OpenAI launches ChatGPT Work: the AI agent that works for hours without you
🔎 The chatbot is dead, long live the executing agent
On July 9, 2026, OpenAI unveiled ChatGPT Work. Not a minor update, not a new, barely perceptible model. A paradigm shift: ChatGPT moves from conversational tool to autonomous work agent, capable of executing complex tasks for hours without human intervention.
The difference is fundamental. Until now, AI generated text that you copy-pasted. Now, it opens your files, navigates your applications, builds spreadsheets, writes documents, assembles presentations. It acts, it no longer just responds.
This is the logical culmination of a trajectory that researchers as early as 2023 described as inevitable: the transition from AIGC (Artificial Intelligence Generated Content) to autonomous agentic execution, as documented in the reference study One Small Step for Generative AI, One Giant Leap for AGI. What was theory three years ago becomes a mainstream product.
The key points
- ChatGPT Work is an autonomous agent launched on July 9, 2026, powered by GPT-5.6 and the Codex engine.
- It can take control of your applications, files, and workflows to produce documents, spreadsheets, presentations, and web apps — for several hours without supervision.
- OpenAI is directly targeting the professional productivity market, where Microsoft Copilot and Google Workspace AI had established themselves.
- The security risks and loss of control are real and not fully resolved at this stage.
- The competition (Anthropic Computer Use, Google) is already reacting.
Recommended tools
| Tool | Main usage | Price (July 2026, check on openai.com) | Ideal for |
|---|---|---|---|
| ChatGPT Work | Autonomous work agent | Included in Pro/Team plans (then Enterprise) | Professionals delegating complex tasks |
| Claude Opus 4.7 (Adaptive) | In-depth reasoning and analysis | Via Anthropic API (check on anthropic.com) | Reasoning tasks requiring nuance |
| Gemini 3 Pro Deep Think | Multimodal analysis and long reasoning | Via Google AI Studio (check on ai.google.dev) | Integrated Google Workspace workflows |
| GPT-5.3 Codex | Code generation and execution | Via OpenAI API (check on openai.com) | Autonomous software development |
What ChatGPT Work actually is
ChatGPT Work is an agent that receives an objective, not a conversational prompt. You don't tell it "write me a paragraph about". You tell it "prepare the Q3 quarterly report from the sales data in this folder, create the charts, format everything into a Keynote presentation".
The agent plans the steps, accesses the necessary files, executes the manipulations, iterates on the results, and delivers a finished product to you. According to Bloomberg, the agent can work continuously for several hours on a single task.
This is made possible by the combination of two engines: GPT-5.6 for understanding, reasoning, and planning, and Codex for technical execution — file manipulation, interface automation, code generation. To understand the power of Codex in this context, see our analysis of OpenAI Codex : record & replay, montrez une tâche une fois, l'agent la répète à l'infini.
The detail that changes everything: ChatGPT Work does not need pre-configured integrations. It interacts with your applications in the same way a human would — via the graphical interface, the file system, APIs when they are available.
GPT-5.6: the engine that makes autonomy possible
An autonomous agent without a robust reasoning model is a disaster waiting to happen. ChatGPT Work relies on GPT-5.6, the latest addition to the OpenAI family presented on the same day. Our article OpenAI GPT-5.6 : Sol, Terra et Luna, la famille de modèles qui change tout details the architecture, but here is what matters for Work.
GPT-5.6 brings three critical capabilities for autonomous execution: reliable multi-step planning, error detection along the way, and the ability to resume a workflow after an interruption. This last point is essential — an agent that loses all its context after a crash is not usable in production.
OpenAI documents these reasoning capabilities in the System Card d'o1, which lays the foundations for what GPT-5.6 pushes to maturity: the internal chain of thought as a real-time verification and correction mechanism.
The concrete result: the agent no longer gets stuck after the third step of a ten-step workflow. It adapts its plan when it encounters an obstacle — a missing file, an unexpected format, an application that isn't responding.
