Google has launched Gemini 4 Hydrogen, an execution-focused lightweight model designed to power everyday coding tasks and agentic workflows across Google Search, the Gemini app, AI Studio, and developer APIs. Replacing the previous Flash naming convention as part of a periodic element branding shift that includes Argon and Carbon, the new model aims to reduce reliance on costly frontier systems. In software engineering assessments on Terminal-Bench 4.1, Gemini 4 Hydrogen achieved a 57% task success rate, marking a 3.1-fold improvement over the 18.2% recorded by Gemini 3.8 Flash. On the Artificial Analysis Intelligence Index, Hydrogen scored 51 under its default high inference setting, overtaking Gemini 4 Argon score of 50, and reached 54 under the xhigh configuration.
To broaden developer adoption, Google established a uniform API pricing tier across a one-million-token context window, setting rates at $1.50 per million input tokens and $7.50 per million output tokens. This flat rate removes tiered cost structures for processing large context batches, providing cost predictability for developers running continuous tool invocations and large-scale codebase evaluations. Koray Kavukcuoglu, vice president at Google DeepMind and Google chief AI architect, explained that while frontier models such as Carbon and Argon handle deep, long-horizon reasoning for complex challenges, Hydrogen is engineered for rapid execution. Within the Google Antigravity orchestrator, Carbon manages high-level planning while Hydrogen carries out repetitive execution duties like code generation and tool calling to curb token consumption and latency.
The deployment also restructures the infrastructure economics supporting consumer web traffic. Robby Stein, vice president of product for Google Search, stated that introducing Gemini 4 Hydrogen across Google Search enables more precise handling of complex, multi-part queries and misspelled searches alongside richer visual exploration. By installing Hydrogen as the default backend for the AI mode in Search, Google aims to improve serving efficiency for hundreds of millions of daily queries. The rollout illustrates a wider industry shift toward multi-model orchestration, combining specialized lightweight engines with large reasoning models to make agentic software viable at enterprise scale.