Summary
Fireship’s Code Report looks at Google’s newly released Gemini 4 Argon and asks whether it finally puts Google ahead of OpenAI and Anthropic. The video covers how the model was built, how it scores, and how it behaves in hands-on tests against Claude Opus 5.5 and GPT6 Astra.
According to the video, Argon is a sparse mixture-of-experts transformer trained from scratch on Google’s TPUs with a March 2026 knowledge cutoff, not a fine-tune of earlier Gemini models. It leads on the deep software engineering benchmark, the Val index and LVBench (91.7%), and tops the text leaderboard on Arena AI. It trails Opus 5.5 on Terminal Bench 4.0 (66 to 57), loses to GPT6 Astra on Frontier S, and ranks only ninth on Code Arena for web development.
Pricing is $2 per million input tokens and $10 per million output tokens until November 30. The rollout was messy: access in Antigravity went first to AI Ultra subscribers. Hands-on tests include generating a promo video, where Claude’s code-generated motion graphics beat Gemini’s native video output, and building a realistic 3D horse in Three.js, where Argon’s result had notable flaws.
📺 Source: Fireship · Published October 20, 2026
🏷️ Format: News Analysis
