Google on November 12, 2026 introduced Gemini 4 Hydrogen, a fast, efficient member of its Gemini 4 family built for everyday coding and agentic workflows. In the launch post, Koray Kavukcuoglu, SVP at Google DeepMind and Google's Chief AI Architect, positions Hydrogen alongside the company's frontier models, Gemini 4 Carbon and Gemini 4 Argon: those two handle deep, long-horizon reasoning, while Hydrogen is designed for rapid execution. It effectively takes over the role of Google's previous fast model, Gemini 3.8 Flash.

Hydrogen is rolling out globally starting on launch day. Developers can use it through the Gemini API in Google AI Studio and Android Studio under the model ID gemini-4-hydrogen, on both free and paid tiers; enterprises get it in Gemini for business, Google Antigravity and Gemini in Google Sheets for eligible subscriptions; and consumers get it in the Gemini app model picker and in Google Search for all users. Free Gemini app users reach it through automatic model selection, and individual Antigravity access requires a paid Google AI Pro or Ultra subscription. VentureBeat reports that Google Cloud lists the model as generally available.

Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking level, more than triples Flash's Terminal-Bench 4.1 score (57 percent versus 18.2 percent) and cuts fabricated claims by more than five times. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default setting, one point ahead of Gemini 4 Argon's 50, and 54 at a new xhigh thinking level. VentureBeat, citing Artificial Analysis's independent testing, puts Gemini 3.8 Flash at 31 on the same index. Hydrogen offers four thinking levels (low, medium, high and xhigh) and a unified context window of about 1 million tokens shared by input, reasoning and output.

The model is not the cheapest option in its tier. Google prices Hydrogen flat across the full context window at $1.50 per million input tokens and $7.50 per million output tokens. VentureBeat notes that this is twice Gemini 3.8 Flash's introductory rate, which runs through December 31, and matches Flash's scheduled standard pricing from January 1, 2027. It also costs 15 times as much per uncached token as the lowest standard rates for Anthropic's Claude Haiku 5.5, OpenAI's GPT-6 Luna and Meta's Muse Ember, though Hydrogen outscored all three on the Artificial Analysis index even at its medium setting, according to VentureBeat.

Search is one of the most visible changes. Robby Stein, Google's VP of Product for Search, called Hydrogen "the biggest model upgrade to Search yet," saying it better understands multi-part questions, typos and highly specific requests, broadens query fan-out to surface a wider range of links, and produces more visual answers such as tables and images. The Verge reports Stein said Hydrogen is now the default model in AI Mode and is used for the hardest questions in AI Overviews. TechCrunch notes the Search update is rolling out to all users globally.

For developers, Google also introduced Orchestrator in Google Antigravity, which switches between Carbon and Hydrogen within a single session. Google says that on Terminal-Bench 4.0, Orchestrator came close to Carbon's task success rate while using 24 percent less quota and finishing 19 percent faster on average. On safety, Google reports that Hydrogen reduces overreach in agentic tasks by up to 81 percent, is more than 2.3 times more resilient on the Gray Swan indirect prompt injection benchmark, and posts a 96.7 percent rate on an internal agent honesty evaluation.

Most performance figures in the launch are Google's own, and the outside numbers vary by test. In Google's Terminal-Bench 4.1 table, Hydrogen's 57 percent still trails Claude Sonnet 5.5's 63.9 percent, according to VentureBeat, which also reports that Artificial Analysis's separate Terminal-Bench 4.0 runs scored Hydrogen at 54 percent at both high and xhigh settings. DigiEditorial has relied on Google's published materials for launch claims and flags that independent results differ by benchmark version.