Google Releases Gemini 4 Hydrogen for Agentic Workflows

• Google released Gemini 4 Hydrogen on 12 November 2026, describing it as a model for agentic workflows. It is rolling out that day across developer, enterprise and consumer products.^[1]
• Developers get it in the Gemini API through Google AI Studio and Android Studio. Businesses get it through Gemini for business, and Google Antigravity gives it to paid Google AI Pro and Ultra subscribers.^[1]
• Consumers can pick it in the Gemini app model dropdown, free users reach it through automatic model selection, and Google Search upgrades to it for all users globally.^[1]
• It follows Gemini 4 Argon (30 September) and Gemini 4 Carbon (28 October), and fills the lower-cost tier below Carbon, whose introductory pricing is $2 and $10 per million tokens.^[8]

Notable features

• API pricing is $1.50 per million input tokens and $7.50 per million output tokens, flat across the full 1 million-token context window, which also allows up to 1 million output tokens.^[1]
• It takes PDFs, spreadsheets, charts, audio and recorded video as input and produces cited reports, code, formulas, presentations and web apps.^[1]
• There are four thinking levels (low, medium, high as the default, and a new xhigh).^[1]
• In Antigravity, the Orchestrator can switch between models in one session, for example Carbon coordinating while Hydrogen runs worker and verifier agents. Google says this gets near-Carbon task success with 24% less quota and 19% faster completion on average.^[1]
• Google says it finishes tasks 2.5 times faster than Gemini 3.8 Flash at the high setting and cuts fabricated claims by more than 5x.^[1]
• Reported scores: Artificial Analysis Intelligence Index 51 at high (Argon is 50) and 54 at xhigh; Terminal-Bench 4.1 57% against 18.2% for Gemini 3.8 Flash; HLE-Diamond 58.8%; Vals Tax Agent Bench 52.3%; Agents' Last Exam v1.1 71.6%. These are Google's own figures^[1], though Artificial Analysis now lists the 54 at xhigh on its own page, at about 190 output tokens per second and $3.16 per index task, and calls the model very verbose.^[2]
• In Google's launch table Hydrogen leads Claude Sonnet 5.5 and GPT-6.1 Terra on seven of ten benchmarks but trails Sonnet 5.5 on Terminal-Bench 4.1 (57.0% against 63.9%), OSWorld-2.1 (77.6% against 83.9%) and Terminal-Bench-Science 0.1 (51.4% against 59.9%).^[4]
• Fello AI's read of Artificial Analysis data puts the default high setting at 51 against 31 for Gemini 3.8 Flash and 60 for Carbon, at $1.87 per index task against $7.77 for 3.8 Flash.^[3]
• The model card says Hydrogen is adapted from Gemini 4 Carbon, has a March 2026 knowledge cutoff, and returns 64K output tokens by default or up to 1M through background long decode. Its automated tests show unjustified refusals up 5.57 points against 3.8 Flash, and the full safety report was still to come.^[4]
• Google also reports safety gains, including agentic overreach down by up to 81% and a 96.7% score on its internal honesty evaluation.^[1]