Gemini 4 Hydrogen release notes
Overview of Gemini 4 Hydrogen
Gemini 4 Hydrogen is a new model designed for rapid, efficient execution across diverse agentic and everyday workflows. While existing frontier models like Gemini 4 Carbon and Argon focus on deep, long-horizon reasoning, Hydrogen is engineered to provide high-speed performance for tasks such as coding, data analysis, and multi-step agentic operations. The model is now available across Google’s ecosystem, including the Gemini API, Google AI Studio, the Gemini app, and Google Search.
Multi-Model Workflows and Orchestration
Hydrogen is designed to function as a standalone model or as a high-efficiency engine within multi-model architectures. To facilitate this, Google introduced the Orchestrator in Google Antigravity, which dynamically switches between models like Carbon and Hydrogen during a single session. On the Terminal-Bench 4.0 benchmark, this orchestration approach achieved success rates comparable to Carbon while consuming 24% less quota and completing tasks 19% faster. This allows Carbon to act as a coordinator for complex architectural decisions, while Hydrogen handles worker and verifier tasks in parallel.
Performance and Developer Capabilities
Hydrogen serves as a significant upgrade over the previous Gemini 3.8 Flash, offering 2.5x faster task completion at its default setting and tripling scores on Terminal-Bench 4.1. Key performance improvements include:
- Coding and Debugging: The model performs focused, surgical edits in codebases and natively batches file-read and search operations in parallel, resulting in 2x to 2.8x faster execution in complex enterprise workflows compared to its predecessor.
- Agentic Execution: Hydrogen ranks #5 on the Agent Arena leaderboard and leads its class on benchmarks including HLE-Diamond (58.8%), Vals Tax Agent Bench (52.3%), and AutomationBench-AA.
- Multimodal Processing: With a 1 million-token input and output window, the model excels at synthesizing dense documents, audio, and video into structured reports, outperforming peers on the Multimodal Knowledge Agent (MKA) benchmark (53.4%).
Adaptable Thinking Levels
Hydrogen features four distinct thinking levels—low, medium, high (default), and xhigh—allowing users to scale reasoning depth based on specific task requirements. This design ensures that users only consume the compute and latency necessary for a given task, while maintaining competitive performance against other models in its class.
Safety and Reliability
Safety improvements in Hydrogen focus on reducing overreach and under-service in agentic environments. The model reduces overreach by 81% and under-service by 79% compared to previous iterations. It also demonstrates a 12x reduction in misaligned actions during tool use and coding tasks. Additionally, the model maintains a 96.7% honesty rate on internal evaluations and shows increased resilience against indirect prompt injection attacks on the Gray Swan benchmark.
Availability
Gemini 4 Hydrogen is available starting November 12, 2026, across the following surfaces:
- Developers: Accessible via the Gemini API in AI Studio and Android Studio.
- Enterprise: Available through Gemini for business, Google Antigravity, and Gemini in Google Sheets.
- Consumers: Integrated into the Gemini app, Google Search, and Google Sheets for paid subscribers.
Sources
- OriginalIntroducing Gemini 4 Hydrogen