Google unveils Gemini 4 Hydrogen, replacing Flash in the consumer app

Google unveiled Gemini 4 Hydrogen on Nov. 12, 2026, replacing Flash in the consumer Gemini app and beginning a worldwide rollout in the app and Google Search at 18:00 UTC, with broader availability starting around 3:00 AM JST on Nov. 13.

Google bills the model as bringing cutting-edge intelligence to everyone. It succeeds Gemini 3.8 Flash, which launched in September. Free users receive Hydrogen through Auto selection when it matches the prompt, while Google AI Pro and Ultra subscribers can run it inside Google Antigravity.

Developers can access it using the Gemini API, Google AI Studio, or Vertex AI, and enterprises can deploy it through the API or choose it from the Gemini Enterprise model menu. It is also coming to Gemini in Google Sheets, starting with Pro and Ultra subscribers. Hydrogen did not go through the Fairwind Program stage and was available to everyone at launch, including free users.

Google describes Hydrogen as a model built for rapid execution and everyday agent tasks, while Carbon and Argon handle deep long-horizon reasoning. The company says Carbon takes the most difficult tasks, Argon covers general programming and knowledge, and Hydrogen handles daily speed-prioritized work. Argon opened to paying developers on Oct. 13 after nearly two weeks with Fairwind Program members. For design decisions and unclear bugs, Google recommends Carbon or Argon when facing them, with Hydrogen assigned to writing or verifying code.

At its default high thinking setting, Google says, Hydrogen finishes tasks at 2.5 times the speed of Gemini 3.8 Flash. On the Artificial Analysis Intelligence Index, Hydrogen scored 51 at default high thinking, edging out Argon at 50, and reaches 54 at its xhigh setting. Artificial Analysis ranks its highest setting one point below Anthropic‘s Claude Sonnet 5.5. Among the top six models in Vals AI’s ranking, Hydrogen is the cheapest and fastest.

Partner tests showed Hydrogen running up to 2.8 times faster on graphs, SQL, and data-heavy enterprise tasks while cutting token use by up to 23 percent.

Google also says Hydrogen beats its own frontier Argon model on the independent Artificial Analysis ranking. On knowledge benchmarks, Google says its model leads OpenAI‘s GPT 6.1 Terra and Anthropic’s Claude Sonnet 5.5. In science and math, computer use, and Terminal-Bench 4.1, it falls behind those rivals, though Google describes the gap as small. Hydrogen posts 57 percent on Terminal-Bench 4.1, more than tripling Flash’s 18.2 percent.

For agentic work, Orchestrator in Google Antigravity moves between Carbon and Hydrogen during one session, and the orchestrator lets Carbon hand off routine steps to Hydrogen while preserving quality and using fewer tokens. For parallel agent work, Carbon handles coordination while Hydrogen runs worker and verifier agents across isolated workspaces. Google says Orchestrator neared Carbon’s success rate on Terminal-Bench 4.0 while cutting average quota use by 24 percent. Finished runs were 19 percent faster, but Google says the figures describe the model mix, not Hydrogen alone.

According to Google, Hydrogen cuts agentic-task overreach by as much as 81 percent and under-service by as much as 79 percent compared with Flash. It reduces misaligned actions and task shortcutting in tool-use and coding tests by 12 times compared with Flash, and cuts fabricated claims by more than five times. It is up to two times less likely to bypass permission boundaries on impossible or blocked tasks than Flash. On Google’s internal agent honesty evaluation, Hydrogen scored 96.7 percent.

Resistance to indirect prompt injection improved by more than 2.3 times, according to Google’s internal evaluation. The safety results come from Google’s internal evaluation, and external verification has not yet arrived.

Hydrogen leads its class on HLE-Diamond at 58.8 percent, takes first on the Vals Tax Agent Bench at 52.3 percent, and ties for first on the Vals Legal Research Bench at 55.8 percent. It leads on AutomationBench-AA and places fifth on Agent Arena for deep research, data analysis, and codebase navigation. It scores 53.4 percent on the Multimodal Knowledge Agent benchmark and 71.6 percent on Agents’ Last Exam v1.1.

Hydrogen serves as the default model for Google Search AI Mode and handles the hardest questions in AI Overviews. Google’s model picker calls it Everyday help, while Flash-Lite covers quick answers and Carbon covers the most complex work. Because it now powers Google Search for every user, Google says answer changes could reach people who are not actively using AI. Information sent through the free API tier is used for product improvement, Google says.

Google says Hydrogen follows tone, format, and rules more closely, and calls the right Connected Apps and skills more often. It also improves understanding of multi-part, typo-filled, or highly specific queries, offering more visual responses and more relevant links.

This is the latest development in Google’s Gemini 4 rollout. Three Gemini 4 models have shipped in just over six weeks, and Hydrogen arrived two days after Carbon became available to paying developers and Ultra subscribers. Anthropic released Claude Fable 5.2 on November 9, and Google released Carbon on Nov. 10; OpenAI had not announced a new model as of that writing.

Koray Kavukcuoglu said in late September that Google wanted Gemini 4 to arrive well before the end of the year. Community testing under the gemini-3.8-flash name took place in late October, and internal staff testing was reported on Nov. 11, 2026. Kavukcuoglu described it as “built for rapid execution.”