How much control are you willing to give a large language model (LLM) to access and act on your digital life? OpenAI engineers argue that to get meaningful, multistep automation, models often need broad access — to email, Slack, calendars and various SaaS apps.
Andrew Ambrosino, lead engineer for OpenAI’s desktop app, said he accepts personal account access for the sake of testing: the app now has access to his inbox, Slack, phone and apps such as Notion and Figma. “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 told TechCrunch. “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.”
ChatGPT Work: release and positioning
OpenAI released ChatGPT Work last month; it is available on the company’s lowest subscription tier at $20 per month. The product is intended to let white-collar workers field AI agents that hook language models up to the digital workflows used by accountants, investors, doctors and other knowledge workers so the agents can complete multistep projects autonomously.
OpenAI’s marketing frames the aim as a world “where [artificial] intelligence goes beyond answering questions to helping everyone turn their biggest ideas into reality.” ChatGPT Work is a modified version of Codex, the company’s coding tool, intended to provide non-engineers with the same agentic functionality software engineers already use.
Adoption gap: inside the company vs. outside
An OpenAI-backed study found that in June 98% of OpenAI employees were using Codex, compared with 17% of organizational subscribers and less than 1% of individual subscribers. That gap — near-total internal adoption versus negligible external uptake — represents both the challenge and the opportunity for OpenAI.
Thibault Sottiaux, who leads OpenAI’s core product work, said the company expects that “the more value and the more utility that we generate for users, the more they will be willing to also pay for some part of that utility,” tying user value to the subscription model.
The harness: making LLMs useful in practice
Every LLM needs a harness — the software layer that decides what information the model sees, which tools it can use, and how it presents results. To turn an LLM into an agent, the harness supplies tools and instructions for long-term tasks.
For developers, a command-line interface (CLI) that enables coding changed how software is built, but most users don’t work via CLIs. Ambrosino stressed that agentic products must “play with the messy world of your life and your tools and websites that were built in 1995 and never updated” to broaden access beyond engineers.
ChatGPT Work offers additional buttons for selecting projects and plug-ins but aims for the same single-box interface as other OpenAI products. Ambrosino compares this to skeuomorphism — design choices that mimic physical objects — arguing that such familiar cues can ease transitions for mainstream users.
Practical uses and user experience
Inside OpenAI, employees use agents to set up weekly metrics reports and turn spreadsheets into planning tools. Venture capitalists assemble communications and analysis into investment memos with agents; ops teams build bespoke dashboards and visualizations; even Sam Altman reportedly uses the tool for vacation planning. The article’s author asked ChatGPT Work to extract a strangely formatted preschool calendar from email and insert it into Google Calendar; the agent completed the task and removed repetitive data entry.
However, granting the model the permissions it needs is not always simple. The author tried multiple times to give the agent “read” access to a cloud drive but encountered errors; eventually the mobile app required full access to make the integration work. Many settings are only available on the web app, forcing users to switch between interfaces. Some specific limitations are puzzling: linked to Google Calendar, the agent can create events but not new calendars. The product can feel like “the worst intern you’ve ever worked with” unless the user is prepared to invest effort.
Joe Gershenson, the engineering lead for OpenAI’s harness, acknowledged that effort settings and other UX elements are not yet intuitive for new users: “there are things that we can do better to help them get the right level of reasoning… Watch this space.”
Measurement, benchmarks and data
OpenAI told TechCrunch it uses a GDPval benchmark drawn from 44 occupations and hundreds of knowledge-work tests and supplements that with user feedback. A less formal source of product direction is internal usage: Ambrosino said the team constantly asks whether they’re building workflows that “everybody else will be doing” or idiosyncratic internal tools. Early adopters’ usage will create valuable traces for product improvement, provided those users do not opt out of making data available for training (the author opted out).
Competition, harness design and the role of the model
OpenAI engineers downplayed close comparisons to competitors like Anthropic’s Claude Cowork and Claude Code, but the products share many UI similarities. Anthropic’s approach, oriented around iterative back-and-forths and frequent checks for user input, proved effective in the market: Claude Code led in downloads until April, after which Codex took a slight lead, and enterprise surveys indicate OpenAI is catching up.
Some external projects demonstrate that different harness-model combinations yield varied performance. Databricks reported that Pi — an open-source harness by Earendil — outperformed Codex while using the same GPT-5.5 model in certain benchmarks. Pi’s creator, Mario Zechner, argues that minimalist harnesses that let models self-modify can work well for coding tasks and warns that labs pushing proprietary harnesses aim to lock in users by owning the full stack.
OpenAI engineers emphasize the continuing importance of model quality: often the differentiator is “the model,” and the app’s design tries to fully leverage the latest powerful and cost-effective models.
Cost and scale considerations
Token consumption is a material concern for scaling agentic products. In one trial, the author used more than 80 million tokens in four days, which the model’s analysis indicated cost $65 — over three times the $20 monthly subscription for only a few days of casual use. OpenAI says it is “working every day to push the frontier on efficiency,” pointing to a recent 80% price cut for users of the Luna model as an example.
Another question is whether extensive data retention and the pain of configuring numerous plugins and permissions create customer lock-in. These friction points could make it difficult for users to switch providers once they have invested significant setup effort.
Where the product stands and next steps
Inside OpenAI’s headquarters — visited by the author in July — engineers appeared busy and focused. Product engineering lead Akshay Nathan said the company is still pursuing “the promise of the magic box,” while acknowledging that too much complexity remains and expressing optimism that it can be solved with better models and AI-native approaches.
In sum, ChatGPT Work is an ambitious push to extend agentic AI beyond engineering teams into mainstream knowledge work. Internal adoption is high, but external uptake lags due to usability, permissions, cost and measurement challenges. OpenAI’s roadmap emphasizes improving harness simplicity, model efficiency and the user experience to broaden appeal and make large-scale, agent-driven workflows viable.