What the agent can actually produce
The use cases validated at launch fall into four categories.
Structured documents: reports, briefs, market analyses. The agent compiles sources, cross-references data, structures the document, and inserts references. Not a draft — a document that you proofread, not rewrite.
Spreadsheets and data analysis: from CSV files, CRM exports, or accessible databases, ChatGPT Work builds tables, applies formulas, creates pivot tables, and generates charts. This is the use case that impresses the most according to the early feedback cited by PYMNTS.
Presentations: the complete workflow — narrative structure, slide content, layout, export in the target format (PowerPoint, Keynote, Google Slides).
Functional web apps: by combining GPT-5.6 and Codex, the agent can generate simple web applications — internal dashboards, forms, calculation tools — and deploy them. This is where the GPT-5.3 Codex model comes into play for the pure execution part.
Android Authority points out that the agent also handles research and information compilation tasks, a domain where reasoning models like Claude Opus 4.7 (Adaptive) or Gemini 3 Pro Deep Think also excel, but without the execution layer.
How it works in practice: the typical workflow
The process takes place in three phases, which the user experiences differently depending on their level of involvement.
Briefing phase: you describe the objective, provide the sources (files, links, folder access), and specify the constraints (format, length, audience). This phase is like a brief you would give to a junior colleague. It takes 2 to 5 minutes.
Autonomous execution phase: the agent works alone. You can close the tab, turn off your computer. ChatGPT Work continues on OpenAI's servers. A notification alerts you when it's finished or if the agent is stuck on a point that requires your decision.
Review and iteration phase: you receive the deliverable. You request adjustments just as you would with a human — "move the Q2 figures up front", "change the tone of the intro", "add a slide on the competition". The agent iterates.
The fundamental difference with classic ChatGPT: you never see the agent work step by step, you never copy-paste an intermediate result. You gave an objective, you received a deliverable.
Security: the acknowledged weak point
Letting an AI agent take control of your files and applications for hours raises security questions that OpenAI cannot dodge. And in fact, the company does not dodge them — it owns up to them with relative transparency.
According to the o1 System Card, OpenAI's reasoning models integrate safeguards against unwanted behaviors, but the document explicitly acknowledges that the ability to act in the world (via tools, via interfaces) mechanically increases the attack surface.
The identified risks fall into three categories.
Data exfiltration: an agent that accesses your files to prepare a report can theoretically expose sensitive data in its requests, its logs, or in its deliverables if they are poorly configured. OpenAI claims that Pro and Enterprise user data is not used for training, but the principle of least privilege is difficult to guarantee when the agent precisely needs to access a lot of things.
Unwanted actions: sending an email to the wrong recipient, deleting a file thinking it is being moved, publishing a draft prematurely. The agent acts with an imperfect understanding of the professional context — who is the boss, which document is confidential, which decision requires human validation.
Systemic dependency: when a team starts delegating its critical workflows to an agent, the ability to take back control diminishes. If the service goes down, if the model changes its behavior, the entire production chain is impacted.
OpenAI offers execution sandboxes and scope limits, but Bloomberg reports that several companies in the beta program have flagged unexpected behaviors requiring manual intervention.
The competition: where do Anthropic and Google stand
OpenAI didn't invent the concept of the autonomous agent. But it is the first to deliver it as a mass-market product at this scale. The comparison is inevitable.
Anthropic and Computer Use: Anthropic paved the way with Computer Use on Claude, which allows the model to interact directly with graphical interfaces. Claude Opus 4.7 (Adaptive), with its agentic score of 94.3, remains a superior reasoning model in certain complex scenarios. But Anthropic hasn't yet packaged Computer Use into a standalone "press a button and let it do its thing" product. It is still a tool for developers. Our comparison Claude vs ChatGPT detailed this difference in approach — Anthropic prioritizes safety and control, OpenAI prioritizes speed to market.
