OpenAI 新 Agent 产品 Dots 的实测体验

Chris @ChrisGPT原推发布于 2026.10.04

在 X 上打开原推

An OpenAI fanboy’s mildly wounded review of Dots I’ve reviewed a lot of products and have been fortunate enough to help shape some that people use every day, including Instagram, Muse, Ray-Ban glasses, and X. Most of that feedback has happened behind closed doors, directly with engineering teams. I wanted to start doing more of that testing publicly, so here’s my honest review of Dots, from someone who’s admittedly biased toward OpenAI. The good • A substantial night of work, five hours through the VM, used just 1% of my reported usage on the $200 plan. • The voice is genuinely flexible and, in my experience, seems slightly more advanced than Advanced Voice Mode. I could change its speaking style and get a softer, slower delivery, and even get it to sing. • It is quite good within its own VM and does a great job of balancing multiple tasks, particularly working with complex spreadsheets. The bad • Computer-use troubleshooting needed constant chasing. It confused a task running on my PC, a local CLI session, and a Local chat inside Codex. • I saw a success-looking checkmark (embedded in chat) in the Dots chat after entering a wrong password, while the site still showed an error. • It sometimes said things were done before verifying them, then contradicted itself. • It couldn’t end a voice call when I asked. (Slight nitpicking, but AVM has done this for me) • It kept using colons after I asked it not to. In my experience, GPT-5.2 is better at instruction following in chat. • Even after I explained the difference, it repeated the same cloud-chat approach. Getting it to open a Local thread for computer use has been nearly impossible in my testing, even with full permissions. There’s real potential here, but a user needs much better follow-through. When GPT-6 launched, my understanding of the blog was that GPT-6 would keep going if it ever got stopped and would infer what you wanted. In my testing, that persistence has been missing. GPT-6 often seems eager to stop. Other users and early testers I’ve spoken to have reported the same thing. It hasn’t been proactive enough about deciding when to continue, and unfortunately, this carries over to Dots. All that said, multiple OpenAI employees have reached out to help with the computer-use issue. I’ve since been able to install updates and loosely get it working. It still, however, is not consistently solved. It feels like a final harness for AGI, but with a subpar, broken orchestrator. I feel as if we were very overhyped for something that, from what I saw, didn’t even work properly on stage. This is my first pass with Dots, and I assume it will become more rapidly more reliant.