Google Introduces Gemini 4 Hydrogen for Fast Agentic Execution

Introduction

Google DeepMind has introduced Gemini 4 Hydrogen as a high-efficiency model for everyday software development and agentic workflows. While Gemini 4 Carbon and Gemini 4 Argon are positioned for difficult, long-horizon reasoning, Hydrogen focuses on rapid execution: editing code, calling tools, running tests, searching repositories and producing work products from large collections of files.

Google says Hydrogen will be available across Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search. It will also be offered to paid Google AI Pro and Ultra subscribers in Google Antigravity.

Key points

• Large context at a flat rate: Hydrogen supports a 1-million-token input context and up to 1 million output tokens. Google lists pricing at $1.50 per million input tokens and $7.50 per million output tokens across the full input context window.
• Four reasoning levels: Users can choose low, medium, high or the new xhigh setting. At high, the model scores 51 on the Artificial Analysis Intelligence Index, compared with 50 for Argon; Google reports 54 at xhigh. These figures are vendor-reported results.
• Built for model orchestration: Antigravity’s new Orchestrator can switch between models in one session. In Google’s Terminal-Bench 4.0 testing, the combination achieved near-Carbon success rates while using 24% less quota on average and finishing 19% faster.
• A substantial upgrade over Gemini 3.8 Flash: Google reports a Terminal-Bench 4.1 score of 57%, versus 18.2% for 3.8 Flash. At the high setting, Hydrogen reportedly completes tasks 2.5 times faster and cuts fabricated claims by more than five times. It can also batch file reads and search calls in parallel.
• Broad agentic and multimodal use: The model targets debugging, repository navigation, research, data analysis, legal and tax workflows, as well as synthesis across PDFs, spreadsheets, charts, audio and video. Google also demonstrated web design, 3D coding, Sheets automation and slide-deck generation.

Why it matters

Hydrogen’s main proposition is not simply a higher reasoning ceiling. It is a division of labor inside practical workflows. Carbon or Argon can handle planning, architecture and difficult judgment, while Hydrogen executes code changes, runs tests and verifies results in parallel. That arrangement could make advanced models more economical for large development teams.

The performance claims should still be interpreted carefully. Most of the figures come from Google’s own evaluations or early partner testing, so they may not transfer directly to every repository, toolchain or production environment. Results will also depend on prompts, tool reliability and quota policies.

Even with those caveats, Hydrogen reflects a broader shift in model competition. The relevant question is increasingly not which single model is strongest, but how multiple models can be coordinated to deliver reliable results with lower latency and cost.

Source: Google DeepMind