Gemini 4 Hydrogen Accelerates Google's Coding and Agents

Google deploys Gemini 4 Hydrogen, its fastest model for daily coding and agentic tasks. It arrives in Search, Gemini, Android Studio, Workspace, and Antigravity, with pricing and usage distinct from Carbon.

Google has launched Gemini 4 Hydrogen, a fast model designed for daily coding and agentic tasks, while Carbon and Argon remain reserved for heavier reasoning. Deployment extends to Google AI Studio, Gemini API, paid Android Studio, Gemini for business, the Gemini app, and Search, with access reserved for AI Pro and Ultra subscribers in Antigravity. The goal is simple: deliver the right model in the right place without making the user wait unnecessarily.

Un modèle rapide, mais pas un remplaçant de Carbon

Hydrogen serves as a fast execution layer, not a single brain to do everything. Google positions it for uses where latency matters as much as quality: repository coding, debugging, search, data analysis, multimodal work, and tool failure recovery. In this logic, Carbon orchestrates agents in parallel, then hands off to Hydrogen for work and verification tasks in Antigravity.

The system relies on a fairly clear division of roles. Carbon plans, Hydrogen executes; the recipe is less glamorous than a keynote speech but often more effective in practice.

Tarifs, contexte et arbitrages pour les développeurs

Google lists Hydrogen pricing at $1.50 per million input tokens and $7.50 per million output tokens, with a context window of 1 million tokens. Reflection levels low, medium, high, and xhigh allow adjustment of cost, quality, and latency. At this stage, the message to technical teams is clear: you can pay less than a large reasoning model for certain tasks while retaining room to scale up.

According to Google's figures, Hydrogen scores 51 on the Artificial Analysis Intelligence Index at high setting, versus 50 for Argon, then 54 at xhigh. Compared to Gemini 3.8 Flash, the company says Hydrogen is 2.5 times faster at high, achieves 57% versus 18.2% on Terminal-Bench 4.1, and reduces fabricated assertions by more than five times. These comparisons remain Google's own [to be verified] in other independent test settings.

Search, Gemini et Workspace gagnent un moteur plus nerveux

In Gemini, Google says Hydrogen improves instruction compliance and tool usage for writing, search, files, and connected apps. In Search, users worldwide should get more intuitive answers, better query understanding, visual responses, and more links thanks to an expanded query fan-out. The search engine remains the same, but its skill at breaking down a question into sub-queries becomes noticeably stronger.

Google also highlights up to 81% less agentic overshoot, up to 79% less underservice, and a 96.7% honesty rate in its internal evaluations, as well as better resistance to indirect prompt injections. In other words, the model is supposed to be less easily distracted, less prone to improvisation, and less likely to be tricked by hidden instructions.

Pourquoi ce lancement compte pour Google

Hydrogen demonstrates Google DeepMind's strategy: separate frontier reasoning from fast execution models, then distribute them in products where end users actually see them. This approach matters for the market because it directly links search, office tools, development, and agents in a single product stack. For Google, the challenge is not just raw performance but the ability to make multiple models work together smoothly. And without turning every query into a computational pilgrimage.

In Antigravity, access to Orchestrator is reserved for Google AI Ultra subscribers, while Hydrogen also arrives on AI Pro and Ultra plans for Gemini in Sheets. Google thus promotes a fine gradation logic between general public, subscribers, and developers, with a fast model for everyday use and heavier models for tasks that justify them.

Key points

  • Global launch of Gemini 4 Hydrogen across multiple Google products.
  • Announced pricing: $1.50 and $7.50 per million tokens.
  • Artificial Analysis: 51 at high, 54 at xhigh according to Google.
  • Antigravity Orchestrator combines Carbon and Hydrogen in one session.
  • Google cites 24% less quota and 19% faster performance.
  • Search receives more visual answers and expanded fan-out.

By the numbers

  • 1 million tokens — Hydrogen's context window, announced by Google.
  • 2.5 times — claimed speed compared to Gemini 3.8 Flash at high.
  • 24% — quota saved in Terminal-Bench 4.0 with Orchestrator.
  • 19% — average completion time reduced with Orchestrator.
  • 96.7% — internal honesty rate claimed by Google.

To read