As of: 2026-10-14 00:22 CEST. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/v/bridgemind-vibe-coding-with-gemini-4-argon-stream/ # Vibe Coding With Gemini 4 Argon BridgeMind, 13 October 2026, YouTube. 44,482 views as of 13 October 2026. Kind: Review. Watch: https://www.youtube.com/watch?v=Xy2FkHoR4Vc ## Why it is here 2 h 16 min launch-day live stream of coding with Gemini 4 Argon; with about 44k views it was the most-viewed independent Argon video on YouTube in the first hours after release. Its description quotes a "1M output token limit", the Sept 30 wording that Google changed at general release to a unified 1M input-output limit. ## Description (written by Gemini from the video) **Summary** In this livestream, BridgeMind creator James conducts a hands-on review and benchmarking session of Google DeepMind's newly released frontier model, Gemini 4 Argon. Throughout the session, he tests the model's coding, multi-agent orchestration, and multimodal capabilities across diverse tasks—including 3D browser games, Blender 3D modeling from video input, mobile app development, and design benchmarks—while comparing its performance against competitor models from Anthropic, OpenAI, and Meta. **What is shown** * **Prompting & Setup [01:13 - 04:10]:** James sets up Antigravity and launches parallel coding tasks for Gemini 4 Argon (at medium and high effort settings), including a 3D horror house game, a single-file HTML Minecraft clone, and a 3D Mario Kart clone orchestrated across 5 sub-agents. * **Pricing & Specs on OpenRouter [06:05]:** Navigating to OpenRouter to inspect Gemini 4 Argon pricing ($2.00 input / $10.00 output per 1M tokens, 50% discount at launch), a 1.0M context window, and native multimodality supporting text, image, audio, and video inputs, running at 78–92 tokens per second. * **Testing 3D Web Games [10:20, 21:30, 82:50, 91:15]:** James launches and plays several games generated by the model in the browser: "Bridge House" (a 3D flashlight horror game), "Super Kart GP" (a Mario Kart clone with functional kart physics and UI), a 3D browser Call of Duty Zombies clone ("Dead Signal"), and "BridgeZone 2" (a multi-player 3D battle royale with live lobby joining). * **BridgeBench Design Gallery [26:20 - 30:50]:** Inspecting UI and 3D shader benchmarks generated by Gemini 4 Argon, including the signature "Lava Lamp" benchmark ($0.18, 1m 58s generation time), a space rocket launch animation, and a sunset ocean render. * **Blender 3D Modeling via Video Scan [64:25 - 66:10, 112:00 - 114:25]:** James films a video walkthrough of his office setup with his phone, air-drops it to his workstation, and prompts Gemini 4 Argon via Blender MCP to reconstruct the entire office layout and desk equipment in Blender 3D. * **Mobile App in iOS Simulator [74:35 - 78:40]:** Launching and playing "Sift - Sand Puzzle", an iOS mobile puzzle game built in React Native/Expo using a specification prompt generated by Claude Opus 5.5 and implemented by Gemini 4 Argon. * **Remotion Code-Generated Promo Video [48:15, 125:00]:** Reviewing full marketing videos for BridgeMind rendered directly via Remotion code with automated sound effects, animation transitions, and music. **Claims & numbers** * The presenter notes Gemini 4 Argon is priced on OpenRouter at $2 per 1M input tokens and $10 per 1M output tokens (with a 50% launch discount), featuring a 1M token context window and 78–92 tokens/sec throughput [06:08]. * On CursorBench 4.0, Gemini 4 Argon High scores 52.9% at $4.57 per task and 109,500 tokens/task, trailing Claude Opus 5.5 (57.8%) and Claude Sonnet 5.5 Extra High (56.0%) [17:48]. * The presenter highlights that Gemini 4 Argon natively supports video as an input modality, unlike Claude Opus 5.5 which accepts text, image, and file inputs [05:08]. * The presenter claims his weekly Antigravity quota fell by 8% and the 5-hour allotment fell to ~81% within roughly 15 minutes of running four agents with Gemini 4 Argon on the $200/month Google AI Ultra subscription [13:16, 42:20]. * On BridgeBench's Lava Lamp benchmark, Gemini 4 Argon completed the task in 1m 58s for $0.18, compared to Claude Opus 5.5 at 9m 35s costing $1.27, and Gemini 3.8 Flash at 1m 03s costing $0.05 [29:25, 45:55]. **Notable quotes** * "It takes in video modality... which is actually a huge strength of Gemini models." [05:01] * "For a Gemini model, this is actually pretty good... it's a nice bump up from Gemini 3.8 Flash. But this is nothing that's going to change the frontier." [56:43] * "I don't think that it's on par with frontier models... I personally am not going to give it the BridgeMind stamp of approval." [124:05] **Assessment** This is an authentic, unedited live review and hands-on developer testing session. The presenter demonstrates real execution, deployment, and in-game playtesting across multiple coding harnesses, highlighting both genuine strengths (fast code generation, video-to-3D Blender workflow, 3D physics games) and weaknesses (inconsistent gameplay design, poor video rendering artifacts, and aggressive subscription token burn). _Described by gemini-3.8-flash on 2026-10-13 from the video's audio and frames._ ## Related - 2026-10-13: [Google releases Gemini 4 Argon broadly](https://postcutoff.com/e/2026-10-13-gemini-4-argon-general-availability/)