Notes from the videos, plus the things that didn't need one.
Google's last model built black blobs and scored a 1. Its new one, Gemini 4 Argon, built see-through, wobbling gummy bears and scored an 8. Here's how it did against Anthropic, OpenAI and xAI.
2026-10-17
Read: Is Google back? Gemini 4 Argon vs six other models, building squishy gummy bears
videoHaiku 5.5 is 90% cheaper, until it isn'tClaude Haiku 5.5 lists at 10 cents per million input tokens. In a coding agent it paid about four and a half times that, and Opus with Haiku helpers cost more than Opus alone.
videoI burned 1.4 billion tokens testing AI effort levelsSix models built Bloons Tower Defense at every effort level. High was the sweet spot, and Sonnet on max spent almost five hours and $163 to score lower than it did on high.