OpenAI's new ChatGPT Work product is an attempt to turn large language models into autonomous digital coworkers. Launched in July on the company's lowest subscription tier, the $20-per-month product connects AI agents to services such as email, Slack, calendars, Notion, Figma and other enterprise software. “How much control are you willing to give an LLM over your digital life?” is the central question the product asks. For Andrew Ambrosino, the lead engineer on OpenAI's desktop app, the answer is plenty. He says his own ChatGPT Work setup has access to his inbox, Slack, phone and design tools, and he accepts the privacy risk as part of testing the product.
“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.”
Thibault Sottiaux, who leads OpenAI's core product work, says the app can complete entire, complicated tasks autonomously. “It's the very mission of OpenAI — to bring everyone along,” he said.
From code to the corner office
OpenAI's marketing describes the goal in grander terms: a world in which artificial intelligence moves beyond answering questions to helping people turn their biggest ideas into reality. In practice, that means hooking up large language models to the workflows used by accountants, investors, doctors and other professionals whose work happens largely on computers.
ChatGPT Work is a variation of the company's coding agent, Codex, modified for people who are not software engineers. Instead of merely generating replies in a chat window, it is supposed to execute multistep projects. It can scan messages, pull records, update spreadsheets, create charts and send follow-ups using whatever tools the user grants it access to.
The commercial stakes are significant. Agents that run for longer stretches consume more tokens, making them more lucrative for OpenAI on a per-user basis. But the industry's current customer base is still narrow. Coding has proved valuable for AI labs, yet software engineers are a small slice of the overall professional workforce. Vertical competitors such as Harvey in law and Clay in sales have pursued those specialty markets with model-agnostic products, plugging in whichever AI performs best at any given time. Some investors and analysts have warned that if frontier labs do not control the distribution and integration points in the market, much of the economic value may flow to those intermediaries.
A study backed by OpenAI found that in June 2026, 98% of OpenAI's own employees used Codex, but only 17% of organizational subscribers and less than 1% of individual subscribers were using the agentic coding tool. The near-total adoption inside the company, paired with negligible use outside it, captures both the challenge and the opportunity for the company.
To understand that disconnect, it helps to understand what OpenAI's engineers are building. Every large language model needs what engineers call a harness — the software wrapped around the model that decides what information it sees, which tools it can call and how it presents results. For developers, a command-line interface was enough to transform software development. But most people left command-line interfaces behind when graphical operating systems replaced DOS. An AI agent meant for the general workplace has to work with what Ambrosino calls the messy world of people's lives: old websites, quirky internal tools, overlapping logins and permissions.
That has direct consequences for interface design. OpenAI pairs Work with a set of buttons for choosing projects and plug-ins, even though some employees argue that a button is unnecessary because users can simply ask the model directly. Ambrosino says discoverability matters in this early phase. He compares the design approach to skeuomorphism, in which digital tools imitate the physical objects they replaced, such as calculator apps styled like pocket calculators. That was not merely decoration, he argues; it helped people make the transition to new capabilities.
What the first users are doing
Early use cases are data-intensive and routine. OpenAI employees use the app to assemble weekly metrics reports, turn spreadsheets into planning tools and build dashboards. Investors have used agents to pull relevant communications and analysis into investment memos. Operations teams have created custom data visualizations. One OpenAI engineer described asking the program to look at a Slack conversation about an engineering problem and “make some charts,” then receiving back a series of insightful plots.
The appeal is easy to see in a workplace flooded with information. Akshay Nathan, who leads OpenAI's product engineering team, says knowledge workers are limited by their ability to parse data across Salesforce, Slack, documents and other records. A model that can traverse those tools effectively makes existing software more useful, he argues.
The product still has rough edges. In a hands-on trial of ChatGPT Work, setting permissions for services like cloud drives was confusing. The model gave unhelpful error messages and the user was eventually told that only complete access would make the integration work. Some features worked only in the web app while others were mobile-only. The tool could create calendar events but could not create new calendars. Users were advised to set the highest effort level, otherwise the agent behaved like an unreliable intern.
