[ITmedia News] Google announces fast and low-cost "Gemini 4 Hydrogen", also available for free users

Google in the US announced its AI model "Gemini 4 Hydrogen" on November 12th (local time), and began offering it around the world from the same day. A model that emphasizes speed and efficiency for everyday coding and work with AI agents. This is the third model in the Gemini 4 series, following the Gemini 4 Argon and Gemini 4 Carbon, which are responsible for long and deep inferences (respectively derived from the element names argon, carbon, and hydrogen). The company positions Hydrogen as its main model for daily use when working with agents.

In the third-party AI performance index "Artificial Analysis Intelligence Index," it scored 51 points at the standard thinking level, exceeding Argon (50 points). Compared to the previous generation's lightweight model "Gemini 3.8 Flash", it is said to finish tasks approximately 2.5 times faster. You can choose from four thinking levels: low, medium, high, and xhigh, and xhigh is a new addition. The initial setting is high. CEO Sundar Pichai explains that latency and costs depend on the depth of inference required for a task.

In a comparison published by Google, it listed Anthropic's ``Claude Sonnet 5.5'' and OpenAI's ``GPT-6.1 Terra'' as models from other companies in the same class. Of the 10 items compared, Hydrogen came in top in 7 items, including ``HLE-Diamond'' (58.8%), which measures reasoning ability, and ``MKA'' (53.4%), which measures multimodal understanding. On the other hand, Claude Sonnet 5.5 outperformed in three categories, including ``Terminal-Bench 4.1,'' a coding evaluation that requires terminal operations, and ``OSWorld-2.1,'' which requires computer operations.

Context window is 1 million tokens. The API price is $1.50 for input and $7.50 for output per million tokens, and the fee is flat regardless of context length. Argon and Carbon will be priced at $2 for input and $10 for output during the introductory period, and after December 1st will be $4 for input and $20 for output.

While the higher-end models were initially limited in usage, Hydrogen was offered in a wide range of applications from the beginning. Developers can use all tiers of Google AI Studio and Gemini API, as well as Android Studio (paid tier) and Firebase AI Logic. For businesses, it will be provided as "Gemini for business". General users can use the Gemini app, and the free version has a function that automatically selects a model according to the prompt content. Google search targets everyone. Google AI Pro and Google AI Ultra subscribers can also use Google Sheets' Gemini and Google Antigravity development tools. Antigravity's "Orchestrator" feature allows you to switch between Carbon and Hydrogen within a single session.

According to the model card released at the same time, Hydrogen is a model adjusted based on Carbon, and can input text, images, audio, and video, and outputs only text. The training data is said to be up to March 2026. Internal safety evaluations compared Gemini 3.8 Flash showed similar rates of safe text and multilingual responses. On the other hand, the proportion of respondents who refused to answer excessively increased by 5.57 points, which was particularly noticeable for prompts that included images. The Frontier Safety Framework, which evaluates serious risks, states that the biological and cybersecurity fields have reached a level that requires caution (warning threshold), but have not yet reached the "serious capability level" at which it is determined that there is a serious risk.