Google releases Gemini 4 Hydrogen and launches automatic model switching in Antigravity…
On November 12, 2026, Google released Gemini 4 Hydrogen, aimed at everyday coding and multi-step agentic tasks, and made it available progressively the same day through channels including the Gemini API, the Gemini app, and Google Search. The model scores 57% on Terminal-Bench 4.1, higher than Gemini 3.8 Flash's 18.2% but lower than Claude Sonnet 5.5's 63.9%. The Antigravity development environment simultaneously introduced Orchestrator, which can automatically switch between Carbon and Hydrogen within the same session based on the task; the API model ID is gemini-4-hydrogen, and thinking_level has four tiers—low, medium, high, and xhigh—with high as the default.
Why read this
SelectedAI score 88
The original text shows Hydrogen and Claude Sonnet 5.5 trading wins across multiple benchmarks, which can be used to gauge its position in coding and agentic tasks.
Topics
Body · AI translation
On November 12, 2026, Google announced the AI model Gemini 4 Hydrogen. It is designed to make everyday coding and tasks in which AI uses tools to carry out multiple steps faster and more efficient. It began rolling out the same day through the Gemini API, the Gemini app, Google Search, and other channels. In the development environment Antigravity, Google also offers an Orchestrator that switches models according to the task.
Introducing Gemini 4 Hydrogen - Google
Meet Gemini 4 Hydrogen: built for rapid coding, multimodal synthesis, and everyday workflows.
It finishes tasks 2.5x faster than 3.8 Flash and surpasses Argon on the @ArtificialAnlys Intelligence Index.— Google DeepMind (@GoogleDeepMind) November 12, 2026
Faster code creation and testing
According to Google, Hydrogen can speed up writing code, running tests, and checking changes. In addition to being used on its own, it can be combined with Gemini 4 Carbon or Gemini 4 Argon, which excel at complex design decisions and reasoning to solve difficult problems, to divide up the work.
In the company's evaluation, Hydrogen scored 57% on Terminal-Bench 4.1, which measures agents' problem-solving ability, surpassing Gemini 3.8 Flash's 18.2%.
The comparison with competing models that Google published in the model card is as follows. Hydrogen surpassed the two comparison models on SWE-sweep and Agents' Last Exam, while Claude Sonnet 5.5 scored higher on Terminal-Bench 4.1, OSWorld-2.1, and Terminal-Bench-Science 0.1 [1].
Benchmark Gemini 4 Hydrogen Claude Sonnet 5.5 GPT-6.1 Terra
Terminal-Bench 4.1 57.0% 63.9% —
SWE-sweep 13.5% 3.1% 5.8%
Agents' Last Exam v1.1 71.6% 60.4% 69.8%
OSWorld-2.1 77.6% 83.9% 74.2%
Terminal-Bench-Science 0.1 51.4% 59.9% 58.3%
According to Google, Hydrogen produces fewer ungrounded responses than Gemini 3.8 Flash, and its handling of cases where a tool call fails or no search results are obtained has also improved. On safety, the company says it has reduced behavior in which an agent tries to exceed the scope of the permissions it has been given when work reaches a dead end. It also says it has increased resistance to attacks that make AI follow malicious instructions hidden in external information (indirect prompt injection).
Automatic model switching in Antigravity, Orchestrator offered
To automate the division of work among models, Google began the gradual rollout of Orchestrator in the development environment Google Antigravity the same day. When a user selects Orchestrator from the model selection menu, the model switches automatically according to the task within a single session.
When combining Carbon and Hydrogen, Carbon examines code dependencies and formulates a plan, and Hydrogen proceeds with editing code and testing. When work reaches a dead end, it returns to Carbon, which checks the situation and revises the plan before work resumes.
With the Antigravity SDK, it is also possible to assign roles to each model and have multiple agents collaborate. In an example Google presented, an agent using Hydrogen proceeds with classifying reported issues and writing code, while an agent using Carbon handles large-scale refactoring, development of new features, and security checks.
In Antigravity, Hydrogen and Orchestrator will be rolled out gradually to paid subscribers of Google AI Pro and Ultra [2]. For enterprises, these can be used via Gemini for business.
API offered even on the free tier; thinking depth can be specified in four levels
In Google AI Studio and the Gemini API, Hydrogen is offered on both the free tier and the paid tier, but the free tier has usage limits. The model ID specified in the API is gemini-4-hydrogen.
