Summary
Sam Witteveen shares his hands-on impressions of Gemini 4 Carbon, Google DeepMind’s new frontier model, after two weeks of early access. Carbon launched at the top of the Artificial Analysis Intelligence Index before Anthropic’s Fable 5.2 reclaimed the lead, but the video focuses on what the model means for developers rather than leaderboard position.
Google positions Carbon, which is the same size as the faster Gemini 4 Argon, around four areas: multi-day agentic engineering, high-precision knowledge work in fields like finance, legal and healthcare, frontier scientific discovery, and full-stack front-end and 3D web development. Google also says a Flash-sized model and a Gemma 4 release are on the way.
Reported benchmarks include 65.2% on Terminal-Bench 4.0 versus roughly 59 for Opus 5.5 and GPT6 Astra, and about 54 on Automation Bench for knowledge work. Witteveen notes that most were published before Fable 5.2 arrived. Introductory API pricing is $2 per million input tokens and $10 per million output tokens through November 30, with 95% off cached input, and sampling settings like temperature are now preset.
The demos include a detailed interactive 3D model of Marina Bay Sands in Singapore, built mostly in one shot, with the model checking and refining its own work along the way.
📺 Source: Sam Witteveen · Published November 10, 2026
🏷️ Format: Review
