Theo times Gemini 4 Carbon: 11 seconds before it replies
Theo put a stopwatch on Gemini 4 Carbon and got 22m21s on low where Opus 5.5 high took 9m33s, and 11 seconds before the first token. He says almost all of it is server side.
Read the whole week in one goFable 5.2 tops the arena and still burns your limit
Receipts
Theo times Gemini 4 Carbon: 11 seconds to first token
Theo followed his early-access impressions with timings and says Gemini 4 Carbon is unacceptably slow in real world usage: 22m21s on low against 9m33s for Opus 5.5 on high, and fourteen hours to write a fix for T3 Code Windows installs.
Almost all of it is server side. Their inference is awful.
How they actually work
Pocock publishes the Home Repo pattern with a setup prompt
Matt Pocock wrote up the Home Repo pattern, his answer to running agents across multiple repos, and shipped a copy-pasteable prompt you can run on your org's repos to build one. The question it settles is where the skills, the steering files and the tooling live when no single repo owns them.
Model and agent watch
Fable 5.2 takes Agent Arena #1 at 1.6x the next model
Arena says Claude Fable 5.2 (Max) took #1 in Agent Arena with +30.4% net improvement across 4.7k real-world agentic sessions, about 1.6x the #2 Opus 5.5 at +19.5%, and #1 in Code, Chat and Work. Median cost per task fell slightly to $7.3 while API pricing dropped 20% to $8/$40 per million tokens.
Anthropic's current dominance compared to other labs is insane.