Google and Gemini: Gemini 3 Pro Deep Think (95.4 on the agentic leaderboard) has comparable reasoning capabilities. And Google owns the Workspace ecosystem — Docs, Sheets, Slides — that ChatGPT Work must conquer from the outside. The question is whether Google will integrate agentic capabilities directly into Workspace or stick to the Copilot model (inline suggestions, not autonomous execution). Our analysis ChatGPT vs Gemini will be updated on this point.
Open source agents: for those who want autonomy without sending their data to OpenAI, the open source ecosystem is making progress. Solutions like AI agents with local Ollama or the best LLMs for AI agents offer alternatives, but with higher implementation complexity and lower performance. Moonshot AI's Kimi K2.6 (88.1) and Z.AI's GLM-5 (82) show that the gap is narrowing, but it still exists.
| Approach | Flagship model | Agentic score | Main advantage | Disadvantage |
|---|---|---|---|---|
| OpenAI ChatGPT Work | GPT-5.6 | Unranked (new) | Finished product, zero config | Data with OpenAI, security risks |
| Anthropic Computer Use | Claude Opus 4.7 | 94.3 | Superior reasoning, safety | No integrated standalone product |
| Google Workspace AI | Gemini 3 Pro Deep Think | 95.4 | Native ecosystem | No autonomous execution to date |
| Open source (Ollama) | Kimi K2.6 / GLM-5 | 88.1 / 82 | Data sovereignty | Complexity, lower performance |
What this changes for daily work
The concrete impact depends on your role, but patterns are already emerging.
For analysts and consultants : the most time-consuming part — compiling data, formatting it, building deliverables — can be delegated. The analyst shifts from producer to editor. They validate, refine, contextualize. Their time shifts from 80% production / 20% reflection to 20% production / 80% reflection. This is a change in profession, not just a productivity gain.
For managers : ChatGPT Work can prepare meeting materials, synthesize team reports, compile metrics. The manager spends less time chasing information and more time deciding. But they also delegate part of their judgment — a summary made by the agent is not a summary made by a human who attended the meeting.
For creatives : the agent handles technical production — layout, formatting, export — but not the creative conception. The copywriter still writes the core ideas, the agent inserts them into the template, manages versions, and exports for different platforms.
The overall trend: work shifts upstream (problem definition, briefing, validation) and downstream (judgment, iteration, decision). The middle (execution) is absorbed by the agent.
The control paradox: more autonomy, more surveillance
Here is the problem that few commentators raise: a more autonomous agent actually requires more supervision, not less. But supervision of a different nature.
With classic ChatGPT, you monitor in real time — every response is visible, every error is immediate. With ChatGPT Work, the agent works for hours in an opaque manner. Your supervision becomes a posteriori: you check the final deliverable.
This requires new skills: knowing how to brief an agent (being precise without being restrictive), knowing how to audit a deliverable produced by AI (verifying sources, calculations, underlying reasoning, not just the form). Skills that the study How Close is ChatGPT to Human Experts? already identified as critical: the ability to evaluate the quality of an AI output is itself a skill that is not uniformly distributed.
The risk is a scenario where professionals delegate, receive a clean deliverable, accept it because it looks good, and discover subtle errors too late. This is not a theoretical risk — it is already documented in legal and medical contexts where LLM outputs are used as drafts.
The business model: why OpenAI is doing this now
The reason isn't just technological. It's economic. OpenAI needs to monetize beyond the $200/month ChatGPT Pro subscriptions.
ChatGPT Work is an enterprise product. Pro plans have access to a limited version of it, but the real market is businesses that will pay per seat, per task, or per execution volume. This is the classic SaaS model applied to agentic AI.
OpenAI is also sending a signal to Microsoft: ChatGPT Work partially bypasses Copilot. The agent operates in the browser, not within the Microsoft ecosystem. If a user can have ChatGPT Work prepare their Q3 report instead of using Copilot in Word, the value of the Microsoft integration diminishes.
