Published: 13 November 2026
Google released Gemini 4 Hydrogen on 12 November, and the fast model now answers questions in Google Search for all users worldwide. It scores 51 on the Artificial Analysis Intelligence Index, one point above Google’s larger Gemini 4 Argon. For developers it replaces Gemini 3.8 Flash at $1.50 per million input tokens, the price Flash itself moves to on 1 January.
Search answers now come from Hydrogen in every AI Mode language, Dutch included
Google says Hydrogen now runs in Google Search for every user, in all languages that AI Mode supports. AI Mode has worked in Dutch since October 2025, so searches in the Netherlands get the new model too. Google promises better answers to long questions with several parts and to searches with typos. It also promises more links, and more images, tables and small tools built for the question. It gives no figures for these claims and does not say whether AI Overviews also use Hydrogen.
In the Gemini app, Hydrogen sits in the model menu. Free users get it only when automatic model selection picks it for a prompt. Google AI Pro and Ultra subscribers can also use it in Antigravity, Google’s platform for coding agents, and in Gemini in Google Sheets. Companies get it through Gemini for business, and in Sheets on business plans that qualify.
Hydrogen costs what Gemini 3.8 Flash will cost from 1 January
In the Gemini API and Google AI Studio, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Thinking tokens count as output. The price is the same for short and long prompts, up to about one million tokens. Cached input costs $0.075 per million. AI Studio also has a free tier with rate limits.
Gemini 3.8 Flash costs half that until 31 December, through a promotion on Google Cloud. On 1 January 2027 it moves to the same $1.50 and $7.50. The token price is therefore a poor guide. What counts is the cost of a finished task.
The table shows where the two models differ for a team that wants to switch.
| What is measured | Gemini 4 Hydrogen | Gemini 3.8 Flash |
|---|---|---|
| Price per million input / output tokens | $1.50 / $7.50 | $0.75 / $3.75 until 31 December, then $1.50 / $7.50 |
| Default thinking level | High | Medium |
| Terminal-Bench 4.1, agent tasks solved | 57% | 18.2% |
| Average cost per task in Artificial Analysis’ test set, high thinking | $1.87 | $7.77 |
Prices, thinking levels and Terminal-Bench scores come from Google; the cost per task is Artificial Analysis’ measurement as reported by VentureBeat, and your own workload may land elsewhere.
Claude Sonnet 5.5 still solves more terminal tasks than Hydrogen
Google reports a score of 51 on the Artificial Analysis Intelligence Index at Hydrogen’s default high setting, against 50 for Argon. At the new xhigh setting, Hydrogen reaches 54. Artificial Analysis also ran Terminal-Bench 4.0, a test of coding agents that work in a command line. Hydrogen solved 54% of the tasks there, while Claude Sonnet 5.5 solved 64% at its highest setting, The Decoder reported. VentureBeat lists Sonnet 5.5 at $2 and $10 per million tokens, a third more per token.
Google also claims Hydrogen finishes tasks 2.5 times faster than 3.8 Flash and makes up facts more than five times less often. It publishes no method for either figure. Google says partners measured 2 to 2.8 times faster runs and 12% to 23% fewer tokens in early tests, again without test details.
Temperature settings and explicit caching stop working when you switch
Google’s migration guide for developers lists changes that can break code written for 3.8 Flash. Hydrogen ignores custom temperature, top_p and top_k values. Explicit caching, where you store a prompt yourself, is not supported, so cached_content fields have to go. Automatic caching still works for prompts of 4,096 tokens or more. The minimal thinking level returns an error, and the default is high instead of medium. That adds thinking tokens to every call unless you lower it.
A single call returns at most 65,536 output tokens, and longer output needs background execution. The guide gives no retirement date for 3.8 Flash, so existing integrations keep running while you test.
Run real cases on both models before Flash’s discount ends on 31 December
Most organisations will meet Hydrogen first in Search and the Gemini app, without making any choice. Answers with tables and tools look more finished, but they need the same source check before staff reuse them in a report or a client email.
Teams that run document processing, classification or customer service agents on 3.8 Flash should act now. Flash’s price doubles on 1 January either way. Put a sample of real cases through both models in December, and compare tokens, errors and time per finished case. Start Hydrogen at medium thinking: by the same Artificial Analysis measure, an average task costs $0.80 there, less than half the cost at high.
Access is not the problem, because the Netherlands is on Google’s country list for the Gemini API. Where the data is processed is less clear. Google’s launch posts name no European data region for Hydrogen, so confirm the processing location with Google before you send it personal data.
For more information, visit the official announcement of Gemini 4 Hydrogen on the Google blog.
