Gemini 4 Hydrogen accelerates daily and agentic uses
Google unveils Gemini 4 Hydrogen, a model designed for fast execution, agentic workflows, and everyday uses. It arrives in Search, the Gemini app, AI Studio, Sheets, and Google Antigravity.
On November 12, 2026, Google introduced Gemini 4 Hydrogen, an AI model designed to quickly execute everyday tasks, from coding to searches, while remaining accessible to a broader audience. The company is deploying it in the Gemini app, Google Search, AI Studio, Gemini API, Google Sheets, and Google Antigravity. The goal is simple: to offer an AI less focused on slow, deep reasoning and more useful when a clean result is needed without waiting for a flood of bytes. Hydrogen also targets agentic workflows, that is, sequences of tasks carried out by multiple agents or tools.
Un modèle pensé pour aller vite sans perdre le fil
Google describes Hydrogen as a "rapid execution" model for daily uses and agentic tasks, whereas Gemini 4 Carbon and Gemini 4 Argon remain reserved for the most complex problems. The positioning is clear: Hydrogen is meant to serve as a fast execution engine, alone or complementing heavier models. On the Artificial Analysis Intelligence Index, it surpasses Argon with 51 versus 50 at its default thinking level, then 54 with the xhigh setting.
This logic extends to several product surfaces. In Google Antigravity, the Orchestrator tool dynamically switches between Carbon and Hydrogen during the same session. Google reports achieving, on Terminal-Bench 4.0, a success rate close to Carbon’s, while consuming 24% less quota and finishing 19% faster. For teams alternating design, testing, and verification, the promise is clear: lower cost, less waiting, less manual juggling between models.
Des gains annoncés pour les développeurs et les entreprises
For development uses, Hydrogen replaces Gemini 3.8 Flash as the upgrade option for teams seeking more speed and greater accuracy. Google claims that at its default setting, it completes tasks 2.5 times faster, while more than tripling Terminal-Bench 4.1 scores, from 18.2% to 57%. The model also reportedly reduces fabricated assertions by more than five times and performs better when a tool call fails or a search returns nothing.
In tests cited by Google, Hydrogen also improves iterative coding, debugging, and Q&A on large repositories, with the ability to read multiple files in parallel. The company indicates partners have observed speed gains from 2x to 2.8x and token consumption reductions from 12% to 23% compared to Gemini 3.8 Flash. On broader benchmark grounds, Hydrogen ranks 5th on Agent Arena, reaches 58.8% on HLE-Diamond, 52.3% on Vals Tax Agent Bench, and 55.8% on Vals Legal Research Bench.
Google mise aussi sur les usages multimodaux
Hydrogen is not limited to text or code. The model accepts up to 1 million tokens input and output to synthesize dense PDFs, spreadsheets, charts, audio, and video recorded in structured reports. Google presents it as particularly useful for knowledge workflows, with a score of 53.4% on the Multimodal Knowledge Agent benchmark and 71.6% on Agents’ Last Exam v1.1.
This capability also powers very concrete demonstrations in Google Sheets, Slides, or Antigravity. The model can transform financial data into interactive models, generate a complete presentation from a keynote video, and then fix layout issues. In Sheets, Google even announces outputs more than 30% faster than previous models. Here, AI does not just look good on a slide; it produces usable work. The spreadsheet didn’t need it, but it got a facelift anyway.
Une disponibilité large, avec un ticket d’entrée plus bas
Google opens access to Hydrogen in several consumer and professional products: AI Studio, Gemini API, Gemini app, Google Search, Gemini for business, Google Sheets, and Google Antigravity. Google AI Pro and Ultra subscribers can also use it in certain workflows. For developers, the announced price is $1.50 per million input tokens and $7.50 per million output tokens, across the entire one-million-token context window.
The model offers four thinking levels — low, medium, high (default), and xhigh — to adjust latency, cost, and quality according to the task. This choice matters: Google aims to normalize the use of an advanced model without forcing everyone to pay the maximum for tasks that don’t require it. In Search, Hydrogen should also make answers more visual, better understand complex queries, and broaden the scope of suggested links.
Sécurité, alignment et tâches sensibles : le vrai sujet de fond
Google emphasizes model safety, a central point as systems act increasingly autonomously. According to the company, Hydrogen reduces agentic task overflows by up to 81%, underperformance by up to 79%, and misaligned actions by a factor of 12 in some evaluations. Its internal honesty rate reaches 96.7%, while its resistance to indirect prompt injections is more than 2.3 times higher on the Gray Swan IPI benchmark.
Another sensitive point: the accuracy of safeguards on biomedical topics. Google says it reduced false rejections to less than 0.4% on HealthBench Pro to better distinguish real health questions from biosecurity risks. The company also claims to preserve 99.9% readability of the thought chain, with zero edits deemed misleading over more than 200,000 audited sessions. This is probably where part of the market is decided: not only going faster but convincing organizations that fast AI can remain controllable.
By the numbers
- 24 % — less quota on Terminal-Bench 4.0 with Orchestrator, Google.
- 19 % — time saved on Terminal-Bench 4.0, Google.
- 2.5x — speed gained compared to Gemini 3.8 Flash, Google.
- 1 million tokens — announced input and output context for Hydrogen.
- $1.50 and $7.50 — price per million input and output tokens.