OpenAI is making its biggest push yet to bring AI agents to the everyday knowledge worker. The company has released ChatGPT Work, an agentic product designed to sit on top of its large language models and take over digital workflows. For $20 a month, users can connect the system to their email, calendar, Slack, and a range of business tools, allowing it to complete multistep tasks that once required manual effort. The objective goes beyond answering questions: the model is meant to act as a digital employee that helps people turn ideas into results.

The core question for users is how much control they are willing to give a language model over their digital life. Andrew Ambrosino, lead engineer for OpenAI’s desktop app, has already handed over access to his inbox, Slack, phone, Notion, Figma, and other tools. He understands the risk: the model could pull from a private DM or share information it should not, especially if asked to write a document. Nevertheless, he says the reward outweighs the potential embarrassment, and he has not yet lost that bet.

ChatGPT Work is an adaptation of Codex, OpenAI’s coding agent. Codex demonstrated that artificial intelligence can handle complex software engineering tasks when given access to a command line and the right permissions. ChatGPT Work expands that capability beyond software developers. It is designed for accountants, investors, doctors, operations managers, and anyone else whose daily work is dominated by a computer. Thibault Sottiaux, OpenAI’s product leader for the effort, says the model can complete very complicated tasks autonomously and safely. He describes it as the fulfillment of OpenAI’s founding mission: giving everyone access to advanced machine intelligence, not just programmers.

The commercial reasoning is straightforward. An agent that works for hours instead of a few minutes will consume far more tokens, making each subscriber more valuable. OpenAI must reach professions beyond coding, however. Software engineering is a relatively small slice of the white-collar workforce. Meanwhile, specialists such as Harvey, which builds legal AI tools, and Clay, which sells sales workflow tools, are already pursuing those vertical markets. They use whichever model performs best for a given client, which means they do not necessarily need to buy model access from a single large lab.

That dynamic has prompted analysts to warn that OpenAI and its competitors will not automatically capture the value of the AI boom. If the labs cannot control the full workflow, from model to interface to business process, value may shift to companies that own the customer relationship and understand domain-specific needs. ChatGPT Work is partly a response to this threat. By controlling the harness — the software that wraps around the model, manages context, and invokes tools — OpenAI keeps itself close to the user.

A step beyond coding

The internal adoption data suggests why OpenAI needed a product like ChatGPT Work. An OpenAI-backed study found that in June, 98 percent of OpenAI employees used Codex. Among organizational subscribers, only 17 percent had used it, and less than 1 percent of individual subscribers had. That gap between internal and external adoption is the clearest possible indication that coding tools do not naturally appeal to non-engineers.

One reason is the interface. The first versions of Codex were unwelcoming to people who are not programmers. Non-coding employees inside OpenAI encountered messages about empty diffs and other technical terminology. The team spent months making the tool more general purpose, teaching it to understand natural language prompts rather than requiring users to think in code.

Many non-technical users have never touched a command-line interface, and that is fine. A mainstream agent needs to work with the messy world of existing software: web pages built in the 1990s, half-integrated SaaS apps, permission systems, and legacy tools. Ambrosino’s team believes the product experience itself is the key to bringing people along. He compares early agent interfaces to skeuomorphism, the practice of making digital tools resemble physical objects. That approach may look unnecessary over time, but it helps newcomers cross the psychological gap between using software and commanding an AI.

The current product has more buttons than a pure conversational interface. Users can select projects, manage plugins, and inspect the tools the model plans to use. OpenAI’s designers defend these extra controls because discoverability matters in the early stages of a new technology. Eventually, as users learn to trust the model, the buttons may disappear.

What people do with the agents

OpenAI pitches ChatGPT Work as a tool for routine, data-intensive coordination. Within the company, employees are using it to generate weekly metrics reports, turn spreadsheets into planning dashboards, and analyze Slack conversations after engineering debates. Outside OpenAI, early adopters report using the product for investment memos, data visualizations, and a wide range of annoying administrative work. Sam Altman is reportedly using it to plan vacations, a task that would require pulling together flights, hotel offers, and restaurant reservations from many sources.

