Google DeepMind released Gemini 4 Hydrogen on November 12, positioning it as the "fast hand" for everyday coding and agent tasks. It belongs to the fourth generation alongside Gemini 4 Carbon and Gemini 4 Argon. Those two lean toward deep reasoning over long chains of steps, while Hydrogen handles fast execution. It can be used on its own, or as the execution engine in a multi-model setup.
Pricing is $1.5 per million input tokens and $7.5 per million output tokens. Google stresses that the rate is flat across the full 1-million-token context window, with no tiers by length. Input context is 1 million tokens, and output can also run to as much as 1 million tokens.
Benchmarks: against the last generation and against Google's own flagship
Google's results mostly use the previous-generation Gemini 3.8 Flash as the reference point:
- On Terminal-Bench 4.1 it scores 57%, versus 18.2% for 3.8 Flash, more than a threefold jump;
- At the default high setting it finishes tasks 2.5 times faster than 3.8 Flash;
- Fabricated facts drop to less than one-fifth of the previous level;
- In tests by partners on complex charts, SQL and data-heavy enterprise workflows, it ran 2 to 2.8 times faster and used 12% to 23% fewer tokens.
A few other results are listed separately: 58.8% on HLE-Diamond, 52.3% on the Vals tax agent benchmark (first overall), 55.8% on the Vals legal research benchmark (tied for first), and 71.6% on Agents' Last Exam v1.1.
Reasoning levels go from three to four: on top of low, medium and high, there is a new xhigh level. Google says the levels step up "smoothly and predictably," so users wait and pay only for the depth of reasoning a task needs.
The most striking comparison is with Google's own lineup. By the Artificial Analysis Intelligence Index figures Google cites, Hydrogen scores 51 at the default high setting, against 50 for Argon; at xhigh it reaches 54. A model positioned for fast execution beats a same-generation frontier model at its default setting.
Against models at a similar price
Looking at third-party numbers, evaluations that Artificial Analysis published on social media show Hydrogen at xhigh scoring 1 point below both Claude Sonnet 5.5 (max) and GPT-6.1 Sol (max). Its cost per task is about 32% of Sonnet 5.5's, but roughly 2.5 times that of GPT-6.1 Sol. In other words, it is far cheaper than Anthropic's mid-tier model, but holds no price advantage over OpenAI's model in this class.
On the coding-tool side, Cursor says Hydrogen is already live in Cursor, scoring 54.3% on CursorBench at default effort, against 39.6% for 3.8 Flash. These figures come from Cursor and are not included in Google's blog post.
Compared with its siblings on price: Gemini 4 Carbon is $2 per million input tokens and $10 per million output tokens during its promotional period, rising to $4 and $20 from December. At the post-increase standard, Hydrogen's unit price works out to roughly 37.5% of Carbon's. In Google's own orchestrator tests in Antigravity, the company concluded that with Hydrogen as the execution end, task success was "close to Carbon," while using 24% less quota on average and running 19% faster.
Where to use it
Developers can call it through the Gemini API, AI Studio and Android Studio. Enterprise users can get it in the API, the model drop-down menu, Google Antigravity, and eligible commercial editions of Google Sheets. In the Gemini app it appears in the model drop-down menu; free users do not need to pick it manually, as the system assigns it automatically when it judges a prompt to be a good fit. In Google Search, it is rolling out to users worldwide.
Gemini in Antigravity and Sheets is available first to paid Google AI Pro and Ultra subscribers. The blog post does not mention Vertex AI, nor does it list separate API call limits for a free tier.
On safety, Google says that compared with 3.8 Flash, unauthorized actions in agent tasks are cut by up to 81%, deviant behavior and "shortcut-taking" in tool-use and coding evaluations are reduced to one-twelfth, and it scores 96.7% on an internal agent honesty evaluation. All of this is Google's own testing. The model card is public, and no third-party reproductions have appeared yet.
Sources: Google official blog, Artificial Analysis, CocoLoop, Cursor; API pricing, context length and benchmark scores were checked against Google's official blog, and per-task costs for comparable models follow the figures published by Artificial Analysis.