Google Launches Gemini 4 Hydrogen, Faster and Agent-Focused

Key Takeaways

1Google unveils Gemini 4 Hydrogen, an AI model focused on execution speed and reasoning adjustment

2The announced performance comes from Google’s evaluations, with up to 2.5 times the speed of Gemini 3.8 Flash and 57% on Terminal-Bench 4.1

3The model features enhanced safeguards and is immediately available in Gemini, Search, and AI Studio

4API access starts at $1.50 per million input tokens

Why it matters : Gemini 4 Hydrogen aims to automate more tasks while showing improvements in security and multi-agent collaboration, but its performance remains to be confirmed outside of internal testing.

Google has introduced a new model, Gemini 4 Hydrogen, touted as faster and more adaptable in reasoning capabilities. It operates within a multi-agent approach with an orchestrator and comes with enhanced safeguards. Performance figures are based on the company's evaluations, while availability starts immediately across several services.

Google Acknowledges Limitations and Introduces Quantified Safeguards

Google states that it has increased protective measures for Gemini 4 Hydrogen. According to its own metrics, the model can restrict up to 81% of attempts to bypass instructions and shows over 2.3 times greater resistance to indirect prompt injection attacks on the Gray Swan IPI benchmark. However, the company admits that the model is still prone to producing hallucinations, experiencing slowdowns, or exceeding expected execution times. During testing phases on Gemini 4 Argon, some undesirable behaviors have already been observed. Companies can thus require human validation before any sensitive action, even if the agent prepares, analyzes, and tests autonomously.

Declared Performance and Discrepancies According to Benchmarks

Google reports that Gemini 4 Hydrogen completes certain tasks 2.5 times faster than Gemini 3.8 Flash at the default high reasoning level. On Terminal-Bench 4.1, it achieves a success rate of 57% compared to 18.2% for its predecessor. These figures come from the company's evaluations and do not guarantee that applications will be developed 2.5 times faster in practice. On the Artificial Analysis Intelligence index, Hydrogen scores 51 points against 50 for Gemini 4 Argon, and rises to 54 points at the xhigh reasoning level. However, Google DeepMind's technical sheet indicates that it still lags behind Claude Sonnet 5.5 on Terminal-Bench 4.1, with 57% versus 63.9%.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Configurable Execution and Multi-Model Organization

Gemini 4 Hydrogen offers four reasoning modes, ranging from low to xhigh, to tailor the depth of thought to the task at hand. Google Antigravity includes an Orchestrator that allows collaboration among multiple models on a single project: Gemini 4 Carbon is dedicated to complex reasoning, while Hydrogen handles operations and validation of changes. On Terminal-Bench 4.0, this organization reduces average quota consumption by 24% and accelerates tasks by 19%, with a success rate close to Carbon. For example, one agent can write a feature, another can run tests, and a third can verify results before suggesting modifications. However, this functionality depends on connected tools and granted permissions.

Immediate Availability in Gemini, Search, and AI Studio

Gemini 4 Hydrogen is available for trials starting now. The global launch began on November 12 in the Gemini app, Google Search, and Google AI Studio. Free users of Gemini benefit automatically when the service determines that the model fits the query. The Gemini API is also open to developers. Depending on access conditions, Google AI Pro and Ultra subscribers enjoy additional features, particularly in Google Antigravity and Google Sheets.

Pricing and Context Window Announced on the API

On the API, the announced pricing is $1.50 per million input tokens and $7.50 per million output tokens. The context window reaches one million tokens.