Summary
Theo of t3.gg shares his hands-on impressions of Gemini 4 Carbon, Google’s newest flagship model, after getting early access. He starts out unimpressed, noting that Carbon’s headline lead on Artificial Analysis (60 versus 59 for Opus 5.5) was quickly overtaken by the release of Fable, but says deeper testing turned up surprises.
The video walks through Google’s launch claims, including the model’s strong showing on Cursor Bench 4 and its emphasis on cybersecurity and long-horizon agentic work. Theo compares that with his own code-analysis benchmarks, where Carbon made 79 claims in a Rust architecture review and 23 of them were false, an improvement over its predecessor but well behind other models. He also tests 3D generation, finding the first attempts underwhelming.
Other topics include usage limits on Google’s $200 subscription tier compared with Anthropic’s, frustrations with the Antigravity interface, and Carbon’s cybersecurity results through Google’s Fairwind program, such as SEC Bench Pro and Exploit Gym, plus Project Fortify’s bug finds in open source projects. Viewers get a candid, benchmark-driven look at where Gemini 4 Carbon fits among current frontier models and why it may still be worth using.
📺 Source: Theo – t3․gg · Published November 11, 2026
🏷️ Format: Review
