Summary
Matthew Berman argues that Google’s Gemini 4 Neon, released two days earlier, changes the narrative of a race long dominated by OpenAI and Anthropic. He calls it the best model available right now, particularly for real knowledge work and coding rather than flashy demos.
The video starts with benchmarks comparing Neon to Claude Opus 5.5 and GPT 6.2 Soul. Neon scores 61 on the Artificial Analysis Intelligence Index, reaches 65% on Cursor Bench versus 58% for Opus, and leads on Automation Bench, Agents Last Exam and Terminal Bench Science. It trails Opus slightly on Terminal Bench and Frontier Code. Berman then moves past the numbers, running a Rubik’s Cube simulation that Neon completed in about 20 minutes with working scramble and solve.
The core of the video shows Neon working on Scarlet, his open-source AI video editing app. Given deliberately underspecified prompts, such as tidying a model picker from a screenshot, adding Gemini model support through the Antigravity CLI over the agent client protocol, and adding toolbar hover explanations, it infers intent correctly and gives short, clear summaries. Berman also tests its computer use by having it inspect an OBS screen recording for quality issues.
📺 Source: Matthew Berman · Published November 20, 2026
🏷️ Format: Review