Akshay Nathan, who leads OpenAI’s product engineering team, argues that modern employees are drowning in information stored in systems such as Salesforce and corporate wikis. They cannot parse all of it before acting. ChatGPT Work offers a way to search across dozens of pieces of context and distill them into something useful. This vision is not unique. Anthropic’s Claude Cowork and Perplexity’s browsing agent offer similar capabilities, and the competition is fierce.

The utility is real, but limited by the user’s willingness to grant access. In one detailed test by a journalist, ChatGPT Work successfully extracted a preschool calendar from email and placed every event into Google Calendar, saving considerable repetitive work. The same reviewer asked it to compile an auto-updating dashboard of financial metrics and to make a queryable database of space launches that had previously required writing Python code. The program even sent a weekly digest of new academic AI research. But the reviewer deliberately withheld access to personal inbox, source interviews, and bank accounts, which made the model less effective than it could have been.

The experience is not seamless. Connecting a cloud drive can be confusing, and some crucial settings are found only on the web version. Linking Google Calendar allows the model to create events, but not to create new calendars. With the effort level set too low, outputs feel sloppy, like the work of a careless intern. One OpenAI engineer, Joe Gershenson, admits that effort settings are not yet intuitive for new users and says the company plans to improve the reasoning levels in future updates.

Rivalries and lessons

OpenAI’s engineers are reluctant to discuss Claude Cowork or Claude Code, Anthropic’s high-profile agents. Asked what differentiates ChatGPT Work from those products, they mostly say they focus on their own work. One engineering leader joked about the Mad Men scene in which a character claims never to think about his competitors. The claim is hard to believe, because the products are similar and ChatGPT Work’s signup flow invites new users to port data from Claude Cowork.

Anthropic deserves credit for defining the modern coding agent. OpenAI initially built Codex as a web app that attempted to do too much without user input. Anthropic’s Claude Code, by contrast, asked the user for choices, checked back after each step, and generally kept humans in the loop. That iterative approach worked better, and OpenAI eventually redesigned Codex along the same lines.

The competitive balance shifted after months of complaints from Anthropic users about safety restrictions and compute shortages. OpenAI improved its model, cut prices, and invested in making the harness more flexible. By the spring of this year, Codex and Claude Code were roughly matched in download volume, and enterprise adoption of OpenAI’s agent was accelerating.

The model and the harness

Under the hood, every major AI lab must decide how much engineering should go into the harness. OpenAI’s engineers subscribe to what the industry calls the bitter lesson: a larger, smarter model will beat clever software scaffolding in the long run. Gershenson says that adding many custom rules and tools can improve results for a few weeks, but then a new model arrives and makes those extra parts obsolete. His team therefore tries to expose only the information the model truly needs, avoiding unnecessary constraints.

Not everyone agrees that harness engineering is merely a temporary aid. Third-party evaluations by companies such as Composio and Databricks show that different harness and model pairings produce different benchmark results. Databricks found that Pi, an open-source harness developed by the software company Earendil, outperformed ChatGPT Work’s Codex tool even when it used the same GPT-5.5 model underneath. Pi has also been used to build popular tools like OpenClaw and the Cloudflare OS.

Pi’s creator, Mario Zechner, argues that the big labs promote their own harnesses because they want to own the full product stack. If users can swap in any model, he says, the labs become mere model providers and face brutal competition from lower-cost open-source models. Zechner also sees a fundamental reason that coding agents have advanced farther than general white-collar agents: coding generates abundant training data and provides clear feedback. A decision made by a manager might not show its consequences for months, which makes it harder for a model to learn from the result.

Costs and permissions

One important issue for any independent evaluation is the opaque economics of agentic

Source: TechCrunch News