OpenAI launched ChatGPT Work on July 9, 2026 — an AI agent built into a redesigned ChatGPT desktop app that accepts a multi-step goal, gathers context from connected apps and workflows, breaks the job into smaller steps, and completes them independently, staying with complex projects for hours. The launch simultaneously absorbed OpenAI’s Codex coding agent into the same app, creating a unified product with three sections: Chat, Work, and Codex.
What ChatGPT Work Does
ChatGPT Work executes multi-step work goals end to end. A user submits an outcome — for example, “analyze last quarter’s sales data and build an executive summary deck” — and the agent queries connected apps, gathers the relevant information, structures it, and delivers the finished output. Supported output types include documents, spreadsheets, slides, and web apps.
According to OpenAI’s product description, ChatGPT Work “gathers information across your connected apps and workflows, breaks the job into smaller steps, and completes them independently — staying with complex projects for hours.” The agent handles intermediate steps without requiring step-by-step user prompts between each action.
ChatGPT Work runs on GPT-5.6 Sol, OpenAI’s current frontier model. GPT-5.6 Sol supports an ultra mode that spawns sub-agents to parallelize complex tasks and applies maximum reasoning effort to longer-horizon work — the capability that enables multi-hour autonomous execution.
Auto-Review: The Safety Gate for High-Stakes Actions
Auto-Review is ChatGPT Work’s built-in safety mechanism. It pauses execution before any high-stakes action — deleting files, sending emails, or making purchases — and requires explicit human confirmation before proceeding. OpenAI designed Auto-Review to address the core risk of agentic AI systems: irreversible actions executed without user oversight.
The feature reflects a deliberate trade-off: ChatGPT Work is designed to run with minimal interruptions for routine steps but inserts a human checkpoint before any action that cannot easily be undone.
One App, Three Sections: Chat, Work, and Codex
OpenAI merged Codex — previously a standalone coding agent — into the new ChatGPT desktop app for macOS and Windows. The unified app presents three sections: Chat for conversational use, Work for multi-step agentic execution, and Codex for code generation and execution. Codex remains a distinct section within the app; OpenAI has not retired it as a separate product category.
The redesigned desktop app is confirmed for macOS and Windows. Linux support has not been announced. The rollout began July 9 on web and mobile, starting with Pro, Enterprise, and Edu users, with Plus and Business users receiving access over the following days.
Anthropic Launched Claude Cowork on Mobile Two Days Earlier
Anthropic launched Claude Cowork on iOS and Android on July 7, 2026 — two days before ChatGPT Work’s debut. Anthropic also extended doubled Cowork usage limits through August 5 to mark the mobile rollout. Both products target knowledge workers who want to delegate multi-step work to AI rather than submit one-off queries.
According to coverage by BuildFastWithAI, OpenAI described ChatGPT Work as “following Claude Cowork’s approach to orchestrating tools across workplace software.” The framing positions this as a direct competitive response rather than an independent product direction.
Google has no equivalent agentic work-execution product. Gemini 3.5 Pro, which had been expected around July 17, had not launched as of ChatGPT Work’s release, leaving OpenAI and Anthropic as the only two major AI labs with live autonomous work agents in market.
Our Take
OpenAI just turned ChatGPT from a chat tool into an autonomous co-worker. The race to own the AI “employee” category is the defining enterprise software battle of 2026 — and Google’s absence while Gemini delays is costing it the enterprise segment. For business decision-makers choosing an AI productivity platform now, the choice is between ChatGPT Work on GPT-5.6 Sol and Claude Cowork on Fable 5 and Opus 4.8 — two products, two model families, two different security postures. Both are worth evaluating before committing to either.
For Context
- OpenAI GPT-5.6 Sol Launch: Terra, Luna, and Cerebras Speed — the model powering ChatGPT Work, with Terra and Luna variants for different task depths.
- OpenAI GPT-Live Makes ChatGPT Voice Sound Like a Real Conversation — OpenAI’s previous ChatGPT capability expansion in July 2026, adding full-duplex voice to the same desktop product family.
- Claude Cowork Mobile Launch: iOS and Android Now Live — Anthropic’s agentic work product that launched two days before ChatGPT Work.
