affordability.
The second is ChatGPT Work, an agent that can invoke apps and files, work continuously for hours, and turn goals into documents, spreadsheets, presentations, and websites.
The third item appears to be a client-side adjustment. The Codex desktop app, launched in February, has been integrated into the new ChatGPT desktop interface, making Chat, Work, and Codex three parallel entry points, while the original ChatGPT desktop app has been renamed ChatGPT Classic.
Screenshot of the new ChatGPT desktop app in Work mode
In other words, beyond model updates, OpenAI has extracted and extended Codex’s already validated capabilities into a Work entry point designed for general-purpose tasks.
On July 9, He Fan Caijing, in the article "Tencent, Alibaba, ByteDance: Rebuilding Office from Scratch," discussed trends in domestic "Work" products. For instance, WorkBuddy added project collaboration features, QoderWork launched a professional workspace, TRAE renamed SOLO to TRAE Work, and Tencent, Alibaba, and ByteDance have recently been restructuring their AI-powered office products.
The core argument in the text is that the programming agent was the first to successfully complete the closed loop of "understanding tasks, invoking tools, checking results, and delivering files," so it naturally expanded from writing code to handling documents, spreadsheets, research, and office tasks, eventually evolving into an AI-era Office.
This update from OpenAI is an immediate example of this judgment. The name Codex tells users that it writes code; the name Work tells users that it can take on tasks.
Interestingly, another aspect of this update occurred at Microsoft.
On July 9, OpenAI also announced that GPT-5.6 will become the preferred model for Microsoft 365 Copilot in Word, Excel, PowerPoint, Copilot Chat, and Cowork. Nitin Agrawal, President of Microsoft Copilot and Agents Core, and Nikunj Handa, Head of API Product at OpenAI, appeared together in the announcement, with the latter explicitly stating that GPT-5.6 will be integrated into Microsoft 365 via the OpenAI API.
In other words, on the same day, OpenAI both entered the engine room of Office and set up a new cockpit outside of Office.
On the surface, this is a model and client update. From a product narrative perspective, OpenAI has translated previously developer-only parameters and features into tasks—such as presentations and reports—that are more accessible to a broader audience.
Choose a new name and redefine the product story.
OpenAI must be very confident about this update.
In the accompanying product launch materials released alongside the update, the title directly reads, “ChatGPT for Your Most Ambitious Work.” “Ambitious” isn’t enough—it adds “most.”
Leaving aside specific parameters and technical details, from the perspective of the overall product logic, the new version of ChatGPT first changes the product division of labor.
On OpenAI's website, three entry points are clearly defined: Chat is for asking questions, searching, brainstorming, and quick conversations; Work is for researching, analyzing, and creating documents, spreadsheets, presentations, reports, or sites; Codex is for writing code, debugging, running tests and commands, reviewing and editing code, and managing code repositories.
Illustration: He Fan Cai Jing
Three entry points, corresponding to three different human-machine relationships.
Chat continues to use conversation as its basic unit, Work uses a deliverable task as its basic unit, and Codex maintains a professional environment composed of a code repository, terminal, and development tools.
In Work's official demonstration, this difference was deliberately incorporated into specific tasks.
Users can ask it to analyze month-end budget variances, organize raw materials into marketing briefs, or prepare for a sales meeting. Users can also monitor progress, answer questions, adjust direction, and approve critical actions. OpenAI also outlines a longer workflow: Work first converts customer research into campaign briefs, then generates marketing materials based on those briefs, adapts them for different markets, and retains context at every step.
The landing page also includes testimonials from several early users.
Nathan Bolt, Head of Digital Products at Virgin Atlantic, asked Work to study competing airlines and compare service differences based on a passenger journey checklist, then build a dataset for the team to review as part of the company’s five-year planning process.
Screenshot of the OpenAI website
OpenAI says this analysis has been reduced from weeks to hours.
NVIDIA’s market expansion manager, Will Daney, entrusted Work with handling customer registrations, scheduling meetings, and preparing for sales prior to the GTC conference, and then tasked it with consolidating hundreds of meeting transcripts and customer notes afterward.
Screenshot of the OpenAI website
According to Daney, the original Excel process accounted for about 40% of pre-meeting preparation time. Both results come from early test cases selected by OpenAI, not from third-party controlled experiments.
Focusing on continuous execution and final delivery, OpenAI has consolidated various capabilities previously scattered across Codex and ChatGPT around Work.
The plugin connects to Slack, Microsoft Teams, Google Drive, SharePoint, email, calendars, CRM, and project management systems. Scheduled Tasks can execute actions or monitor changes based on time or events. The built-in browser enables access to websites and online files. Computer Use on the desktop allows operation of local applications after user authorization. Sites lets users create dashboards, project trackers, internal portals, and interactive reports that can be shared via links.
