How much control are you willing to give a large language model over your digital life? That is the central question behind OpenAI's newest product, ChatGPT Work. The company wants to move beyond simple chatbots and give every white-collar worker a set of AI agents that can actually do things: read email, update spreadsheets, schedule meetings, and juggle multiple apps.
For Andrew Ambrosino, the lead engineer for OpenAI's desktop app, the answer is simple: enough to test the future. He has granted the app access to his inbox, his Slack account, his phone, and tools like Notion and Figma. He acknowledges the risks. If he asks the model to write a document, it might unknowingly pull from a private message or share something it should not. But he is willing to take that chance. "I'll do it for the job," he says. "I will take the personal hit here and there if I have to. And I haven't had to."
What Is ChatGPT Work?
ChatGPT Work, released last month, is OpenAI's biggest bet on the future of knowledge work. It is available on the company's lowest subscription tier at $20 a month. The product is built to let accountants, investors, doctors, and others connect AI agents to the digital workflows they use every day. Instead of just answering questions, the model can complete multistep projects on its own.
OpenAI's marketing describes the goal as a world where artificial intelligence goes beyond answering questions to helping everyone turn their biggest ideas into reality. The product is a modified version of Codex, OpenAI's coding tool. Software engineers already use agents to write, test, and deploy code; ChatGPT Work aims to bring that same capability to people who do not write code.
"In this new factor, ChatGPT can actually do entire, very complicated tasks for you all autonomously in a way that is delightful and safe," said Thibault Sottiaux, who leads OpenAI's core product work, including Work. "It's the very mission of OpenAI — to bring everyone along."
Adoption Numbers Tell a Mixed Story
The gap between internal enthusiasm and external adoption is striking. An OpenAI-backed study found that in June, 98% of OpenAI employees were using Codex. But only 17% of organizational subscribers and less than 1% of individual subscribers had used the agentic coding tool. That difference represents both a challenge and an opportunity.
Commercially, agents that work longer are more lucrative because they consume more tokens. Reaching new professions is crucial, not just for OpenAI but for the entire AI industry. Coding has been profitable territory for AI labs, but it is a tiny slice of the professional work that AI tools need to enable if companies are to justify massive investments in training and computing.
Vertical-specific competitors like Harvey for law and Clay for sales have been chasing those customers with a model-agnostic approach. They can plug in whichever AI model works best at any given time. Industry analysts warn that if AI labs cannot rapidly control the key complementary assets needed to scale AI in the market, value will accrue elsewhere.
Making AI Intuitive for Non-Engineers
Most people do not use command-line interfaces. There is a reason Windows replaced DOS. To reach a mass audience, AI products need to be more intuitive. Ambrosino describes an agentic product as "something that plays with the messy world of your life and your tools and websites that were built in 1995 and never updated."
OpenAI's non-engineering workforce, including communications and finance teams, initially struggled with Codex. They were asked about code and shown empty diffs, a technical readout meant for software changes. The team spent months making the tool more general purpose.
Ambrosino compares the design approach to skeuomorphism, the practice of making digital tools look like physical objects they replaced, like a calculator app designed to look like a pocket calculator. "That stuff wasn't just cringe design. That actually helped get people into this and make the transition," he says.
ChatGPT Work has a few more buttons for selecting projects and plug-ins, but it aims for the same magic box interface as other OpenAI products. Ambrosino says discoverability matters at this early stage. "We push back on [the idea that buttons are unnecessary] — because it's very early," he explains. "At some point we won't have the button."
What Can ChatGPT Work Do?
OpenAI employees are using the tool for routine, data-intensive tasks. They set up weekly metrics reports and turn spreadsheets into planning tools. Venture capitalists use it to assemble communications and analysis about companies into investment memos. Operations teams spin up bespoke dashboards and data visualizations. Sam Altman, OpenAI's CEO, uses it to plan vacations.
Akshay Nathan, who leads the product engineering team at OpenAI, says workers face a deluge of information. "We're actually quite limited by our ability to parse everything that's available to us, and then take action on it," he says. The value of ChatGPT Work is that it gives users true access to the information stored in systems like Salesforce and other records tools.
In practice, the system can be impressive. One reporter asked ChatGPT Work to get a preschool calendar out of an inbox and put it into Google Calendar. It did, saving a lot of repetitive data entry. The same reporter used it to do financial analysis on publicly traded companies, create a queryable database of space launches, and send a weekly email about new AI research.