Developers can use the API's thinking_level to adjust how deeply Hydrogen thinks before answering. The setting has four levels: low, medium, high, and xhigh, with high as the default. Users can choose, depending on the use case, whether to prioritize response speed or to have the model think deeply using time and tokens [3].
In addition to coding, Hydrogen can read PDFs, spreadsheets, images, audio, and video, and compile the information obtained into reports with sources. The context window is about 1 million tokens, and this limit includes not only the input but also the model's thinking and output.
The standard pricing for the API's paid tier is $1.50 per 1 million input tokens and $7.50 per 1 million output tokens. Tokens used for thinking are also billed at the same rate as output. The rate does not change with context length.
Also rolled out to the Gemini app and Google Sheets
In the Gemini app, Hydrogen more accurately reflects detailed instructions about tone and format in its answers and more precisely calls the tools suited to the task, according to Google.
For personal Google accounts, subscribers to Google AI Plus, Pro, and Ultra can select Hydrogen in chat. For free users' chats, if automatic model selection is enabled, Hydrogen may be selected automatically depending on the input.
We confirmed that even with Google Workspace accounts, "4 Hydrogen" and "4 Hydrogen Thinking" are displayed in the Gemini app's model selection menu.
Google has also begun offering Hydrogen in Gemini within Google Sheets. Hydrogen creates calculation models from financial data, and users can verify results while changing forecast conditions. For individuals, it will be rolled out gradually starting with Google AI Pro and Ultra subscribers, and for enterprises it is available under eligible business contracts.
For enterprises, Hydrogen is also available through the Gemini for business API and model selection menu.
Also introduced to Google Search; improves understanding of questions and presentation of answers
Google has also introduced Hydrogen to Search and began rolling it out gradually to users worldwide, including free users. The company positions this as the largest model update to Search to date.
Google Search gets a major upgrade with Gemini 4 Hydrogen - Google
Robby Stein, who is in charge of Google Search products, explained in a post on X that in AI Mode, the conversational search feature, Hydrogen has become the new default model. It is also used to answer particularly difficult questions in AI Overviews, which display AI-generated summaries in search results.
In Google's official demo, "Everyday help (4 Hydrogen)" and "Most complex work (4 Carbon)" are displayed in the model selection menu. On the Japanese search screen we checked as well, a "4 Hydrogen" item was displayed in the "+" menu in the search box.
(1/6) The biggest model upgrade to Search yet has landed. Gemini 4 Hydrogen is here, and brings cutting-edge intelligence to all Search users. It’s now the new default model in AI Mode & being used for your hardest questions in AI Overviews.
> Breaks down complicated, multi-part Qs & precisely follows your prompts > Reasoning advancements improve query fan-out to show more links to the web > Much more visual, scannable & highly tailored responses (with comparison tables, charts & more) > Rich understanding of all supported languages
I’ve been testing the model behind the scenes and it’s really incredible. It’s one of my favorite models I’ve ever used, and can’t wait for you all to try it! Here are a bunch of examples I found helpful 👇— Robby Stein (@rmstein) November 12, 2026
In terms of question understanding, the company says it can more accurately grasp user intent even when multiple questions or detailed conditions are entered together, or when there are hard-to-parse input mistakes. It has also improved instruction following and tool calling, returning answers that address each of the questions contained in a query. These language understanding improvements reportedly extend to all languages supported by AI Mode.
The mechanism for finding relevant information after understanding the question has also been strengthened. Google Search uses a method called query fan-out to simultaneously search multiple topics related to a question. Hydrogen's reasoning capabilities improve this mechanism, and by presenting links to a broader range of related sites, it is said to make it easier for users to reach information they could not find before.
In terms of how answers are presented, it incorporates images, comparison tables, dedicated tools, and more according to the content of the question, making it easier to find the necessary information. In a post on X, Google's Nick Fox also cited graphs and simulations as examples. Regarding interactive diagrams and simulations that users can operate, the official help page states that they are provided in English for logged-in users.
In a thread of posts, Stein introduced an example showing a table with star ratings and a map when choosing an area to live in a destination city. He also showed examples such as explaining a statistical concept with a graph, and suggesting outfits that go with shoes you just bought along with images.
(3/6) Detailed response w/ graphs to help you breakdown a statistical concept pic.twitter.com/HVBf9VC9Cw— Robby Stein (@rmstein) November 12, 2026
(5/6) Get styling tips and visual outfit inspo for the new shoes you just bought pic.twitter.com/2zyO2yxnNb— Robby Stein (@rmstein) November 12, 2026