What Is Gemini 4 Hydrogen? The New AI That's 2.5× Faster — for Free
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever thought, "Free AI must be less capable than paid AI?" Gemini 4 Hydrogen, announced by Google on November 12, 2026 (US time), may be the model that changes that impression. This article clearly breaks down its features, pricing, how to use it, and what to watch out for.
Gemini 4 Hydrogen is a new AI model Google announced on its official blog. It is the third model in the Gemini 4 series, following Argon and Carbon.
The name comes from the element hydrogen. It is positioned as a fast, efficient model suited for everyday programming and AI agent tasks (AI that carries out work on behalf of people).
Google calls it the "workhorse model" for agentic work. It is the successor to the lightweight model "Gemini 3.8 Flash," and Google recommends it as the migration target from 3.8 Flash.
According to Google, at the standard setting (high), it completes tasks approximately 2.5× faster than Gemini 3.8 Flash. Verification by partner companies showed speeds of 2 to 2.8× faster, with token usage (the unit of text AI reads and writes) reduced by 12 to 23%.
On the programming benchmark "Terminal-Bench 4.1," the score rose from 18.2% with 3.8 Flash to 57%.
According to ITmedia NEWS, on the Intelligence Index compiled by third-party evaluation firm Artificial Analysis, the model scores 51 at the standard setting — slightly edging out the higher-tier model Argon (50 points). At the highest setting, "xhigh," the score reaches 54.
Google's announcement also includes the following figures. These are all from Google's own announcement and have not been verified by independent parties.
Thinking levels come in four tiers: low, medium, high, and xhigh. The newly added xhigh tier improves scores but increases both cost and time. The default setting is high.
API pricing for developers is $1.50 per million input tokens and $7.50 per million output tokens. The per-unit price remains the same even for long contexts of up to one million tokens.
Argon and Carbon are scheduled to move to $4.00 input and $20.00 output starting December 1. Compared to those prices, Hydrogen's per-unit cost is roughly 37.5%.
Free users of the Gemini app can access it through automatic selection based on the content of their query. According to gori.me, users cannot manually specify it from the model selection menu. Google assigns it through the "Auto" setting.
In Google Search, it is available to all users. Among the Gemini 4 series, Hydrogen is the only model eligible for the free tier of the Gemini API. However, since rollout is gradual, not everyone will have access on the day of the announcement.
On the other hand, Gemini in Google Sheets and the development tool Google Antigravity are limited to subscribers of paid Google AI Pro or Ultra plans.
Even within the same Gemini 4 family, the roles are divided. Carbon handles deep-thinking tasks such as design and auditing, while Hydrogen handles editing, testing, and parallel execution.
A feature in Antigravity called "Orchestrator" switches between Carbon and Hydrogen within a single task. According to Google, this maintains a success rate close to Carbon's while reducing average quota consumption by 24% and speeding up completion by an average of 19%.
Consider a scenario where an engineer at a small company needs to rewrite legacy code. The smart Carbon handles the overall architecture, while the fast Hydrogen runs fine-grained edits and tests in parallel. The more expensive model is only used for the parts that require deep thinking.
Google compared Hydrogen against Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra across 10 benchmark categories. Hydrogen ranked first in 7 categories, including HLE-Diamond and MKA.
Claude Sonnet 5.5 ranked first in 3 categories, including Terminal-Bench 4.1 and OSWorld-2.1. This suggests that competing models outperform Hydrogen on computer operation and terminal tasks.
It has also been pointed out that the figures for competing models in the comparison table come from mixed sources depending on the category. It is most natural to read Google's table as results measuring Hydrogen at its highest setting. The safest approach is to try it yourself and see whether it suits your use case.
For everyday users in Japan, the main change is that a faster model will be working behind the scenes in their regular Gemini app and Search experience. This is the biggest difference, and there is nothing users need to switch on themselves.
There are some points to be aware of. Content sent via the free tier may be used by Google to improve its products. According to reports, Japan — unlike the EEA (European Economic Area), Switzerland, and the United Kingdom — is not included in the exception that grants free users the same treatment as paid users.
For this reason, please do not enter confidential business information or personal data into the free tier. For business use, the paid API or Gemini for Business is the safer option.
For example, consider a startup developer building a prototype chatbot for internal use. The free tier may be sufficient through the testing phase, but once the production version handles customer data, switching to a paid plan becomes necessary.
No specific figures regarding Japanese-language performance were found in Google's announcement. Please try it yourself to evaluate Japanese language quality.
According to Google's model card, the model reached the warning threshold in the biological and cyber domains. However, it did not reach the Critical Capability Level (CCL) for dangerous capabilities.
Compared to 3.8 Flash, safe responses to inputs containing images improved by 7.00 points. At the same time, "over-refusal" — where the model declines even safe questions — also increased by 5.57 points. A detailed safety report is listed as "coming soon."
Additionally, when migrating from 3.8 Flash, note that explicit caching is not supported, and parameters such as temperature may be ignored.
A. Yes. Free users of the Gemini app can access it through automatic selection. A free tier is also available for the Gemini API. The usage limit is not publicly disclosed; you can check it on the AI Studio interface.
A. Free users cannot select it by name from the model selection menu. Google automatically assigns it based on the content of the query.
A. Through December 31, pricing is $0.75 input and $3.75 output. From January 1, 2027, it will move to $1.50 input and $7.50 output.
A. With the free tier, content you submit may be used to improve Google's products. Do not enter confidential information or personal data. A paid plan is recommended for business use.
Start by sending your usual questions in the Gemini app and see how fast the responses are.
This article is a cross-post from AI Friends.