Google Unveils Fast and Affordable 'Gemini 4 Hydrogen' - Free Users Can Also Use It - ITmedia NEWS

The third-party AI performance metric "Artificial Analysis Intelligence Index" scores Hydrogen at 51 points at the standard thinking level, surpassing Argon (50 points). It can complete tasks approximately 2.5 times faster than the previous generation lightweight model "Gemini 3.8 Flash." The thinking level can be selected from four tiers: low, medium, high, and xhigh, with xhigh being newly added this time. The default setting is high. CEO Sundar Pichai explained that the system is designed so that waiting time and cost occur based on the required depth of reasoning for the task.

In a comparison published by Google, they mentioned "Claude Sonnet 5.5" from U.S.-based Anthropic and "GPT-6.1 Terra" from U.S.-based OpenAI as other models in the same class. Hydrogen ranked first in seven out of the ten compared items, including "HLE-Diamond" (58.8%) for measuring reasoning ability and "MKA" (53.4%) for multimodal understanding. On the other hand, it was outperformed in three items, such as "Terminal-Bench 4.1" for coding evaluation involving terminal operations and "OSWorld-2.1" for computer operations.

The context window is 1 million tokens. API pricing is $1.5 per input token and $7.5 per output token for 1 million tokens, with pricing being uniform regardless of context length. For Argon and Carbon, the pricing during the introduction period is $2 per input token and $10 per output token, and after December 1st, it increases to $4 per input token and $20 per output token.

Unlike top-tier models that were initially limited to specific users, Hydrogen is available broadly from the start. Developers can use it through Google AI Studio and all tiers of the Gemini API, as well as from "Android Studio" (paid tier) and "Firebase AI Logic." For enterprise users, it will be provided through "Gemini for business." General users can utilize it via Gemini apps, and the free version allows usage through an automatic model selection feature based on the content of the prompt. All users are eligible for Google Search. "Google AI Pro" and "Google AI Ultra" subscribers can also use it in "Google Sheets" and development tools "Google Antigravity." The "Orchestrator" feature in Antigravity allows switching between Carbon and Hydrogen within a single session.

According to the model card released together, Hydrogen is a model fine-tuned based on Carbon, capable of taking in text, images, audio, and video, and outputting only text. The training data is up to March 2026. In internal safety evaluations compared to Gemini 3.8 Flash, the proportion of safe responses for text and multilingual content remained almost the same. However, the proportion of overly rejecting responses increased by 5.57 points, particularly noticeable in prompts that include images.

The "Frontier Safety Framework" for assessing significant risks states that while the levels in the biological field and cybersecurity field have reached a level requiring caution (caution threshold), they have not yet reached the "severe capability level" deemed to pose serious dangers.