Limitations and Frustrations
Setting up permissions is not always easy. Giving the model read-only access to a cloud drive can be confusing and circular. Many important settings are only available on the web app, forcing users to work in both web and mobile at the same time. Some limitations seem arbitrary: when linked to Google Calendar, the model can create events but not new calendars.
The effort setting also matters. If the effort level is not set to high, the model can feel like the worst intern you have ever worked with. Joe Gershenson, the engineering lead for OpenAI's harness, admits that effort settings are not intuitive for new users. "There are things that we can do better to help them get the right level of reasoning," he says. "Watch this space."
Evaluation Is Harder Outside Coding
One of the biggest challenges is that white-collar work is not as measurable as code. Software either works or it does not. A good presentation, business strategy, or sales pitch is much harder to evaluate or trace. OpenAI uses a benchmark called GDPval, drawn from 44 occupations and hundreds of knowledge work tests, but it also relies heavily on user feedback.
Ambrosino says OpenAI employees themselves are an important source of insight. "We have to always parse out ... are we doing the workflow that everybody else will be doing, or are we weird?" As early adopters use the app, their behavior creates valuable traces for training and improvement, assuming they do not opt out of data collection.
The Rivalry With Claude Cowork
OpenAI's engineers are reluctant to talk about competitors. But the influence of Anthropic's Claude Code is hard to ignore. Claude Code defined the market for AI coding and launched a revolution in how software engineers work. It is a sensitive topic at OpenAI because the company had the idea first but did not execute it perfectly.
When OpenAI first developed Codex as a web app, the engineers were, in Ambrosino's words, "a bit more AGI-pilled." They bet on the model being smart enough to handle tasks entirely on its own with minimal user input. Anthropic's Claude Code took a different approach. It was oriented around a back-and-forth conversation with the user, offering options and checking in frequently. That approach proved more effective, even though it demanded more work from users.
OpenAI eventually followed suit by adding more opportunities for users to interact with the model. Download statistics suggest Claude Code was more in demand until April of this year, but Codex has since taken a slight lead. Surveys of enterprise use also suggest OpenAI is catching up. Some of that gain comes from product-market fit, while some comes from complaints about safety restrictions and compute shortages on Anthropic's side.
What Makes a Good Harness?
Every large language model requires a "harness" — the software wrapped around the model that decides what information it sees, which tools it can use, and how it presents answers. A good harness gives the model just enough context and no more. Gershenson says his team focuses on the simplest ways to expose the model to the tools it needs. He points to the "bitter lesson" of AI research: a better general model is more important than specific domain experience.
"You could get good results in the short term by adding a whole bunch of extras — if and thens and tools — but like, come on, the next model is going to come out in a couple of months and make that obsolete," Gershenson says.
Not everyone agrees that a model-specific harness is the best approach. Open-source harnesses like Pi have outperformed Codex on coding benchmarks while using the same model. Pi's creator, Mario Zechner, says his minimalist harness succeeds because it can modify itself and build its own interfaces. He believes AI labs are trying to own the entire stack to avoid competing with cheaper Chinese models.
Zechner also notes that coding agents benefit from abundant training data. "Everything is coding agent shaped ... the reason is that they only have training data for coding agent tasks," he says. Tasks in management or other vague fields are harder to capture in simple traces of user-agent interaction.
The Cost Question
Token costs remain a major issue. A reporter using a $20-a-month subscription consumed more than 80 million tokens in four days, which the model estimated cost $65. That is a subsidy of more than three times the subscription price for just a few days of casual use.
Sottiaux says OpenAI is working to push the frontier on efficiency, pointing to an 80% price cut for users of the Luna model. "If you wake up six months from now, you should be able to do all of the same [tasks] with less spend."
There are also concerns about lock-in. Once users configure all their plug-ins and permissions, they may be reluctant to switch products. Data retention and the sheer pain of reconfiguring access could keep customers tied to one ecosystem.
Inside OpenAI's headquarters, the atmosphere is calm but slightly tense. Engineers constantly monitor laptops and rush between meetings. Nathan says the focus remains on "the promise of the magic box, but I still think there's too much complexity." He is optimistic that the company can solve it "with the model and in a truly AI-native way."
Source: TechCrunch News