AI News

Google Releases Gemini 4 Hydrogen Model and Launches Orchestrator Scheduling System

Phoenix TechnologyUpdated Nov 13 13:26

AI summary

On November 12, Google officially released Gemini 4 Hydrogen, the successor to Gemini 3.8 Flash. It is natively multimodal, with input, thinking, and output sharing a 1 million token window. On the Artificial Analysis Intelligence Index, it scores 51 points in the default high tier and 54 points in the xhigh tier, surpassing the larger-parameter Argon (50 points). API pricing is $1.5 per million input tokens and $7.5 per million output tokens. Google also launched the multi-model scheduling system Orchestrator, which is already available to Ultra users in Google Antigravity, and uses Hydrogen as the default model for AI Mode in its search engine.

Why read this

SelectedAI score 86

The original text provides Hydrogen's benchmark scores, pricing, and the ways it is opened across multiple entry points, so readers can judge its execution positioning within the Gemini 4 matrix.

Topics

Tags

Body · AI translation

Google's new Gemini 4 beats Argon on benchmarks, free to use in search engine | Machine Heart Download the client for exclusive early access November 13, 2026 10:30:38 From Beijing Updated November 13, 2026 11:13:05

Overnight, Google has released the Gemini 4 model again.

On November 12, Google officially released Gemini 4 Hydrogen and simultaneously launched the multi-model intelligent scheduling system Orchestrator.

This is a Flash-level large model (and the successor to Gemini 3.8 Flash), natively multimodal, with input, thinking, and output sharing a 1 million token window, but according to Gemini's new naming convention, Google chose the brand-new Hydrogen as the tier name for it.

In the Gemini 4 product matrix, Carbon and Argon focus on deep reasoning and complex tasks, while Hydrogen plays the role of a "fast execution engine." Hydrogen focuses on efficient operation and fast response, making it very suitable for everyday coding, agentic workflows, and multi-step data processing. Autoplay

With Hydrogen for instant 3D coding, Geminisland allows anyone to generate custom structures in real time on a shared pixel island through simple instructions. Players can freely roam and explore in first-person view.

On multiple benchmarks, Hydrogen is very competitive on knowledge work, automation, and code repair tasks:

It is benchmarked against Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra. In actual tests involving complex graph computation, SQL queries, and high-concurrency enterprise-level data workflows, Hydrogen's running speed is 2-2.8 times faster than the previous generation; at the same time, when cross-modal understanding of long videos and large codebases is required (with a million-level context fully loaded), the time to first token (TTFT) still remains at a low level.

On the Artificial Analysis Intelligence Index, Hydrogen scored 51 points at the default high setting and 54 points at the highest xhigh setting, a result that also surpasses Google's own larger-parameter Argon (50 points).

It can be seen that the Hydrogen model has added an xhigh reasoning intensity. Google recommends starting main tasks from high, and only using xhigh when encountering difficult multi-step problems, because the latter may increase latency and token consumption, and overthink on simple tasks.

With the Hydrogen model, Google has greatly promoted the Gemini 4 series, making it available to everyone, including all tiers of Google AI Studio, Gemini API, Gemini for Business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity.

However, the ways different entry points are opened differ. In the Gemini app, free users get Hydrogen responses through Auto automatic model selection, while Google AI Plus, Pro, and Ultra users can select it manually. Developers can use the free tier of AI Studio and Gemini API, but are subject to rate limits.

The Gemini 4 Hydrogen API is priced at $1.5 per million input tokens, $0.075 for cache hits, and $7.5 for output tokens.

Google said that in the actual software development process, although developers need to use frontier models at the Carbon and Argon levels for in-depth architectural judgment to design systems or solve subtle bugs, they still need a fast, standardized execution model for code writing, running tests, and verifying changes. Hydrogen was designed to serve both as a standalone model for daily work and, in combination with Carbon and Argon, as an efficient execution engine in multi-model workflows.

To help automate this multi-model collaboration, Google introduced Orchestrator in its own coding tool Google Antigravity; this mechanism can dynamically switch between models such as Carbon and Hydrogen within the same session. In the Terminal-Bench 4.0 test, Orchestrator's task success rate was close to Carbon's, while average quota consumption was reduced by 24% and completion speed increased by 19%. For workflows requiring parallel agents, Carbon can serve as the coordinator, while Hydrogen can support work agents and verification agents, thereby simultaneously building and testing changes in isolated workspaces, helping teams deliver complex projects faster without increasing token costs.

The Orchestrator feature has first been rolled out to Ultra users.

In addition to multi-model workflows, Hydrogen is also a major standalone upgrade over Gemini 3.8 Flash. It has achieved significant improvements in both speed and intelligence: under the high default setting, task completion speed increased 2.5 times, the Terminal-Bench 4.1 score more than tripled (57% vs. 18.2%), hallucinations were reduced by more than 5 times, and it can seamlessly adapt when tool calls fail or search results are empty.

Efficiency improvements are also reflected in the way the model uses tools. When exploring large codebases, Hydrogen can batch and parallelize file reads and search calls, reducing the waiting caused by inspecting files one by one. When necessary context is missing, it is also more inclined to point out information gaps rather than guess.

Google demonstrated a more intuitive office case: in Antigravity, give Hydrogen a YouTube keynote video and a design template, and it can directly generate a multi-page presentation in Google Slides, then check and fix layout issues such as overlapping text on its own. From understanding the video and organizing content to invoking applications and checking results, multimodal capabilities run through the entire task. Autoplay

What does it feel like for Hydrogen to be even faster than 3.8 Flash? It seems hard to describe anymore. Google has now made Hydrogen the default model for AI Mode in its search engine, and also uses it to support the underlying reasoning for complex queries in AI Overviews on the main search interface, providing more useful and intuitive search results to all users worldwide.

With the Hydrogen model, people can ask questions in the search box according to their natural train of thought; even if your query contains multiple parts, hard-to-understand typos, or very specific requirements, search can intelligently understand what you are looking for. In the search results, you will receive AI-generated feedback that is highly adaptive, visual, and easy to browse.

With Hydrogen's advanced reasoning technology, Google has also upgraded Query Fan-Out in search (the mechanism in which an LLM breaks down a complex question into parallel parts, retrieves and generates separately, and finally merges them), providing people with a broader range of useful links than ever before and helping them find websites they had not discovered before.

With the addition of Hydrogen, Gemini 4 has now covered Google's entire suite of products, and more people can now enjoy frontier AI.

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of Dafeng Hao, a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."

Source

Phoenix Technology