So yesterday I fell down a rabbit hole watching someone put ChatGPT’s newest update through its paces, and I couldn’t stop taking notes. The short version? ChatGPT just shipped one of its biggest updates ever, and it’s called ChatGPT Work. The person who put it to the test is a seasoned AI creator who got early access, and they ran five real prompts to show exactly what it can do. I was genuinely surprised by how far this pushes ChatGPT past simple chatting.
Here’s the twist that grabbed me: this isn’t a chatbot that answers you. It’s an agent that goes off, thinks in the background, gathers context, and hands you a finished file. Reports, spreadsheets, presentations, web apps, even hosted websites with a shareable link. That’s the leap.
What’s actually new
- 🛠️ ChatGPT Work: an AI agent that completes tasks and can even access files on your computer through the desktop app.
- Three new models: Sol, Terra, and Luna, built specifically to power Work.
- A new desktop app for Mac and Windows that bundles chat, Work, and Codex together.
- Hosted sites: instead of a canvas preview, you now get a real URL you can share.
The models matter here, so let me pass along how the expert explained them. Sol is the most capable and most expensive, best for complex coding and deep reasoning. Terra is the balanced everyday model and the new default replacing 5.5. Luna is the fastest and cheapest, but the least capable. Free and Go accounts get Terra automatically. Paid plans unlock the model picker plus ChatGPT Work itself.
The twist most people will miss
Here’s the clever move the original poster shared. You’d think you’d just crank everything to Sol on max reasoning all the time. Don’t. Even inside the website, better models and higher reasoning effort burn credits fast, and you’ll hit a wall.
So the trick they used: before running a big prompt, they’d start a normal chat and simply ask ChatGPT which model it recommends for that specific task. Then they’d follow the suggestion. Sometimes it said Sol with max effort, sometimes Terra at medium. That one habit stops you from torching your credits on jobs that didn’t need the heavy model.
The five prompts in action
- 📊 Data cleanup into a dashboard: they uploaded files and asked Work to validate, clean, and audit the data, then publish a hosted dashboard with region filters. ChatGPT recommended Sol. It took about 15 minutes and spit out a real website at a shareable .site URL.
- ✈️ Trip planner: four files in, and the ask was a mobile-first hosted trip hub with a day-by-day itinerary and budget. ChatGPT suggested Terra at medium effort. The neat surprise: it offered three design directions to pick from before building. Both desktop and mobile versions looked clean.
- PowerPoint from a template: they fed it an old deck for style plus a CSV of data. It picked Sol at max reasoning and returned an 11-slide deck that matched the sample design almost exactly when downloaded as a real .pptx.
- 🚀 Full marketing plan: this was the showstopper. One prompt asked for a launch strategy, a 10-slide deck, a five-email sequence, 12 social posts, and three ad concepts. The agent produced a branded website, a 30-day launch doc, the presentation, and a full campaign kit, all in one go. It took 33 minutes on high reasoning effort.
- Daily command center: using plugins, it reviewed inbox, calendar, and Slack, then built a one-page PDF with priorities, urgent items, and decisions to make. Connect your real Gmail and Slack and it pulls live instead of from exported samples.
That last one struck me as the sleeper hit. A single page every morning telling you exactly what needs to happen. I can see myself actually using that.
Pro tips worth stealing
- Ask before you run. Let ChatGPT recommend the model for each prompt so you spend credits wisely instead of defaulting to max.
- Use the desktop app for file work. The website creates files and gives download links, but only the desktop app can read and write to actual folders on your machine.
- Feed it a style sample. For the deck, handing over an old presentation got the design matched almost perfectly. Same trick works for brand guidelines on the marketing plan.
- Lean on plugins for context. Connecting Gmail, Calendar, and Slack turns generic output into something built around your real day.
- Match effort to the job. Higher reasoning means longer waits. The marketing plan took 33 minutes, so save max effort for the big, multi-part asks.
One honest note the creator made: on the presentation, the output sometimes packs in almost too much text. So treat these as strong first drafts you refine, not finished-to-perfection work.
The expert also lined this up against Claude’s Co-work, which has been their go-to for a while, and teased a side-by-side comparison on output quality and credit usage coming next. I’m curious to see who wins.
If you use ChatGPT for business, content, planning, or just wrangling your daily chaos, this update is worth a real look. Watch the full video to see all five prompts play out step by step and catch the exact settings they used for each one. 🎯