Gemini 4 Carbon Goes Fully Public: 1M Context, 65K Output, and a Multi-Agent Workflow That Reversed a 1994 Game

For developers leaning on AI coding agents, Carbon is the first mainstream model shown to handle hours-long autonomous tasks like binary reverse engineering and CUDA kernel authoring without a human stepping in. Access still gates the benefit: Antigravity users need a Google AI Ultra subscription, and API users only get price parity until November 30, 2026.

Google has put Gemini 4 Carbon on general release, and it is already plugged into Antigravity. The model keeps the 1M-token context of Argon but pushes output to 65,536 tokens and adds three thinking-depth levels (Low, Medium, High). On the latest CursorBench 4.0 leaderboard, Carbon's High setting scored 60.4%, taking first place. API pricing matches Argon until November 30, 2026, making the upgraded model a free swap during the promotion window.

The post walks through seven demonstrations. A /swarm command in Antigravity spins up a four-agent team (a Coordinator, two Builders, a Verifier) that built an interactive Three.js 3D supercar showroom in about ten minutes, with the Coordinator planning module interfaces in 26 seconds. Another case reverse-engineered the 1994 game Theme Park from its binary, rewrote the rendering pipeline in browser-friendly code, and added a new DeepMind-branded ride in one pass. Additional demos cover a WebGPU fluid simulation running 1,048,576 particles, a CUDA-only grassland with procedural wind audio, and an interactive 3D fiber mechanics lab from a 2026 knitting-mechanics paper.

The breadth suggests Carbon can hold a long context while switching between reading old code, rearchitecting in a new language, and adding features. The author chose Medium thinking depth for the showroom rather than High, which points to the latency and cost trade-offs developers will manage as they pick a reasoning budget for each job.