OpenAI has spent years turning large language models into assistants that can summarize, write and generate code. Now it wants those models to take action across the tools that dominate office work. ChatGPT Work, the newest product in the company’s desktop and mobile app, connects AI agents to email, Slack, cloud documents and enterprise software, so a single prompt can produce a report, update a dashboard, schedule meetings or sort through an inbox. The product is a bet that the next phase of artificial intelligence will be measured less by what models say and more by what they do on behalf of their users.

Key facts:

• ChatGPT Work is available to paying subscribers on OpenAI’s $20-per-month tier, making agentic AI a mainstream consumer feature.
• The product is essentially a non-engineer version of Codex, OpenAI’s popular coding agent, capable of using internal and external tools over long, multi-step projects.
• OpenAI’s own research shows that while 98 percent of its employees used Codex in June, only 17 percent of organizational subscribers and less than 1 percent of individual subscribers were regular users.
• OpenAI reports over one billion total ChatGPT users, while the app that includes ChatGPT Work and Codex has around 20 million users, highlighting how early the agent market still is.
• The company sees model efficiency, product design and harness engineering as the keys to closing that gulf before competitors capture value.

The control trade-off

Getting the full value from an AI agent means granting it access to accounts and documents that were once considered private. Some users may hesitate to let an algorithm read an inbox or inspect a sales pipeline. For Andrew Ambrosino, the lead engineer for OpenAI’s desktop app, that control is not a boundary but a requirement. Ambrosino has linked ChatGPT to his email, Slack, phone, Notion and Figma, and he acknowledges the risks. A model might pull a private message into a public document, or it might act on incomplete information when permissions are misconfigured.

“If I’m asking it to write a document, is there a possibility that it’s going to pull from a private DM on that subject and not know that it’s not supposed to share some info? Yes,” Ambrosino said. “I’ll do it for the job. I will take the personal hit here and there if I have to. And I haven’t had to.”

That trade-off helps explain why OpenAI is moving cautiously. The company has made ChatGPT Work available on its lowest paid subscription tier, with a familiar chat interface wrapped around deeper automation. For users who are skeptical, allowing an AI system to act in the messy world of enterprise software feels like a large leap. OpenAI is betting that useful automation will outweigh those concerns, but early adoption data shows the market is still finding its balance.

The adoption gap

An OpenAI-backed study on Codex adoption found a striking disconnect inside the company. In June, 98 percent of OpenAI employees used Codex regularly. By contrast, only 17 percent of organizational subscribers and less than 1 percent of individual subscribers had adopted the agentic coding tool. The gap between near-total internal enthusiasm and negligible external uptake is both a warning and an opportunity. If OpenAI cannot convince normal knowledge workers to use agents, its expensive infrastructure will depend on a small slice of the software engineering market.

The company’s consumer ChatGPT service has more than one billion users, but the app that includes ChatGPT Work and Codex is used by about 20 million people. That means the vast majority of ChatGPT users still prompt a model for answers rather than trust it with multi-step tasks. OpenAI’s core product lead, Thibault Sottiaux, sees the transition as inevitable. He said during an interview that ChatGPT can now perform “entire, very complicated tasks for you all autonomously,” and that the core mission is “to bring everyone along.” Whether users are ready for that level of autonomy is still being tested.

The commercial case for agents

There is a direct financial incentive for OpenAI to make agents work outside programming. Agents burn through more tokens because they run longer and touch more tools, and that creates more revenue per user on the same subscription. But if coding is already a proven use case, it remains a small subset of all knowledge work. To justify massive investment in models and data centers, labs need accountants, doctors, marketers, lawyers and operations teams to use agents as part of their daily jobs.

That is why vertical competitors are moving into the gap. Companies focused on law, sales, research and customer support are building model-agnostic products, which lets them switch to whatever AI works best. If the biggest labs cannot quickly secure the “complementary assets” needed to deploy AI across the economy, value may accrue elsewhere. Analysts point to the distribution and workflow relationships that are being built now, not simply to the raw models, as the key battleground.

OpenAI sees ChatGPT Work as a direct answer to that threat. The product is a modified version of Codex, but it is designed for people who have never written a software test. Instead of asking users to embrace a command-line interface, the team has built a visual, conversational experience that offers a “magic box” for delegating tasks. But making the tool intuitive has required a lot of internal iteration. OpenAI’s non-engineering teams were among the first testers, and early versions were often hostile to them, showing empty technical diffs and asking them to edit code. Engineers gradually made the interface more general until it could handle finance, communications and operations work.