Model names also evolve with the product narrative: the focus is no longer solely on parameter scale, but rather on aligning capabilities, speed, and cost with specific tasks.
On July 9, GPT-5.6 ended its limited preview and began full rollout across ChatGPT, Codex, and the API. Sol is the flagship model, Terra emphasizes the balance between capability and cost for everyday tasks, and Luna prioritizes cost efficiency; the top-tier Ultra mode defaults to coordinating multiple agents to work in parallel.
Compared to the past, where a single model name covered multiple tasks, Sol, Terra, and Luna are defined by OpenAI as distinct, long-term capability tiers that can each evolve independently; selection criteria now go beyond just "new vs. old" to include task difficulty, speed, and cost.
From the client perspective, Codex has not been replaced by Work. After July 9, developers can still set Codex as the default interface for desktop applications and even replace the app icon with the Codex logo. Meanwhile, existing Codex users will receive the new ChatGPT desktop version only after updating their application normally.
What has truly changed is that OpenAI has added access points for general-purpose tasks beyond Codex, and has reshaped user expectations using the names Chat, Work, and Codex.
This also explains why the old desktop version was named ChatGPT Classic. OpenAI’s most successful product format remains the chat interface, but in the new desktop version, chatting is now just one of three entry points. The answer is no longer the sole product—execution and delivery have been elevated to the same level.
People who misused "Codex" invented Work for OpenAI.
This update is primarily a reorganization based on AI product concepts and requirements.
Codex has not disappeared. The new desktop version of ChatGPT offers two optional modes: Codex, designed for developers and technical tasks, and Work, which helps you get your work done.
Screenshot of the new ChatGPT desktop app
On its July 9 release page, OpenAI explicitly stated that ChatGPT Work is powered by GPT-5.6; the primary distinctions between Work and Codex lie in their task scenarios, tooling environments, and delivery methods, not in being tied to specific model generations.
However, many of the execution methods emphasized by Work have previously appeared in Codex: working around files, invoking tools, continuous execution, checking results, and ultimately delivering the final output. This update is more like abstracting these capabilities from programming scenarios and repackaging them as a formal entry point for knowledge work.
There are many other supporting examples, such as customer cases released by OpenAI.
In May 2026, at the Madrid private bank Singular Bank, Juan José Guerrero was already doing something that appeared unrelated to programming.
In the past, he had to pull position data from multiple systems, manually verify the figures, and piece together a complete portfolio view before client meetings. After the bank built its internal assistant, Singularity, using ChatGPT and Codex, Guerrero can now analyze portfolios in real time during meetings, shifting his focus to client conversations. Singular Bank reports that pre-meeting preparation time has been reduced from about 20 minutes to under one minute, saving bankers 60 to 90 minutes per day; within a 30-day period, the team executed over 3,500 operations across 19 workflows.
Why would a tool originally designed for programmers appear in a private bank’s internal system? The answer lies at the origin of Codex.
On May 16, 2025, OpenAI publicly unveiled the cloud-based Codex research preview, with its scope clearly defined: a cloud-based software engineering agent capable of writing functions, answering questions about codebases, fixing bugs, and submitting pull requests.
And the code happened to provide the Agent with a rare training ground. On February 2, 2026, when the Codex desktop app launched on macOS, the product was still called the "Agent Command Center."
What does it mean?
In simple terms, users can have multiple agents work in parallel across different threads and workflows, review and modify their outputs, and continue asking follow-up questions. During this phase, OpenAI also introduced Skills and Automations. Skills bundle instructions, documentation, and scripts into reusable workflows, while Automations enable Codex to execute tasks in the background according to a schedule, such as categorizing issues, summarizing continuous integration failures, and generating release briefs.
Interestingly, at the end of the February post introducing this feature, OpenAI left a note: coding ability "lays the foundation for a broader range of knowledge tasks."
At the time, it disclosed that Codex had over one million developer users in the past month; four months later, the user base began to reshape the product definition.
On June 2, OpenAI released a set of more direct data showing that Codex had over 5 million weekly active users, more than six times the number at the time of the desktop app’s release in February; non-developers—including analysts, marketers, operations staff, designers, researchers, investors, and banking professionals—accounted for approximately 20% of users, growing at more than three times the rate of developers. On July 9, the Work landing page further stated that over 1 million people were using Codex outside of software development.
A product designed for programmers has been prematurely validated by users as a testing ground for general work. This phenomenon has been discussed in the article "Tencent, Alibaba, and ByteDance Redo Office."
For example, the same migration has also occurred at Anthropic.
