Google DeepMind Launches Gemini 4 Hydrogen, a Fast Model for Everyday Agentic Workflows

Google DeepMind introduced Gemini 4 Hydrogen on November 12, 2026. It is the mainstream, everyday model in the Gemini 4 family. Google calls it its "daily driver for agentic workflows" and says it brings cutting-edge intelligence to everyone. The announcement was written by Koray Kavukcuoglu, SVP of Google DeepMind and Google's Chief AI Architect, and the post was updated on November 13.

Focusing on fast execution and tool invocation

Google splits the Gemini 4 line into two roles. Gemini 4 Carbon and Gemini 4 Argon are the frontier models. They handle deep, long-horizon reasoning on the hardest problems. Hydrogen is built for rapid execution across agentic workflows and everyday coding. Google argues that teams still lean on Carbon and Argon for architectural judgment when they design systems or chase subtle bugs. They also need a fast, disciplined model for writing code, running tests and verifying changes. Hydrogen can work alone, or it can act as the high-efficiency execution engine inside a multi-model workflow.

Price and benchmark claims

At its default high thinking level, Hydrogen scores 51 on the Artificial Analysis Intelligence Index, against 50 for Argon. At the top xhigh setting it reaches 54. Pricing is flat across the whole 1 million-token context window: $1.50 per million input tokens and $7.50 per million output tokens. Google adds that Hydrogen often finishes tasks faster than models in its class. All of these figures come from Google's own announcement. External benchmark rankings or independent reproductions were not verified for this report, so these results cannot be independently confirmed.

Google compares Hydrogen with its predecessor, Gemini 3.8 Flash. At the high default setting, tasks finish 2.5 times faster. The Terminal-Bench 4.1 score rises from 18.2% to 57%, more than triple. Fabricated claims fall by more than five times, and the model adapts better when a tool call fails or a search returns nothing. On SWE-sweep, where a model must find and repair bugs without breaking existing code, Google says Hydrogen beats similar-tier models. In early partner tests on complex graph, SQL and data-heavy enterprise workflows, Hydrogen ran 2x to 2.8x faster than 3.8 Flash and used 12% to 23% fewer tokens.

On open-ended agentic work, Hydrogen ranks fifth on Agent Arena, the live benchmark run by arena.ai. For domain reasoning, Google reports 58.8% on HLE-Diamond, which it describes as the lead in its class. It reports 52.3% on the Vals Tax Agent Bench (first overall) and 55.8% on the Vals Legal Research Bench (tied for first). It also claims a class lead on AutomationBench-AA. On multimodal work, it scores 53.4% on the new Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1. The model accepts up to 1 million input tokens and can produce up to 1 million output tokens. It reads PDFs, spreadsheets, charts, audio and video.

Orchestrator in Google Antigravity

To automate multi-model work, Google introduced Orchestrator in Google Antigravity. Instead of using one model from start to finish, Orchestrator switches between models such as Carbon and Hydrogen inside the same session. On Terminal-Bench 4.0, Google says it reached near-Carbon task success while using 24% less quota on average and finishing 19% faster. In parallel-agent setups, Carbon can coordinate while Hydrogen powers worker and verifier agents that build and test changes in isolated workspaces at the same time.

Four thinking levels and customer feedback

Hydrogen offers four thinking levels: low, medium, high (the default) and a new xhigh tier. Google says quality, latency and cost scale in a smooth, predictable way, so users wait and pay only for the reasoning depth a task needs.

The announcement quotes several early customers. Box reports that accuracy improved clearly over Gemini 3.8 Flash, with about twice the speed and 12% fewer tokens. STAN reports matched accuracy, 100% format compliance, 28% lower end-to-end latency and 23% lower token use. Pixelcut reports that across 276 runs the model swept every design brief, ran 1.8 times faster and produced 30% cleaner code. LatchBio says Hydrogen passed 67% of attempts on the toxicology and safety section of its antibody development benchmark, against 37% for the best non-Gemini model, at about 13 minutes per task. These are vendor-selected testimonials, so treat them as supportive claims, not audits.

Gemini app and Search

In the Gemini app, Google says Hydrogen follows tone, format and specific rules more closely, and pulls context from uploads, Gmail, Google Drive and connected apps such as Airtable and Dropbox. In Google Search, it is meant to understand multi-part questions, typos and highly specific requests, return more visual and easy-to-scan answers, and widen the query fan-out so users see a broader range of links.

Safety figures

Google says Hydrogen uses the latest Gemini 4 safety and security stack. In comparison with Gemini 3.8 Flash, it reports up to 81% less overreach and up to 79% less under-service in agentic tasks, and a 12x cut in misaligned actions and task shortcutting across tool-use and coding evaluations. When a task is impossible or blocked, the model is up to 2x less likely to bypass permission boundaries, and it reaches a 94% score on an internal agent honesty evaluation. Across more than 200,000 audited turns, chain-of-thought legibility stayed at 99.9%, with zero deceptive reasoning edits. Under the Gray Swan IPI benchmark test conditions, Google says resilience to indirect prompt injection is more than 2.3 times that of its own comparison configuration. That does not mean the model cannot be attacked. False refusals on HealthBench Pro stay under 0.4%. These are internal evaluations. More detail sits in the model card.

Availability

Hydrogen began rolling out globally to developers, enterprises and consumers on the day of the announcement. Developers can use it through the Gemini API in Google AI Studio and Android Studio, and paid Google AI Pro and Ultra subscribers can use it in Google Antigravity. Enterprises can deploy it in Gemini for business through the API or the model drop-down, in Antigravity, and in Gemini in Google Sheets for eligible business plans. Consumers get it in the Gemini app model drop-down and in Google Search. Free Gemini app users receive it through automatic model selection when it suits the prompt. Gemini in Google Sheets starts with Pro and Ultra subscribers.

What to watch

The point of this release is lower cost of execution, not a higher ceiling. Google keeps its strongest models for judgment, puts Hydrogen on execution, and links the two with Orchestrator. Whether that design holds up depends on whether independent tests reproduce the speed, cost and accuracy numbers that Google published.