Google demoed Gemini 4 Hydrogen as its fast execution model across AI Studio, the Gemini API, Gemini for business, the Gemini app, Search and Antigravity, with attackers modeled as remote actors planting indirect prompts in tool output rather than local intruders.
Update, Nov 12, 19:06 UTC · via Google Blog
Google Blog reported that "Starting today, developers can build with Gemini 4 Hydrogen" on Nov 12, 2026, and that it is "Built for rapid iteration, Hydrogen handles repository coding and debugging, interactive web development, and multimodal workflows" while "running standalone across four configurable thinking levels or alongside frontier models like Gemini 4 Carbon and Gemini 4 Argon".
Google Blog reported that developers can "Start building with the Gemini API in Google AI Studio and Android Studio" and, "if you are a paid Google AI Pro or Ultra subscriber, try it in Google Antigravity".
Koray Kavukcuoglu, SVP at Google DeepMind and chief AI architect at Google, said Hydrogen handles rapid execution while Carbon and Argon handle long horizon reasoning, per Google.
Business Insider reported Hydrogen was tested internally as a speed and efficiency model after Carbon and Argon, which Google did not confirm in its launch note.
Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at default high thinking versus 50 for Argon and reaches 54 on xhigh, while often finishing faster in its class.
Pricing is flat across the 1 million token window at 1.50 dollars per million input tokens and 7.50 dollars per million output tokens as of launch, per Google.
Google says Hydrogen runs 2.5 times faster than 3.8 Flash at default high and jumps to 57 percent on Terminal-Bench 4.1 from 18.2 percent while cutting fabricated claims by more than 5 times.
Gemini 4 Carbon can coordinate while Hydrogen powers workers in Orchestrator, which reached near Carbon success on Terminal-Bench 4.0 while using 24 percent less quota and finishing 19 percent faster, per Google.
Google says Hydrogen cuts agentic overreach by up to 81 percent and under service by up to 79 percent and proves more than 2.3 times more resilient to indirect prompt injection on the Gray Swan IPI benchmark.
That speed claim sits uneasily beside the intelligence claim because a worker tuned for quota and latency rarely beats a frontier reasoner without tradeoffs Google has not detailed.
Who audits the switch when Orchestrator trades reasoning depth for speed inside a customer repo with live credentials?