What the product does today

Use cases from early adopters range from practical to playful. Some venture capital investors ask the agent to pull together relevant emails and analyst research into an investment memo. Operations teams use it to build live dashboards and data visualizations without waiting for help from engineering. Sam Altman, OpenAI’s chief executive, reportedly uses it to plan vacations. One engineer asked the program to examine a Slack conversation about an engineering problem and “make some charts”; the agent returned useful plots without further instruction.

In a hands-on evaluation by reporters, ChatGPT Work handled a mundane personal task: extracting a child’s preschool calendar from an email and putting it into Google Calendar. It also built an auto-updating dashboard for financial metrics and created a searchable database of space launches, work that once demanded Python scripts. The friction came when setting permissions. Giving the agent the right access to a cloud drive was confusing and circular, with error messages that did not explain the problem. Many critical settings appeared only in the web app, forcing users to switch between desktop and mobile experiences.

The effort level is also an issue. OpenAI’s agent has multiple reasoning settings, and users who leave the effort level on low can quickly discover that the tool behaves like a careless intern. Joe Gershenson, engineering lead for OpenAI’s harness, acknowledged that the settings are not yet intuitive. “There are things that we can do better to help them get the right level of reasoning,” he said, adding that users should watch the space for improvements.

The Anthropic rivalry

OpenAI does not build in a competitive vacuum. Anthropic’s Claude Code helped define the modern coding agent category. Anthropic oriented its product around back-and-forth conversation, giving users choices at every stage and checking back before making large changes. OpenAI’s first Codex web app was more autonomous, betting that the underlying model could complete a project from a single prompt. That proved too far ahead of the capabilities of both the model and the harness. OpenAI later adjusted, adding more interactions and checkpoints, and the product began to catch up.

By some download measures, Claude Code was more popular until April of this year, and since then Codex has taken a slight lead. OpenAI’s engineers insist they do not study the competing interfaces closely, but the products mirror each other in obvious ways. When ChatGPT Work is launched, it even prompts new users to import data from Claude Cowork, Anthropic’s equivalent workplace agent. OpenAI says the real differentiator is its latest models, which are cheaper and more capable, allowing simpler harnesses to do more. The bitter lesson of AI research, they argue, is that a more powerful general model beats an elaborate set of domain-specific rules.

Harness engineering matters

Not everyone is convinced that a proprietary harness is worth owning. The layer of software around the model, called a harness, decides which tools the model can use, what context it receives and how it presents results. OpenAI employees say the goal of good harness engineering is to give the model exactly the information it needs and nothing more. They want to be “more precise” about context, not to build a complicated decision tree, because the next model update will make those rules obsolete.

Open-source alternatives show that model choice matters as much as the interface. Databricks and other evaluators have found that an open-source harness called Pi, published by the software company Earendil, can outperform Codex while using the same GPT 5.5 model. Pi has already served as the foundation for projects like OpenClaw and Cloudflare OS. Its creator, Mario Zechner, says his intentionally minimalist harness shows that an autonomous approach can work for software engineers. He also believes the big labs need to own the entire stack, from models to interfaces, or they will face intense price competition from Chinese model makers.

Zechner points out that coding is easier to train for because success is measurable and immediate. A manager who makes a decision today may not know the outcome for months, making it difficult for AI systems to learn from that work. That is one reason OpenAI’s push into white-collar knowledge work is hard. The workflows are messier, the data is incomplete and the outcomes are less traceable than code. OpenAI has built benchmarks from 44 occupations to evaluate progress, but the real test is how quickly ordinary subscribers learn to trust the agent.

The cost question

Agentic AI is expensive. In one reporter’s test, four days of moderate use on the $20 subscription burned more than 80 million tokens. The model itself estimated the cost at $65, more than three times the monthly price. OpenAI appears willing to subsidize that use while it builds habits and gathers data, but the economics will have to improve. Sottiaux says the company is pushing on efficiency every day, pointing to an 80 percent price cut for users of its Luna model. “If you wake up six months from now, you should be able to do all of the same things with less spend,” he said.

There is also the lock-in effect. Once a user connects every calendar, cloud drive and SaaS tool to an agent, switching to a rival involves real work. OpenAI is building the entire stack, including the model, the harness and the desktop app, and hoping that the convenience of the “magic box” keeps users inside its ecosystem. The same factors that make agents powerful, broad access and deep context, could eventually become the reason users hesitate to grant them even more control.

Inside OpenAI’s headquarters, the mood is busy and optimistic. Akshay Nathan, who leads the product engineering team, says knowledge workers are overwhelmed by the information stored in Salesforce, spreadsheets, email and chat platforms. The value of ChatGPT Work, he argues, is that it turns that previously hidden context into action. Nathan still acknowledges that complexity remains. “I still think there’s too much complexity,” he said. “I’m very optimistic that we can solve it, with the model and in a truly AI-native way.”

Source: TechCrunch News