ChatGPT Work Launched July 9 — Here's What It Actually Does (and What It Doesn't)

What ChatGPT Work Is

On July 9, 2026, OpenAI launched ChatGPT Work — an autonomous agent mode built on GPT-5.6 that does something earlier ChatGPT versions could not: it stays in the loop long enough to finish the job.

Where previous ChatGPT interactions were transactional (you prompt, it responds, you move on), Work is designed to run multi-step workflows over hours, gather context from connected apps like Slack, Google Drive, and email, and hand you back finished deliverables — documents, slide decks, spreadsheets, even Sites — without requiring you to write a single line of code.

The practical mechanics: Work connects to your apps through plugins, executes tasks in sequence using its built-in browser and Computer Use capabilities on desktop, and includes Scheduled Tasks so workflows can trigger automatically. It merges what was previously Codex — OpenAI's code-execution environment — into the same unified interface alongside standard Chat mode.

On access: if you have the desktop app on Mac or Windows, you have Work right now, regardless of whether you are on a Free plan. Web and mobile rollout began July 9 for Pro, Enterprise, and Edu accounts, with Plus and Business users phased in after that. The desktop availability for Free-tier users is the detail most coverage has buried — this is not a Pro-only feature, at least not on desktop.

The 95% Claim, Examined

The 95% figure comes from a specific place: a July 16, 2026 Entrepreneur article that walked through seven job categories using live prompts inside Work mode. The demos covered content dashboards, 90-day marketing campaigns, SEO blog production, and lead audits — the kind of preparatory, repeatable work that consumes most of a solo operator's week. The author's framing was that a one-person business could delegate all of this through single prompts, no hiring, no coding required.

That is a useful frame. It names concrete task types and gives you a starting mental model for what Work actually targets. The problem is the number itself. "95%" is a headline, not a benchmark. No independent study tested it across business types, industries, or operator skill levels. It reflects what one demonstration produced under controlled conditions, written to illustrate the tool's upside.

Real-world adoption data from OpenAI tells a more grounded version of the story. Over 5 million users were running Codex workflows weekly before it merged into Work. Companies including Zapier, RingCentral, Virgin Atlantic, and NVIDIA have used the underlying capabilities to compress work that previously took weeks into hours. That pattern — multi-step automation compressing time-intensive but repeatable tasks — is where the legitimate value is.

So treat the 95% as directional shorthand for the task types Work targets well, not as a performance guarantee.

What It Still Cannot Do

they build a prompt, hand it to Work, and then wait for a finished strategy.

Work does not produce strategy. It executes against one. OpenAI is explicit that the tool is designed for repetitive and preparatory work — and that distinction matters more than most coverage suggests. Work requires user input and approval at key decision points. It is not running unsupervised toward conclusions you have not already reached. You are the one who decides the positioning. You are the one who knows whether the tone of a campaign is right for a specific client relationship. You are the one who can tell the difference between a lead worth chasing and a name on a spreadsheet.

Consider a few specific scenarios. Work can draft a 90-day content calendar — but it cannot tell you whether your audience is ready for a pivot in message direction. It can produce a lead audit from your CRM data — but it cannot read the history behind why a particular account went cold six months ago. It can generate a slide deck — but it cannot decide whether the deck positions your offer the way your best salespeople would frame it in a room.

Creative direction lives in the same category. Work can generate variations. Deciding which one sounds like you — or like your client — is still a human call. That is not a limitation the next model update is likely to dissolve.

Where to Start

Download the desktop app first. It is free, available on Mac and Windows, and you do not need a paid plan to access Work mode. That removes the most common excuse for putting this off.

Once you have it open, the useful exercise is not to explore every feature — it is to pick one workflow you already run on a recurring basis. Something weekly, something multi-step, something you find yourself doing by hand that produces a predictable output. A competitive recap. A weekly content brief. A client status summary pulled from notes and email threads. It does not need to be complicated. It needs to be something you know well enough to evaluate the output honestly.

Run it end-to-end in Work mode and treat the result as a calibration, not a verdict. What did it get right on the first pass? Where did it need correction? Where did it ask for a decision that only you could make? Those gaps are the actual data. They tell you where Work saves time and where it still depends on your judgment — which, based on the previous section, is going to be more places than the promotional framing suggests.

That one workflow test gives you a concrete reference point before you build any assumptions about what Work can or cannot handle for your specific operation.