Element replaced the "Flash" era in Gemini

Google announced the Gemini 4 Hydrogen model and did not use the name "Flash" in this series. The model offers support for 1 million token input and 1 million token output. But it comes right after Carbon, which was released to Ultra subscribers this week, and is positioned not in its place, but in the daily use layer below it.

Performance of Gemini 4 Hydrogen

Google describes Gemini 4 Hydrogen as a "major, standalone upgrade" over Gemini 3.8 Flash, which came out in September. At the high default setting, it completes tasks 2.5 times faster. Moreover, the Terminal-Bench 4.1 score increases from 18.2 to 57 percent.

The rate of fabricated information drops by more than 5 times. Because the model adapts without disrupting the flow when a vehicle call fails or the call returns empty.

It scores 51 points on the Artificial Analysis Intelligence Index at the assumed high level of thinking, beating Argon's 50 points. At the highest xhigh level the score goes up to 54. According to partners' measurements, it runs 2 to 2.8 times faster than 3.8 Flash. For this reason, token usage decreases between 12 and 23 percent.

5th place in Agent Arena

Leader in its class with 58.8 percent in HLE-Diamond

1st place overall in Vals Tax Agent Bench with 52.3 percent

Vals shares first place with 55.8 percent in Legal Research Bench

53.4 percent in the Multimodal Knowledge Agent test

71.6 percent in Agents' Last Exam v1.1

It is also assertive on the security side. According to Google, Gemini 4 Hydrogen reduces task over-extension by 81 percent and under-servicing by 79 percent compared to 3.8 Flash. Unaligned actions and shortcuts decrease by 12 times. The probability of exceeding the permission limits decreases by 2 times, and it reaches 96.7 percent in the internal honesty test.

Gemini app and Search

In the Gemini app, Gemini 4 Hydrogen fits the tone, style, and conventions better. Tables and visual breakdowns are also clearer. For example, according to Google, email drafts, study guides, and text edits come ready to use on the first try. The model also calls on the right tools, like connected apps, more often.

Hydrogen is listed as "Daily help" in the model selector. There's 3.5 Flash-Lite for quick answers, 3.1 Pro (in AI Pro) for detailed research, 4 Argon for code and analysis, 4 Carbon for the most complex work. Free users access Hydrogen through Automatic model selection when available for prompting.

On Google Search, Gemini 4 Hydrogen became the default model of AI Mode. In AI Overviews, it is used for the "hardest questions". Therefore, questions with incorrect spellings, multiple parts or very specific questions are better understood. Responses are becoming more visual and include a wider range of connections.

Antigravity and price

In Google Antigravity, Gemini 4 Hydrogen comes with an "Extra" thinking layer next to the Low, Medium and High levels. The new Orchestrator dynamically switches between Carbon and Hydrogen in the same session. Thus, Carbon delegates the routine steps to Hydrogen, resulting in fewer tokens being spent while the quality remains the same.

In jobs that require parallel agents, Carbon can work as the coordinator and Hydrogen can work as the worker and verifier agent. Gemini 4 Hydrogen is available for paid Google AI Pro and Ultra subscribers in Antigravity. Developer price is $1.50 per million input tokens, $7.50 per million output tokens.