Anthropic disclosed on the Claude Cowork product page that internal non-technical teams—such as marketing and data—have begun bypassing the standard chat interface to use Claude Code for handling complex, multi-step tasks; similar usage patterns have also emerged among external users. Cowork was thus created, retaining Claude Code’s ability to call files, tools, and applications, while offering a simpler interface tailored for non-technical users.
Domestic manufacturers have taken a more direct approach.
On June 9, 2026, ByteDance’s AI development tool TRAE announced the official upgrade of "TRAE SOLO" to "TRAE Work." TRAE stated that the product was initially designed for individual developers, but users soon began using it to write product requirements, analyze data, draft marketing plans, organize research reports, and coordinate cross-departmental projects. The official explanation for the name change was that the work users were accomplishing had surpassed the scope of the original name.
QoderWork, from the outset of its product design, aims to extend the capabilities of its Coding Agent to non-programmers. Qoder states that the product is built on its CLI Coding Agent, which can autonomously plan tasks, manipulate local files, and invoke skills.
On June 2, OpenAI took this expansion further by launching six job-specific plugins at once, covering data analysis, creative production, sales, product design, stock investing, and investment banking—integrating 62 applications and encompassing 110 skills. The announcement at the time specifically noted that these plugins "require no programming."
A month later, ChatGPT Work was released.
Competition and Collaboration: Supplying models to Office while bypassing Office
As the product narrative expands from "writing code" to "getting work done," the competitive landscape shifts. OpenAI is no longer just competing with coding tools, but with established productivity platforms like Microsoft 365.
On July 9, another announcement from OpenAI elevated this update from technology and product to business and competition.
On that 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 of Microsoft said this model will help users generate more refined documents, analyses, and presentations within their existing tools; Nikunj Handa of OpenAI emphasized that Microsoft 365 is where millions of people write, analyze, create, and collaborate every day.
This means that OpenAI has not abandoned reaching users through Office.
Microsoft owns Word, Excel, and PowerPoint, along with enterprise accounts, files, permissions, auditing, and distribution systems; GPT-5.6 accessing these applications via API means that OpenAI models may be involved in work regardless of whether the user opens ChatGPT.
Of course, Microsoft did not sit idly by.
On June 16, 2026, Microsoft Executive Vice President Charles Lamanna announced the global commercial availability of Copilot Cowork.
According to its official website, Cowork also emphasizes complex, time-consuming, multi-tool tasks: users define workflows, and Cowork executes them end-to-end to return completed results, not just suggestions or drafts. Microsoft reports that during its three-month preview period, more than half of the Fortune 500 companies used Cowork; one team used it to compare nearly 4,000 files between two product versions, and a sales leader condensed a week’s worth of stalled opportunity reviews into a single morning.
Microsoft also lists "multimodal" as one of Copilot's five differentiators.
On June 16, at commercial launch, Cowork ran Anthropic’s Opus 4.8 and Sonnet 4.6, while GPT-5.5 was available as an option in the Frontier project; Microsoft’s internally developed Cowork 1 was planned for release shortly thereafter. Twenty-three days later, Microsoft and OpenAI announced that GPT-5.6 would become the preferred model for Microsoft 365 Copilot.
But OpenAI’s Work, launched the same day, is not about adding a writing button to Word or an analysis sidebar to Excel. From a product logic standpoint: users must first describe their goal to Work, after which the Agent invokes plugins, local files, browsers, Microsoft 365, or Google Workspace to complete the task.
In other words, Office can still be involved, but it doesn't have to be the starting point.
Illustration: He Fan Cai Jing
This relationship of collaboration while maintaining independent space did not emerge only in July this year.
On July 22, 2019, Microsoft announced a $1 billion investment in OpenAI, with both parties planning to jointly build Azure AI supercomputing capabilities, making Microsoft OpenAI’s exclusive cloud provider and preferred commercialization partner.
By April 27, 2026, the two parties revised the agreement again: Microsoft will continue as OpenAI’s primary cloud partner, and the license for model and product intellectual property will be extended until 2032, but the license will now be non-exclusive. At the same time, OpenAI gains the ability to offer its products through other cloud providers. The key terms repeatedly used by both companies in their announcement have shifted from the earlier “exclusive” to include “flexibility.”
Of course, at least in the short term, both parties' interests remain aligned. A more powerful GPT-5.6 enhances the capabilities of Microsoft 365 Copilot, while Microsoft 365 distributes OpenAI's models to a large number of enterprise users.
As the boundaries of AI’s capabilities in real-world work become clearer, OpenAI and Microsoft are transitioning from a clear relationship of model provision and application distribution to an overlapping competition and collaboration among models, agents, and work entry points. If users continue to begin their work in Word, Excel, and PowerPoint, GPT-5.6 resides within Office; if users gradually become accustomed to first telling an agent what they want to accomplish, work happens outside of Office.
Now, OpenAI has made room for both outcomes.