Gemini 4 Hydrogen focuses on speed for daily use

Google introduces Gemini 4 Hydrogen, a model designed to execute coding, research, and agent tasks faster. It is available in Search, the Gemini app, AI Studio, and Google Antigravity.

On November 12, 2026, Google unveiled Gemini 4 Hydrogen, a generative AI model designed for rapid execution of coding, research, and agent tasks. The company is deploying it widely across Google Search, the Gemini app, AI Studio, Gemini for business, and Google Antigravity. The goal is simple: make AI more useful daily without increasing latency or cost per query.

A model built for speed, not just thinking

Gemini 4 Hydrogen sits below Google's more “deep” models, Gemini 4 Carbon and Gemini 4 Argon, but is designed for fast execution. Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect at Google, sums it up with a promise: “Hydrogen brings cutting-edge intelligence to everyone.”

According to the announcement, the model combines four thinking levels — low, medium, high by default, and xhigh — to adjust quality, latency, and cost depending on the task. On the Artificial Analysis Intelligence Index, it surpasses Argon at its default high level, scoring 51 versus 50, and reaches 54 at xhigh. Google also offers it at a fixed price with a context window of one million tokens: $1.50 per million input tokens and $7.50 per million output tokens.

This positioning matters because part of the generative AI market now plays on execution speed. For many professional uses, users do not always seek the most “cerebral” answer but one that produces a clean, usable result immediately. AI that saves two minutes quickly becomes the one you keep. Models may be brilliant, but teams mostly have deadlines.

In workflows, Google wants to make multiple models cooperate

Google does not present Hydrogen as a replacement for its top models but as an execution engine in mixed workflows. In Google Antigravity, a new Orchestrator can automatically switch between Carbon and Hydrogen during the same session, depending on the task nature. The idea: reserve heavy reasoning to Carbon, then assign production, testing, and verification to Hydrogen.

On Terminal-Bench 4.0, Google says this Orchestrator achieves success rates close to Carbon while consuming on average 24% less quota and finishing 19% faster. In development environments, this can translate into shorter cycles and fewer tokens consumed, thus lower costs at the team scale.

Hydrogen also stands out on iterative coding, debugging, and navigating large repositories. Google claims it completes some tasks 2.5 times faster than Gemini 3.8 Flash’s default high setting, while tripling scores on Terminal-Bench 4.1, at 57% versus 18.2%. In enterprise workflows related to graphs, SQL databases, and heavy data, partners reportedly observed speed gains of 2x to 2.8x and token consumption reductions of 12% to 23% compared to 3.8 Flash.

Search, Sheets, and the Gemini app are also involved

The release is not only aimed at developers. In the Gemini app, Hydrogen is presented as more reliable for following instructions, calling the right tool, and producing ready-to-use answers, notably for email drafts, study guides, or document summaries. In Google Search, it should better understand complex queries, including typos or mixed intents, then display more visual answers and more varied links.

Google also pushes Hydrogen into Sheets, where the model would be able to exploit financial data faster to build interactive models, with outputs more than 30% faster than previous models in the Sheets canvas. In Google Antigravity, the company demonstrates web prototyping, automatic presentations, and even scene generation in spreadsheets. The message is clear: AI must leave the showcase and enter the tools teams already use.

Security, honesty, and broad deployment

Google finally emphasizes security, a central topic as models take control of more actions. According to the announcement, Hydrogen reduces overactions in agent tasks by up to 81% and under-executions by up to 79%, while decreasing some inappropriate actions or shortcuts in tool and code evaluations by 12x. When facing impossible or blocked tasks, it is reportedly up to twice less likely to bypass authorization limits.

The company also claims a 96.7% honesty rate in its internal evaluation, 99.9% chain-of-thought readability over more than 200,000 audited exchanges, and better resistance to indirect prompt injections. On HealthBench Pro, false refusals would fall below 0.4%. These figures are provided by Google [to be verified] within its own evaluation framework but show where the battle now lies: making AI work fast without letting it improvise nonsense.

Hydrogen is deployed worldwide from the announcement for developers, businesses, and individuals via the Gemini API, AI Studio, Android Studio, Gemini for business, Google Antigravity, the Gemini app, and Search. Google is thus advancing on two fronts simultaneously: the general public on one side, professional uses on the other. And, for once, both product lines seem to be playing the same tune.

Key points

  • Launch announced on November 12, 2026, by Google.
  • Hydrogen targets fast coding and agent tasks.
  • Orchestrator switches between Carbon and Hydrogen in Antigravity.
  • Terminal-Bench 4.0: 24% less quota.
  • Search, Sheets, and the Gemini app are involved.
  • Google cites 96.7% internal honesty [to be verified].

By the numbers

  • 1 million tokens — context window announced by Google.
  • 24% — average quota saved on Terminal-Bench 4.0.
  • 19% — completion speed gained on Terminal-Bench 4.0.
  • 57% vs. 18.2% — Terminal-Bench 4.1 score versus Gemini 3.8 Flash.
  • 0.4% — false refusals on HealthBench Pro, according to Google.

To read