que paso
TechCrunch reports that OpenAI released ChatGPT Work last month as a $20-a-month product designed to let non-engineers use AI agents across workplace tools. The product can access services such as email, Slack, Notion, Figma and calendars, and is intended to complete multistep tasks rather than only answer questions.
TechCrunch reports that OpenAI’s ChatGPT Work, released the previous month, is a modified version of the company’s Codex coding tool aimed at people outside software engineering. The product connects an AI model to digital workflows and services, allowing it to read information, use tools and carry out multistep projects. OpenAI employees told TechCrunch that the system is being used for weekly metrics reports, planning spreadsheets, investment research, dashboards, data visualizations and vacation planning. The account describes the product as part of OpenAI’s broader attempt to make agentic software useful to accountants, investors, doctors and other workers whose jobs depend heavily on computer systems.
The report identifies a sharp difference between internal and external use of OpenAI’s existing coding agent. According to an OpenAI-backed study cited by TechCrunch, 98% of OpenAI employees used Codex in June, compared with 17% of organizational subscribers and less than 1% of individual subscribers. TechCrunch says OpenAI would not disclose how many people used ChatGPT Work compared with Codex. The joint app was described as having about 20 million users, while OpenAI says more than one billion people prompt ChatGPT online. Those figures come from the company or an OpenAI-backed study and are not independently verified in the source.
TechCrunch also reports hands-on testing of ChatGPT Work. The reporter says the system transferred information from a child’s preschool calendar in email into Google Calendar, created a financial-analysis dashboard, built a queryable database of space launches and sent a weekly email about new AI research. The same testing found confusing permissions, repeated errors when attempting to grant limited read access, a requirement for complete access in at least one case, and settings split between web and mobile applications. TechCrunch reports that the product could create calendar events but not new calendars, and that lower effort settings produced poor results. These are the reporter’s observations, not an independent product audit.
Lea la fuente principal: techcrunch.com ↗
Por qué es importante
The report describes OpenAI’s effort to move agentic AI beyond software development into routine professional work. It also documents a substantial gap between adoption inside OpenAI and usage among outside subscribers, alongside unresolved questions about permissions, privacy, reliability and operating costs.
TechCrunch’s report shows why AI agents could have a different economic and practical impact from chatbots that only generate answers. An agent connected to workplace systems can retrieve information, transform it and take action, potentially reducing repetitive coordination work. The report says longer-running agents consume more tokens, making them more valuable per user to AI companies, while expansion beyond coding is important to justify the industry’s large investments in training and computing. The practical benefit, however, depends on whether users can grant appropriate access and understand what the system may do.
The article places the product-design challenge in the software surrounding the model, often called a harness. TechCrunch reports that this layer determines what information the model receives, which tools it can use and how it interacts with users. OpenAI engineers told the publication that mainstream users need clearer buttons, project controls and guidance because most people do not work through command-line interfaces. The report contrasts OpenAI’s approach with Anthropic’s Claude products, which it says more often present options, request feedback and proceed in smaller steps. TechCrunch also cites comparisons suggesting that different combinations of models and harnesses can produce different results, meaning the model alone may not determine performance.
The risks become more consequential when an agent receives access to personal or organizational systems. TechCrunch reports that OpenAI’s desktop-app engineering lead acknowledged the possibility that an agent could draw on a private message and share information inappropriately, even though he said that had not happened in his use. The article also describes uncertainty about how to evaluate presentations, business strategies and sales work, which lack the clear pass-or-fail criteria available for much software. TechCrunch reports that OpenAI uses GDPval, a benchmark covering 44 occupations and hundreds of knowledge-work tests, alongside user feedback, but the source does not establish how well that benchmark predicts real-world outcomes. Privacy protections, retention practices and the consequences of mistakes therefore remain important unknowns.
Qué ver a continuación
The key tests are whether ordinary users can configure access safely, whether agents can complete less measurable knowledge-work tasks reliably, and whether subscription pricing can support long-running use. OpenAI has not provided a public breakdown of ChatGPT Work adoption, and the report’s usage and cost figures are not independently confirmed here.
Adoption evidence will be the first meaningful test. TechCrunch reports that OpenAI has not disclosed separate usage figures for ChatGPT Work and Codex, leaving unclear whether the product is reaching nontechnical users at scale or mainly attracting existing AI enthusiasts. Independent measures of active use, repeat use, task completion and user abandonment would help distinguish a widely useful workplace product from a heavily marketed experiment. The internal-versus-external gap reported in the article is significant, but the underlying study and company figures should be treated as attributed claims rather than settled market data.
Permissions and data governance deserve close scrutiny as the product expands. The report describes difficulty granting narrow access to a cloud drive and a need to use both web and mobile interfaces for important settings. Future reporting should establish whether OpenAI adds clearer permission scopes, activity logs, confirmation steps, data-retention controls and accessible ways to opt out of training use. These safeguards matter because an agent that can read email, messages, files and business applications can create privacy or operational harms even when its underlying model is generally capable.
Cost and reliability will determine whether the service can scale. TechCrunch reports that its reporter used more than 80 million tokens in four days and that the model estimated a $65 cost, despite testing a $20 monthly subscription; that calculation was not independently verified by the source. OpenAI said it was working to improve efficiency and cited an 80% price cut for users of its Luna model. Watch for transparent usage meters, predictable pricing, independent evaluations across occupations and evidence that agents can handle long-running tasks without excessive supervision. The report leaves unresolved whether the economics work for users, OpenAI or both.
