ChatGPT Work: OpenAI Turns the Chatbot Into a Multi-Hour Work Agent

From Chat Window to Work Layer

On July 10, OpenAI announced ChatGPT Work, a new autonomous agent that uses Codex and GPT-5.6 to pursue tasks across applications and documents for hours, turning goals into completed work. The move is not merely a feature upgrade; it is a repositioning of ChatGPT from a conversational interface to an execution layer. In the same wave, the company is folding the Codex coding assistant into the ChatGPT brand as ChatGPT Codex, erasing the old distinction between chat and code and signaling that all agentic capabilities will live under one product roof.

OpenAI chief Sam Altman also promoted GPT-5.6 Sol as one of the best models the company has shipped, framing the launch as a model-plus-narrative product event. The practical question is less whether the model is the best on benchmarks, and more how it connects to ChatGPT Work and Codex as a continuous product line.

What Multi-Hour Autonomy Changes

Until now, most AI interaction has been a ping-pong conversation: the user asks, the model answers, the user asks again. ChatGPT Work breaks that rhythm. It can stay with a project for extended periods, operating across apps and files without a prompt for every step. This converts the user from a continual questioner into an occasional validator, a shift Ethan Mollick has described as the twilight of the chatbots and Jack Clark has reported inside Anthropic as the move toward mass delegation of tasks to agents.

Long-running autonomy is also why the chat window was arguably a safety playpen. The dialog box forced a human approval at every turn. Once an agent is allowed to act across systems for hours, the interface changes from a stream of replies into a stream of task updates. That makes the work faster, but it also makes the harness around it essential.

Why OpenAI Is Pulling the Stack Together

ChatGPT Work arrives alongside the merger of ChatGPT and Codex under one brand. The simplification is deliberate: instead of asking users to choose a chatbot or a coding agent, OpenAI is putting all agentic functionality behind a single ChatGPT entry point. This follows the pattern of vertical integration Peter Thiel has identified as the source of real competitive moats: proprietary model, embedded tool, and the workflow itself all bound together.

Competitors are building similar vertical stacks. Meta is pitching Muse Spark 1.1 as a powerful yet cheap agent and coding model through its API and Meta AI. xAI is reportedly training Grok 4.5 with help from Cursor. OpenAI is even releasing a Codex plugin for Claude Code, allowing its model to enter a rival terminal workflow. The message is that the boundary between models and tools is dissolving, and the platform that owns the workflow may matter more than the model that powers it.

Output Is Cheap; Comprehension Is the Bottleneck

More autonomous execution does not remove the hard parts of software or knowledge work. It accelerates them. Frederick Brooks’s argument that adding people to a late project makes it later applies in a new form: AI can generate code far faster, but the essential complexity of systems remains, and the risk is that understanding falls behind output. The result is a maintenance backlog, or, in the words of one developer, a situation that can feel like maintaining a public restroom after rapid construction.

The response is harness engineering: strict sandboxes, version control, work trees, review gates, and rollback paths. Far from slowing AI down, these constraints let it run overnight across multiple windows, because the human can safely validate later. In this division of labor, AI becomes a supercharged System 1 — fast, generative, associative — while humans are forced into the System 2 role of verification and judgment. The problem is that System 2 is exhausting, and a user can only verify so many parallel streams at once.

Data, Sovereignty, and the New Workplace

Long-running agents that read documents and update apps also raise questions about who owns the data. The All-In podcast recently argued that companies should avoid handing proprietary data to frontier labs, precisely because those labs can absorb the data and then turn around and launch competing products. ChatGPT Work’s value proposition depends on access to files, calendars, codebases, and messages, which is exactly the kind of access that triggers sovereignty concerns.

Tools such as ZeroGPU, which claims to handle 70 to 80 percent of production inference with small, fast, edge-hosted models, show one way out: a tiered stack where routine tasks stay on premises or at the edge and only high-stakes reasoning goes to a frontier model. But whether sovereign AI is a genuine architecture or a marketing label depends on whether the vendor controls the model, the data, and the hardware.

The Bottom Line

ChatGPT Work is OpenAI’s clearest attempt to move beyond the chatbot and become a layer in which work actually happens. It is not simply a better model; it is a bet that users will want their agent to stay inside a project, use their tools, and deliver finished results. The technology is impressive, but the harder test will be trust: the safety harnesses, the data boundaries, and the willingness of humans to hand over tasks that last longer than a single conversation.