Observed failures

  • During early Skatebench testing, API errors were frequent enough that Theo changed his benchmark infrastructure. He leaves post-launch reliability unresolved. 8:33
  • In another run, a gateway connection fails after one hour and two minutes of file reading. Theo says the thread refuses to continue and he must restart without its history. 34:31
  • Antigravity's ACP exposes effort levels as separate model slugs, such as Argan High, Medium, and Low, rather than separate model and effort choices. Theo says this complicates integration and that Antigravity accounts for a disproportionate share of T3 code bugs. 35:06
  • He says T3 code embeds Google's official server-side ACP bindings and rejects claims that the app itself explains Gemini's slowness. This is his account of the integration, not an independently isolated performance test. 30:22

Account coupling

  • Theo worries that a Google-account ban could affect his YouTube channel, Gmail, Chrome passwords, and other Google sign-ins, making the potential consequences larger than losing access to a standalone model account. 33:45 34:10
  • He acknowledges Google's promise to limit bans to Gemini/Antigravity, but says he has still seen reports of whole-account bans. The video establishes his concern, not the current enforcement policy or frequency of such bans. 34:10 34:31

What the tooling changed for him

  • T3 code finally gives him a Gemini interface he finds reasonably usable. Seeing Gemini beside competing models still leaves him less satisfied because the interface does not resolve his concerns about output quality. 35:44