OpenAI Launches ChatGPT Work: Long-Running Agentic Task Automation

OpenAI launched ChatGPT for Work on July 9, 2026 — a rebranding of Codex as a unified desktop super-app — alongside GPT-5.6 in three capability tiers (Sol, Terra, Luna) [1]. The tool is designed to run multi-step professional tasks over hours, integrating with Slack, Teams, Google Drive, SharePoint, email, and CRMs via plugins, and includes Scheduled Tasks and a Sites feature for publishing interactive web apps [3][2]. The launch drew broad media coverage [4][6] but also immediate expert confusion: both Ethan Mollick and Simon Willison said publicly they could not parse the distinction between ChatGPT for Work and Codex [1][7]. OpenAI published enterprise case studies from Australian Payments Plus and Deutsche Telekom around the launch to support adoption claims [8][9].

ChatGPT for Work is OpenAI's clearest attempt to shift from a reactive chat tool to a sustained autonomous agent embedded in enterprise workflows. The product identity confusion at launch — acknowledged even by close expert observers — points to a communication problem that could slow enterprise evaluation, regardless of the underlying capabilities.

OpenAI launched ChatGPT for Work on July 9, 2026, as part of what The Neuron called a 'Super Thursday' that also included the release of GPT-5.6 in three tiers — Sol (most capable), Terra (balanced), and Luna (affordable) [1]. ChatGPT for Work is a rebranding and expansion of Codex into a unified desktop super-app with web browsing, file editing, multi-agent coordination, and scheduling built in [1]. The tool is designed to stay engaged with professional projects for hours, directly addressing a limitation in the earlier Atlas Agent Mode, which stopped automated tasks after a few minutes [2]. It integrates with Slack, Microsoft Teams, Google Drive, SharePoint, email, and CRM systems via plugins, and includes a Scheduled Tasks feature for recurring automation and a Sites feature, in public beta, for publishing interactive web apps [3]. OpenAI requires human approval before the tool executes important actions [2].

The launch drew wide media attention [4][5][6] but also produced immediate confusion about what the product actually is. Grant Harvey at The Neuron noted that both Ethan Mollick and Simon Willison — two observers who follow the space closely — said publicly they could not understand the distinction between ChatGPT for Work and Codex [1]. Willison separately identified a concrete UX fragmentation: Work conversations started on web or mobile run in the cloud and do not appear in the desktop Work interface, while the desktop app can additionally access local files and desktop applications [7]. Willison found OpenAI's own explanation of this split insufficient. Harvey also flagged memory reliability as a structural problem, writing that until memory becomes reliable enough, the app will feel like 'dude where's my chat because somebody rustled all the papers around on your desk' [1].

OpenAI has positioned the launch with enterprise validation evidence. Australian Payments Plus, in a case study published July 7, reported that 77% of surveyed employees save more than two hours per week using ChatGPT Enterprise, and that Codex reduced a complex reconciliation investigation from four hours to thirty minutes [8]. AP+ employees created more than 300 custom GPTs and more than 1,000 Projects, suggesting broad organizational rollout rather than isolated experimentation [8]. Deutsche Telekom announced a partnership with OpenAI around the same time, with a stated goal of becoming an 'AI-native' telco across customer service, employee workflows, and network operations [9]. Both are OpenAI-published promotional pieces with no independent verification. OpenAI also reports more than 5 million weekly Codex users, with over 1 million using it for non-software tasks, and says nearly 100% of internal teams across finance and sales now rely on the product [3] — figures that have not been independently verified.

The competitive backdrop is shifting. The Neuron noted that Meta released Muse Spark 1.1 with a 1M-token context window and computer use at $0.80 per million input tokens, applying direct pricing pressure on OpenAI and Anthropic in the enterprise agentic space [1]. Harvey characterized the competitive moment as 'Anthropic's game to lose,' reflecting a view that OpenAI's UX and naming confusion create an opening for rivals [1].