Gemini 4 Hydrogen: Google unveiled a new AI that completes tasks up to 2.5 times faster

Google introduced Gemini 4 Hydrogen, a new model of artificial intelligence that aims to combine greater speed, reasoning ability, and efficiency. The company describes it as a significant upgrade over Gemini 3.8 Flash, with improvements in both performance and execution of complex tasks.

Among its main features, it includes a context window of up to one million input tokens and another million output tokens. It will also be available in various Google products, such as the Gemini app, the search engine, and the developer platform Antigravity.

Gemini 4 Hydrogen: what are the improvements over Gemini 3.8 Flash

Google presents Hydrogen as an independent evolution of Gemini 3.8 Flash, launched in September. According to tests released by the company, the new model offers advancements in speed, accuracy, and handling of external tools.

Among the improvements highlighted by Google are:

• Speed: completes tasks 2.5 times faster with its default high-performance setting.

• Tool usage: more than triples the results of Terminal-Bench 4.1, with 57% compared to 18.2% of Gemini 3.8 Flash.

• Accuracy: reduces invented claims by more than five times.

• Adaptability: responds better when a tool fails or a search returns no results.

• Efficiency: Google partners recorded speeds between two and 2.8 times higher, with a reduction in token usage of between 12% and 23%.

In the Artificial Analysis Intelligence Index, Hydrogen scored 51 points with the default high reasoning level and reached 54 with the most demanding setting, called XHigh. In this way, it surpassed Argon, which scored 50 points.

How Gemini 4 Hydrogen performs in complex tasks

Google also highlighted the model's results in research, programming, and professional work evaluations. In Agent Arena, it ranks fifth in multi-step tasks, such as deep research, data analysis, and navigating large codebases.

The company also reported the following results:

• HLE-Diamond: 58.8% in reasoning and specialized analysis.

• Vals Tax Agent Bench: 52.3%, with the overall first place.

• Vals Legal Research Bench: 55.8%, a result that shares the first place.

Compared to Gemini 3.8 Flash, Hydrogen also reduces excessive actions in tasks with agents by up to 81% and cases where it provides less assistance than necessary by up to 79%. According to Google, it decreases misaligned actions and shortcuts in programming evaluations and tool usage by 12 times.

What changes in the Gemini app with the new model

In the Gemini app, Hydrogen improves the tracking of instructions related to tone, format, and rules indicated by the user. Google claims it can produce drafts of emails, study guides, rewrites, and ready-to-use documents from the first attempt.

It also more frequently selects the appropriate connected tools and applications for each task. In the model selector, the company positions it as the option intended for everyday assistance.

The model offerings are organized as follows:

• Quick responses: Gemini 3.5 Flash-Lite.

• Everyday help: Gemini 4 Hydrogen.

• Detailed research: Gemini 3.1 Pro, available in AI Pro.

• Programming and analysis: Gemini 4 Argon.

• More complex tasks: Gemini 4 Carbon.

How much does Gemini 4 Hydrogen cost for developers

Google reported that the use of Hydrogen for developers will have the following prices:

• Input tokens: US$ 1.50 per million.

• Output tokens: US$ 7.50 per million.