Sam Altman posted the number last week: 900 million weekly active users. Nearly every one of them uses ChatGPT the same way. Open a tab, type a question, copy the answer, paste it somewhere else.

The tab that never touches your work

I tracked my own ChatGPT usage for a week back in January, which I wrote about in The Problem with ChatGPT for Work. And the pattern was depressingly consistent. I’d be writing an email in Gmail, alt-tab to ChatGPT, describe the email I was looking at (because ChatGPT cannot see it), get a response, copy it, tab back, paste it, then edit out the parts that didn’t fit because the model was guessing at context I never gave it.

The same thing happened with spreadsheets, LinkedIn messages, Jira tickets. Every single task required me to play middleman between ChatGPT and whatever browser tab I was actually working in. I counted roughly 40 minutes a day lost to the context-switching alone, and that does not include the cognitive tax of translating visual information into text prompts that ChatGPT could understand. You’re essentially doing the AI’s seeing for it. A colleague called it “being ChatGPT’s seeing-eye human” and I haven’t managed to get that phrase out of my head since.

Because the weird part is nobody treats this as a problem. Reddit threads about ChatGPT productivity tips are full of people sharing elaborate prompt templates to describe their screen contents more efficiently. We’ve normalized the workaround instead of questioning why the tool cannot just look at what we’re looking at.

And that’s how nearly a billion people use the most powerful AI tool ever built. Copy, paste, repeat.

900 million people, one workflow

OpenAI has done something remarkable with distribution. ChatGPT is on phones, desktops, in Slack, inside Apple’s ecosystem. But the core interaction model has not changed since GPT-3.5 launched in late 2022. You talk to it in a box. It talks back. And you manually shuttle information between that box and wherever your actual work lives.

The subscription itself is solid. GPT-4o is fast, capable, great for brainstorming or writing drafts. So I am not arguing against paying for ChatGPT Plus. But the 900 million number, impressive as it is, masks a pretty fundamental gap between what the tool can do and how people actually deploy it during their workday, which mostly involves browsers, tabs, and web apps that ChatGPT cannot see or interact with on its own.

You already pay for it

And this is the part that bugs me most. People pay $20/month for ChatGPT Plus or $200 for Pro, and then they use maybe 30% of what it could do because the interface keeps the AI separated from their work. It’s like buying a really nice drill and only using it as a paperweight because nobody mentioned drill bits.

dassi recently shipped something that addresses this directly: you can log in with your existing ChatGPT subscription. No API key, no separate billing, no configuration screens that make non-technical people want to close their laptop. Just your ChatGPT login, and suddenly the model you already pay for can see your browser, read the page you’re on, and actually do things with that context.

What changes when AI can see your screen

So instead of describing an email to ChatGPT in a separate tab, the AI reads it alongside you. Instead of copying a table from a spreadsheet into a prompt, it looks at the spreadsheet. And instead of pasting a cleaned-up response back into a form field, it drafts directly where you’re working.

The difference sounds incremental but it compounds across a workday. I wrote about this in our dassi vs ChatGPT comparison — the real distinction isn’t intelligence, it’s presence. ChatGPT is smart but blind to your context. A browser agent uses that same intelligence while actually seeing what you see.

The missing piece

OpenAI built the engine. But they did not build the steering wheel for browser-based work, which is where most knowledge work happens in 2026. That’s not a dig at OpenAI so much as an observation about what happens when a product scales to 900 million users before the interface catches up with the use case.

So dassi slots into that gap. Bring your ChatGPT subscription, skip the API key dance, and use the model inside the browser where your work already lives. For the millions of people who alt-tab to ChatGPT dozens of times a day, this might be the upgrade that actually changes how their afternoon feels.

And if 900 million people are already paying for the AI, maybe the next step isn’t a better model. Maybe it’s just putting the damn thing where the work is.