Complete ChatGPT Work guide: Plan Mode, Plugins, Computer Use, Sites, Scheduled Tasks and Templates. All features tested.
what it does, how to get access.
I disagree with the statement. I’m a brand manager and have used Claude before Chats upgrade. I run the analysis Claude does and Chat goes so much deeper and actually listens to my directions while Claude needs every tiny thought exactly defined or it makes assumptions.
Good to hear from someone running both on real work. Going deeper and following directions are the two things I'd test next. What kind of analysis are you running?
A bit of everything- catalog, sales, forecasting, inventory and more. I’ve recently paired the two with metas Muse and a Linux OS cloud connect and there’s much better power usage and accuracy overall. The core is still chat though. I hate tokens it’s like texting back in the day when you have to pay for data…
Something about the Claude Code harness still has a bunch of alpha in it. It makes connectors more reliable, tasks more in my voice, and better long-coordination over huge 5+h tasks. How do you compare the two?
Thanks for asking, Tara! The three gaps I feel most are the ones I named above: connector reliability, drafts that land in my voice, and coordination over 5+ hour tasks. Close those and I'd move more of my week over.
The cloud tasks from desktop to mobile are the part I use constantly. And ChatGPT is now where I'd point people instead of Codex.
This mirrors what I've seen building with clients: the model matters less than everyone thinks, the orchestration layer is where the real performance difference shows up. We switched a client's support automation between three different 'best' models last year and the output quality delta was under 8% each time, but harness/prompt engineering swings were 30%+. Worth remembering when the next 'genuinely better' model drops next month.
Same model, different harness, very different results. That's why I spent most of the post on how the ChatGPT Work harness is set up and not on the model.
"I don’t think ChatGPT work is better than Claude Cowork yet.."
Curious what you find better about Cowork!
I disagree with the statement. I’m a brand manager and have used Claude before Chats upgrade. I run the analysis Claude does and Chat goes so much deeper and actually listens to my directions while Claude needs every tiny thought exactly defined or it makes assumptions.
Good to hear from someone running both on real work. Going deeper and following directions are the two things I'd test next. What kind of analysis are you running?
A bit of everything- catalog, sales, forecasting, inventory and more. I’ve recently paired the two with metas Muse and a Linux OS cloud connect and there’s much better power usage and accuracy overall. The core is still chat though. I hate tokens it’s like texting back in the day when you have to pay for data…
Something about the Claude Code harness still has a bunch of alpha in it. It makes connectors more reliable, tasks more in my voice, and better long-coordination over huge 5+h tasks. How do you compare the two?
I lead the ChatGPT Work product, so probably a bit biased :) but I’d love your feedback for anything we can do better.
Thanks for asking, Tara! The three gaps I feel most are the ones I named above: connector reliability, drafts that land in my voice, and coordination over 5+ hour tasks. Close those and I'd move more of my week over.
The cloud tasks from desktop to mobile are the part I use constantly. And ChatGPT is now where I'd point people instead of Codex.
I’d love to see your work girl
This mirrors what I've seen building with clients: the model matters less than everyone thinks, the orchestration layer is where the real performance difference shows up. We switched a client's support automation between three different 'best' models last year and the output quality delta was under 8% each time, but harness/prompt engineering swings were 30%+. Worth remembering when the next 'genuinely better' model drops next month.
Same model, different harness, very different results. That's why I spent most of the post on how the ChatGPT Work harness is set up and not on the model.
I’ve love to look into this you know