Junious Media NetworkMonday, November 16, 2026
Juni/Tech
Advertisement
AI

Google Launches Gemini 4 Hydrogen for Search and Developers

Google replaced Gemini 3.8 Flash with Gemini 4 Hydrogen across Search and developer tools, adding faster query handling and agentic coding support.

Key points

  1. Google made Gemini 4 Hydrogen available to all Search users globally, replacing Gemini 3.8 Flash as the default model in AI Mode.
  2. At its default thinking level, Gemini 4 Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash and reduces fabricated claims by more than 5 times.
  3. The Orchestrator feature in Google Antigravity automatically switches between Gemini 4 Carbon and Hydrogen, finishing tasks 19% faster while using 24% less quota.
  4. Developers can access Gemini 4 Hydrogen through Google AI Studio and the Gemini API, with input tokens priced at $1.50 per million and output tokens at $7.50 per million.

Google has rolled out Gemini 4 Hydrogen to all Search users worldwide, replacing the previous Gemini 3.8 Flash as the default model in AI Mode. The company called it the biggest model upgrade to Search in the product's history.

The release extends to Google AI Studio, the Gemini API, Gemini for business, and Google Antigravity for paid Google AI Pro and Ultra subscribers. Free users in the Gemini app will get access through automatic model selection when Hydrogen is the best fit for a prompt, Google DeepMind said.

What Changed in Google Search

The main change in Search is how the model reads difficult, layered questions. Users can type queries with multiple parts, typos, or narrow requirements, and the system figures out the real intent before forming a reply, according to Google's Search team. That covers travel rules, shopping criteria, and specialized topics phrased in plain language.

Robby Stein, VP of product at Google Search, said the model is now the default in AI Mode and handles the most complex questions in AI Overviews, The Verge reported. Responses now carry more visuals, including images, comparison tables, graphs, and maps. The model runs queries across related subtopics at the same time, connecting users to a wider set of websites than before.

Language coverage expanded as well. Users searching in any language that Google's AI models support now get more precise, nuanced answers, the company said.

What Gemini 4 Hydrogen Means for Everyday Users

For people who rely on Google products daily, the model delivers a more consistent experience. It tracks a user's tone, format, and specific instructions more closely, so email drafts, study guides, and document rewrites come back ready to use on the first attempt, Google DeepMind said. The Gemini app will also more often call the right tools for a given task, cutting back-and-forth.

Advertisement

The model draws context from a user's uploads, Google apps such as Gmail and Google Drive, and connected services like Airtable and Dropbox. When comparing products or working through dense research material, the model builds clear tables and visual summaries without leaving out key details, Google DeepMind said.

Gemini integration in Google Sheets is available to users, though access begins at the Google AI Pro subscription tier and above. The model generates outputs in Sheets canvas more than 30% faster than earlier models, Google DeepMind reported.

Advertisement

Developer Access and Coding Capabilities

Developers can build with Gemini 4 Hydrogen through the Gemini API in Google AI Studio and Android Studio, as well as through Gemini for business, Google said. The model runs across four thinking levels - low, medium, high (the default), and a new xhigh tier - letting developers control the tradeoff between speed, cost, and reasoning depth.

The model handles repository coding, debugging, and iterative changes with focused edits that do not touch unrelated files. It groups file-read and search calls together rather than inspecting one file at a time. In early benchmarks involving SQL pipelines, graph workloads, and large enterprise data tasks, partners saw it run 2x to 2.8x faster than Gemini 3.8 Flash. Token use also fell by 12% to 23%, Google DeepMind said.

For multimodal work, the model combines a 1 million-token input window with up to 1 million output tokens. It synthesizes dense PDFs, spreadsheets, charts, audio, and recorded video into structured reports. Each million input tokens costs $1.50 and each million output tokens costs $7.50, with that rate held flat across the full context window, according to Google DeepMind.

Advertisement

For web development, the model generates functional layouts in fewer prompts. Google DeepMind said it handles live UI prototyping and design-to-code tasks faster than tools in a comparable class, which can reduce revision rounds before a working prototype is finished.

Speed and Benchmark Gains Over the Previous Model

Several direct measurements favor the new model. At its default high thinking setting, Gemini 4 Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash. Its Terminal-Bench 4.1 score rose to 57%, up from 18.2% for the older model. Fabricated claims dropped by more than 5 times compared to the predecessor, Google DeepMind reported.

On the Multimodal Knowledge Agent benchmark, the model scored 53.4%, a measure of end-to-end agentic tasks involving dense visual files. It scored 71.6% on Agents' Last Exam v1.1, which tests end-to-end professional workflows. For open-ended agentic work, it ranks fifth on Agent Arena, a live benchmark run by arena.ai, Google DeepMind said.

Advertisement

It also ranks first on the Vals Tax Agent Bench with a 52.3% score and shares the top spot on the Vals Legal Research Bench at 55.8%. On the Artificial Analysis Intelligence Index, it scores 51 at the default high setting versus 50 for Gemini 4 Argon, and reaches 54 at the xhigh level, Google DeepMind reported.

Multi-Model Workflows and the Orchestrator Feature

Google also introduced Orchestrator in Google Antigravity. Instead of locking one model in for the whole session, Orchestrator picks between Carbon and Hydrogen based on what each step demands. On Terminal-Bench 4.0, Orchestrator reached close to Carbon-level success rates while using 24% less quota and finishing 19% faster on average, Google DeepMind said.

Advertisement

For teams running parallel agent setups, Carbon takes the coordinator role while Hydrogen handles worker and verifier agents. That arrangement lets teams build and test changes at the same time across separate workspaces, which can speed up complex projects without raising token costs, according to Google DeepMind.

While Hydrogen is built for fast execution, Gemini 4 Carbon is designed for complex architectural decisions and difficult bug analysis. Teams can use Hydrogen for writing code, running tests, and verifying changes, then call on Carbon when a problem needs broader context, Google DeepMind said.

Safety Updates and What to Watch

Google DeepMind said Hydrogen includes updates from its latest safety stack. For agentic tasks, the model cuts unwanted overreach by as much as 81% and reduces under-service by as much as 79%, compared to Gemini 3.8 Flash. Misaligned actions and shortcuts during tool-use and coding work drop by 12 times, the company said.

In Google's internal evaluation for agent honesty, the model scores 96.7%. When blocked or impossible tasks arise, it is 2 times less prone to crossing permission boundaries. It keeps 99.9% of its chain-of-thought steps legible, with no deceptive edits found in more than 200,000 turns reviewed by auditors, Google DeepMind reported. The model is also more than 2.3 times more resistant to indirect prompt injection attacks on the Gray Swan benchmark, the company said.

On HealthBench Pro, the rate of incorrect refusals stays below 0.4%, meaning the model handles medical questions without blocking legitimate queries, Google DeepMind said. The Verge noted that Google pointed to Claude Sonnet 5.5 and GPT-6.1 Terra as competing models in the same class. How those comparisons hold up under real-world use will become clearer over the coming months.

Advertisement

Sources