OpenAI dropped GPT-5.3 Instant this week and my timeline went feral over the benchmarks. Fair enough. But I’ve been living with GPT-5.2 as my daily driver since January, and the thing that nags at me has nothing to do with how smart the model is or isn’t. Most people are running the most capable reasoning system we’ve ever built and feeding it information through a straw.

chatgpt.com cannot see your screen

Last Tuesday I was deep in a quarterly report in Google Sheets, the kind where every tab spawns two more, and I needed GPT-5.2 to spot revenue trends. So I tabbed over to chatgpt.com, copied a column of numbers, pasted them in, typed “what trends do you see,” and immediately realized I’d forgotten the damn column headers so the model was just guessing. Back to Sheets, copy the headers, paste those too. Four minutes to set up what should have been a ten-second question.

And this ritual is everywhere. There’s a Reddit thread from last week where people are trading elaborate copy-paste workflows and prompt tricks for describing their screen contents to the model, as though the bottleneck is that nobody’s found the right incantation for manually transcribing a browser tab into a chat window. But the real problem is simpler and dumber than that: chatgpt.com cannot see what you’re looking at. We built arguably the best reasoning system on the planet and then asked users to be its eyes.

Five tasks, zero clipboard

I ran a few things through dassi, which puts GPT-5.2 inside the browser with full page context. Summarizing articles, comparing prices across tabs, drafting an email reply in Gmail. Same model, completely different experience.

Where I changed my mind

I expected “saves a few clicks” territory. I wrote about the copy-paste tax back in January, where I tracked roughly an hour per day lost to the overhead of shuttling context between my actual work and chatgpt.com. Running GPT-5.2 inside the browser, that hour mostly disappeared — nice, but not exactly a revelation.

But last week was different. A customer complaint came in and I asked GPT-5.2 to help draft a response. On chatgpt.com I would’ve pasted the most recent message, maybe the one before it, and gotten back something polished that completely missed the point. But the browser agent read the full thread on its own, all seven messages going back two months, and surfaced a shipping date promise from early January that I had completely forgotten we made. And that one forgotten detail reshaped the entire reply. So instead of the usual “we apologize for the inconvenience” boilerplate, it wrote something that addressed what the customer was actually upset about, which was that we’d broken a specific promise rather than just delivered late.

I keep thinking about the gap between those two drafts. The chatgpt.com version would have been grammatically perfect and emotionally tone-deaf, because it only knew what I remembered to paste in, which was the most recent angry message with none of the backstory. And I do not think I’m unusual in this. Most people’s frustration with ChatGPT probably comes more from context starvation than from model capability, but because context starvation looks the same as stupidity from the output side, nobody frames it that way.

Also, vendor lock-in is annoying

Dassi runs bring-your-own-key. So it’s GPT-5.2 for quick tasks, Claude Opus 4 when I want more careful reasoning, Gemini 3.1 Pro for that million-token context window. Or skip the API key entirely and log in with your existing ChatGPT subscription.

I have been bouncing between GPT-5.2 and Claude for different kinds of work and honestly stopped thinking about model selection as something that matters. Or maybe I just haven’t hit the task yet where it makes a difference.

Go try it

Install dassi. Open whatever tab you’re actually working in. Ask GPT-5.2 about it. You do not have to keep copying your work into a separate chat window like it’s 2024.