Asteria: This is basically OpenAI saying chat was the appetizer and the actual meal is, can the thing finish the deck.
Draco: Yeah. And more specifically, can it finish the deck by pulling from the mess your team already made elsewhere instead of making you restate everything.
Asteria: Which, honestly, is the real claim here. Not “look, another agent.” It’s that the bottleneck at work is scattered context across Slack, files, desktop apps, business tools, and half-finished drafts, and ChatGPT Work is supposed to gather that, plan the approach, then turn it into polished spreadsheets, docs, slides, even a little web page in Sites.
Draco: Right.
Asteria: That is a much bigger product argument than a feature page usually admits. It’s trying to become the work surface where the artifact gets assembled, not just the place you ask for help.
Draco: The page keeps repeating the same pattern through examples: account plans from a data room plus Slack, reconciliation across Oracle, Salesforce, and Databricks, recurring business reviews and campaign tasks, and Plan mode asking whether the output should be a Google Doc, a slide deck, or Markdown. That last one matters because format choice is part of the system, not just text generation.
Asteria: Mm-hm.
Asteria: That part I actually like. The family story made sense when we talked about GPT-5.6, and here it clicks even harder because “work” is not one workload. Some jobs want the fancy reasoning ceiling. Some just need the spreadsheet cleaned up before lunch. This is your favorite thing, Draco, boring model routing wearing a nice shirt.
Draco: It is, annoyingly, my favorite thing.
Asteria: Also, my week has had weirdly high document entropy, so this page hit me right in the product nerve. Too many tabs, too many half-true drafts, too many “can you turn this into something executive-ready” moments.
Draco: That is such an Exploring Next disease. We keep rediscovering that the expensive part is not typing words, it’s exposing the shape of the work clearly enough that a system can carry it.
Asteria: Yeah, the human-becomes-the-interface thing again.
Draco: Where I do push back a little is the evidence. The page is persuasive as product theater, but technically it’s thin. It gives polished scenario screenshots and broad claims, but it does not show success rates, failure cases, review burden, or how often these cross-tool workflows quietly go wrong in the middle.
Asteria: Sure.
Draco: And with work products, the hard part usually isn’t the last mile formatting. It’s state fidelity. Did you pull the right Salesforce account? Did Slack context include stale decisions? Did the desktop app action mutate the wrong sheet? None of that is visible here. So I buy the direction more than I buy the implied generality.
Asteria: I think that’s fair, but I don’t think it kills the argument. The practical buyer for this is not asking whether OpenAI published an error bar for deck assembly. They’re asking whether this removes a miserable workflow where somebody manually stitches notes, CRM context, spreadsheets, and old templates into one artifact every single week.
Draco: Yeah.
Asteria: If it does even a decent first pass, that is real. Especially for ops, finance, sales, program people, the teams who live in status docs and business reviews. This is not really about replacing judgment. It’s about collapsing the janitor work between scattered systems and a presentable output. That’s why the “stay in control” line is load-bearing.
Draco: And the rollout detail matters. The page says desktop is available today, and that web and mobile are rolling out over the next few days to Plus, Pro, Business, Enterprise, and Edu. So this is basically day-one-ish availability, not some vague coming-soon promise.
Asteria: Right, which changes who should care. If you’re already paying and your team uses ChatGPT seriously, this is not abstract road map vapor. You can at least start testing whether Work actually tracks your templates and connected tools well enough to matter.
Draco: There’s a funny little subtext too. They do not want the model story living as benchmark discourse alone. They want it attached to an application layer fast.
Asteria: Which is probably smart. Otherwise you end up in the same trap we complained about with harness bundling, where the model headline absorbs all the credit. Here they’re at least being blunt that the value is context plus tools plus output shaping plus action. It’s the stack, not just the weights.
Draco: Wait— and that means the argument mostly holds up, just at a different level than the page pretends. As a proof of universal agent competence, no. As a claim that work software is turning into context assembly plus artifact generation with a human approval loop … yeah, I think that’s basically correct.
Asteria: Exactly. And the Sites screenshot made me laugh a little because every company now wants the model to make the landing page, the slide deck, the spreadsheet, and maybe your personality if there’s time.
Draco: If it starts offering to reconcile my tabs emotionally, then I’m in.
Asteria: No, that’s Sol Ultra Therapy and it’s one hundred dollars a million tokens.
Draco: Terrible. Keep it.
Asteria: Honestly, Asteria-level optimism here, I’d put maybe seventy-thirty on this finding a real home in teams that already have repeatable document-heavy workflows. Not magic. Just enough friction removal that it sticks.
Draco: I’m a little lower, maybe sixty-forty, only because integration reliability is where these things go to die. But if the connected-tool layer is solid, you’re right, this is the kind of boring win that actually ships.
Asteria: Okay. Go reconcile your tabs emotionally, Draco. I’m calling this episode six-thirty-four and getting out before Work assigns us homework.