AI Sunday
EN

Gemini 4 Hydrogen: Google launches a fast model for code

2–3 minutes
Gemini 4 Hydrogen : Google lance un modèle rapide pour le code

Google announces Gemini 4 Hydrogen, a model designed for speed in writing code and in agent workflows. It is presented as an everyday model that can also work alongside the heavier Carbon and Argon models.

According to the press release, it is offered at $1.50 per million input tokens and $7.50 per million output tokens, at a flat rate across its one-million-token context window.

What’s happening

According to Google, Hydrogen is for executing: writing code, running tests, checking changes. Carbon remains the model in charge of architecture decisions. A feature called Orchestrator switches automatically from Carbon to Hydrogen within a single session. On a test the press release doesn’t name, Google reports a success rate close to Carbon’s, with on average 24% less quota and 19% less time.

Google also presents Hydrogen as an upgrade for teams using Gemini 3.8 Flash. The model would finish tasks 2.5 times faster at the “high” setting, its default. Partners who tested the model on enterprise workflows, with graphs, SQL and large amounts of data, saw it run 2 to 2.8 times faster, with 12 to 23% fewer tokens compared with 3.8 Flash.

The press release cites three demonstrations carried out in Google Antigravity: an online shop generated in a single attempt, a multi-slide presentation created in Google Slides, and a Google Sheets spreadsheet turned into an animated pixel canvas. In the Gemini app, Google says the model follows instructions better and calls tools better.

Key figures

  • Terminal-Bench 4.1: 57%, versus 18.2% for Gemini 3.8 Flash
  • HLE-Diamond: 58.8%, first place in its category according to Google
  • New benchmark of agentic work on visual files: 53.4%
  • arena.ai, live test: 5th place on deep research, data analysis and code navigation tasks
  • Made-up errors: divided by more than 5, according to Google
  • Output: up to 1 million tokens

In context

The press release points to another announcement in the series, Gemini 4 Argon, presented as the leading model for real-world code, enterprise knowledge and cyber defense. It is announced as “rolling out soon”, with no date.

On the pricing front, Anthropic cut the price of Claude Fable 5.2 by 20% in early November, as we reported in our article on Claude Fable 5.2. Google’s press release does not compare Hydrogen’s prices with those of other models.

What we don’t know yet

  • Availability: the text provided is truncated. It announces Gemini 4 for “all” access tiers, without detailing which plans are concerned.
  • The names of several tests: the press release gives neither the name of the benchmark where Orchestrator approaches Carbon, nor that of the test where Hydrogen comes first, nor that of the new 53.4% test.
  • The competing model for the score of 51 versus 50 (54 at the highest setting) isn’t named in the text provided.
  • The figures come from Google and its testing partners. The material provided contains no independent verification.

This news brief was written automatically by an AI system, for greater responsiveness and to bring you ever fresher news.

Sources

À découvrir aussi sur AI Sunday

Portrait de Daniel, fondateur d'AI Sunday

Écrit par

Daniel

Professionnel des achats, j'utilise l'IA au quotidien dans mon travail et je raconte ici ce qui marche vraiment, et ce qui coince. En savoir plus