That kind of friction is common among early AI products. Joe Gershenson, the engineering lead for OpenAI's harness, admits that the effort settings are not intuitive for new users. There are, he says, things the team can do to help people choose the right level of reasoning. He also expects those choices to become simpler as models improve.
The evaluation problem is also different outside coding. Software either works or it does not, giving developers a clear signal about agent quality. There is no equivalent test for a good presentation, a sensible business strategy or an effective sales pitch. OpenAI says it uses a benchmark called GDPval, drawn from 44 occupations and hundreds of knowledge work tasks, supplemented by user feedback. Less formally, the company relies on its own employees to test workflows before external users do. Ambrosino acknowledged that his team has to constantly ask whether the workflows they are building are genuinely useful to a wide population or only to a group of unusually technical people.
The competition that shaped its design
OpenAI's largest rival came at the same market from a slightly different direction. Anthropic's Claude Code defined the market for agentic coding and started a revolution in how software engineers work. Its harness was built around a back-and-forth conversation with the user. When someone gave it a problem, it would survey the possibilities and suggest several approaches. After the user chose one, it would proceed a little further and check back again, leaving less room for large errors.
OpenAI's earlier version of Codex was more autonomous and, according to Ambrosino, slightly ahead of what the underlying model could reliably handle. “A bit more AGI-pilled” is how he describes that original design. The interface assumed the model was smart enough to handle an entire task with minimal input. Anthropic's more collaborative loop proved more effective in practice, even though it demanded more work from users. OpenAI eventually followed suit by adding more opportunities for people to interact with the model and correct its course. Download statistics suggest that Claude Code was more in demand until early this spring, when Codex took a slight lead.
The same dynamic is now playing out in workplace agents. OpenAI engineers insisted they do not spend much time studying rival harnesses, pointing instead to the strength of the company's latest models. Yet the resemblance between products is noticeable. When ChatGPT Work first launched, its onboarding flow asked users whether they wanted to migrate their data from Claude Cowork, a subtle acknowledgment of the competitive landscape.
Whether a proprietary harness is even necessary remains a contentious question. Benchmarks run by companies like Composio and Databricks show that different combinations of models and harnesses produce different performance on coding tasks. Databricks found that Pi, an open-source harness published by the company Earendil, outperformed Codex while using the same GPT-5.5 model. Pi has been used to build tools such as OpenClaw and Cloudflare OS.
Mario Zechner, Pi's creator, argues that the intentionally minimalist harness is evidence that a less decorated approach can work well for software engineering. He also thinks frontier labs are limited in their ability to train on non-coding professional work because much of it leaves no digital trace. “Everything is coding-agent shaped,” he said. Management decisions, for example, unfold over months and cannot be captured in a simple transcript of a user and agent talking and acting.
Zechner and other open-source developers say that big labs want to own the entire stack so they do not become model providers competing with cheaper alternatives from China. The labs, for their part, contend that a well-designed harness simply lets the model do its best work. Gershenson says his team deliberately avoids adding many brittle rules and shortcuts because a better model will make them obsolete. The goal of good harness engineering, he says, is to give the model precisely the context and tools it needs, then let it solve the problem.
Cost remains an open question for agentic software. In one test on a $20-per-month subscription, a user consumed more than 80 million tokens in four days. The model calculated that same usage would have cost about $65 on listed prices, more than three times the subscription fee for a short period of casual use. Sottiaux says OpenAI is working on efficiency, pointing to an 80% price cut for users of its Luna model. He argues that the same set of tasks should become significantly cheaper to run over the coming months.
At OpenAI's headquarters in late July, the atmosphere was busy but tempered. Engineers monitored laptops between meetings, and product leaders acknowledged that complexity remains. Nathan says the team is still focused on the promise of a simple magic box, but wants to move beyond the current level of setup. He is optimistic that future models and truly AI-native design will solve the problem.
Source: TechCrunch News