It's a risky but calculated move. OpenAI is betting that autonomous execution is a strong enough competitive advantage for users to change their workflows, even if they have to leave their usual tools.
The link with Codex and the future of the universal agent
ChatGPT Work is not an isolated product. It is the visible face of a strategy that connects Codex, reasoning models, and multimodal capabilities.
The Lance model published on arXiv shows that the trend in research is moving toward multimodal unification — a single model that processes text, image, interface, code. ChatGPT Work is the product application of this trend: an agent that is not specialized in one type of task but that orchestrates multiple capabilities.
In the longer term, the logic points toward a universal agent: you give it a complex objective involving research, analysis, writing, code, design, and it orchestrates everything. Work on dexterous manipulation (Learning Dexterous In-Hand Manipulation) even shows that robotics research and software AI are converging toward the same paradigm — an agent that acts in the physical or digital world with the same principle of planning and execution.
ChatGPT Work is a step. Not the last.
❌ Common mistakes
Mistake 1: Briefing too vague
Giving a vague objective like "make a good presentation about our results" guarantees a mediocre deliverable. The agent cannot guess your mental context. Be specific about the sources, format, audience, and tone. Brief the agent as you would brief an external contractor who knows nothing about your company.
Mistake 2: Not auditing calculations and sources
The agent can produce tables with incorrect formulas or cite data that is not in the provided files (hallucination). Systematically verify the figures, formulas in spreadsheets, and the origin of key data.
Mistake 3: Letting the agent access everything without boundaries
The more permissions you open up, the higher the risk. Limit the agent's access to the strict minimum necessary for the task. A specific folder, not the entire drive. One application, not the whole system.
Mistake 4: Confusing autonomy with reliability
The fact that the agent works alone for hours does not mean the result is reliable. Autonomy applies to execution, not judgment. The last mile — validation, human context, accountability — always belongs to you.
❓ Frequently Asked Questions
Does ChatGPT Work replace Copilot or built-in AI assistants?
Not exactly. Copilot suggests in-context (in Word, in Excel). ChatGPT Work executes outside, autonomously. The two can coexist, but for different use cases — Copilot for inline help, Work for end-to-end tasks.
Are my data safe with ChatGPT Work?
OpenAI claims that Pro and Enterprise data are not used for training. But the agent accesses your files, which creates an attack surface. For sensitive data, local solutions like AI agents with Ollama remain safer, despite their performance limitations.
Can ChatGPT Work be used on mobile?
The launch is desktop-first. But the logic behind OpenAI's mobile strategy — like what we analyzed with Codex in ChatGPT Mobile: coding from your phone while the agent works on your machine — suggests that the Work agent will soon be controllable from mobile, with execution on desktop or server.
Which model is best suited for autonomous agents?
GPT-5.6 is optimized for agentic execution in the OpenAI ecosystem. But for pure reasoning scenarios, Claude Opus 4.7 (94.3) and Gemini 3 Pro Deep Think (95.4) remain competitive. The choice depends on your priority: integrated ecosystem (OpenAI) or raw reasoning quality (Anthropic/Google). Check out our guide to the best LLMs for AI agents for a detailed comparison.
Will ChatGPT Work monetize through advertising?
This is a legitimate question as OpenAI just launched ChatGPT Ads: targeted advertising for all US advertisers. For now, ChatGPT Work is a premium, ad-free product. But the tension between the subscription model and advertising revenue at OpenAI is real and warrants monitoring.
✅ Conclusion
ChatGPT Work marks the tipping point: AI shifts from a content tool to an execution tool. It is no longer an assistant that helps you work, it is an agent that works for you. The implications — on the nature of work, on data security, on OpenAI's business model — are massive. The challenge for professionals is no longer learning how to prompt, but learning to delegate to a machine that has neither intuition nor responsibility. Those who master this skill will gain a disproportionate advantage. The others will simply endure the deliverables.