Google announced Gemini 4 Hydrogen on November 12, 2026. The company says the model is built for everyday tasks that need quick answers, unlike the more heavyweight Gemini 4 Carbon and Argon which target deep reasoning. Hydrogen runs at the default "high" thinking level but can also be pushed to a new "xhigh" tier for extra intelligence.
In Google’s own tests, Hydrogen scores 51 on the Artificial Analysis Intelligence Index at the high level and up to 54 at xhigh, edging out Argon’s 50. On the Terminal‑Bench 4.1 benchmark, Hydrogen reaches 57 % compared with Flash’s 18.2 %, a 2.5‑times speed improvement at the same level. Hallucinations – the model’s tendency to make up facts – dropped by a factor of five, and the model recovers better when tool calls fail.
The model can read and write up to one million tokens in a single request. Pricing is $1.50 per million input tokens, $7.50 per million output tokens, with cache hits at $0.075. At the high tier a typical task costs about $1.87; at xhigh the cost rises to $3.16 because it uses roughly twice as many output tokens.
Who can use Gemini 4 Hydrogen?
Hydrogen is being rolled out globally. Developers can call it through the Gemini API in AI Studio or Android Studio. Paying Pro and Ultra subscribers access it via Google Antigravity, while businesses get it through Gemini for Business. The Gemini app and Google Search also expose the model to anyone with a Google account.
In the Gemini app the model is said to stick to tone and format longer, making drafts and summaries more usable on the first try. It can pull context from Gmail, Drive, Airtable and Dropbox, and Google says it improves search results, especially for multi‑part or typo‑laden queries.
How does Hydrogen compare to rivals?
Independent testing by Artificial Analysis puts Hydrogen on par with Claude Sonnet 5.5 at its highest level and slightly behind GPT‑6.1 Sol at the default high level. On the AA‑Omniscience benchmark Hydrogen fabricates only 8 % of its output, a lower hallucination rate than the other Gemini models.
However, Hydrogen lags on pure terminal work. It scores 54 % on Terminal‑Bench 4.0 and 50 % on Terminal‑Bench‑Science 0.1, while Claude Sonnet 5.5 reaches 64 % and 53 % respectively. Price‑performance analysis shows Hydrogen costs about a third of Claude Sonnet 5.5 for a comparable intelligence level, but about 2.5 times more than GPT‑6.1 Sol.
Why it matters
For developers, Hydrogen offers a cheaper, faster alternative to larger Gemini models when building chatbots, summarizers or code assistants. For everyday users, the model’s tighter adherence to prompts and better search handling should make Google Search and the Gemini app produce more useful answers with less tweaking.
