Google launches Gemini 4 Hydrogen, a cheaper model for coding, agents and Search

A faster model for everyday work

Google released Gemini 4 Hydrogen on 12 November 2026, a model built for quick execution rather than deep reasoning. The company's frontier models, Gemini 4 Carbon and Gemini 4 Argon, handle hard problems that call for careful judgment, while Hydrogen is aimed at agentic workflows, in which an AI carries out several steps on its own. Google describes a split in which Carbon coordinates parallel agents and Hydrogen runs the working agents and checks that their changes execute correctly. The model replaces Gemini 3.8 Flash across Google's products, according to Ars Technica. Google also launched Orchestrator in its Antigravity coding tool, which switches between Hydrogen and Carbon and, by Google's account, gets close to Carbon's results while using 24% less quota.

Benchmarks and speed

Google puts Hydrogen's speed at 2.5 times that of Gemini 3.8 Flash on its default setting. On the Artificial Analysis Intelligence Index, Hydrogen scored 51 at its default high reasoning level, one point above Gemini 4 Argon at high (50) and well above Gemini 3.8 Flash at high (31). At the top of its four thinking levels, xhigh, Hydrogen reached 54, against 50 for Argon, according to Ars Technica.

Gemini 4 Hydrogen51

Gemini 4 Argon50

Gemini 3.8 Flash31

Gemini 4 Hydrogen

51 points

Gemini 4 Argon

50 points

Gemini 3.8 Flash

31 points

Data from the article text

In Google's own testing, reported by Europa Press, Hydrogen leads Claude Sonnet 5.5 and GPT-6.1 Terra in seven of ten benchmarks, all centered on organizational and developer workflows. The widest gap appears in agentic coding on SWE-sweep, where Hydrogen scored 13.5% against 3.1% for Claude and 5.8% for GPT. On Terminal-Bench 4.1, Hydrogen completes 57% of tasks, up from 18.2% for Flash, but Google's table puts Claude Sonnet 5.5 at 63.9%.

Gemini 4 Hydrogen57%

Claude Sonnet 5.563.9%

Gemini 3.8 Flash18.2%

Gemini 4 Hydrogen

57 %

Claude Sonnet 5.5

63.9 %

Gemini 3.8 Flash

18.2 %

Data from the article text

In an early-access test, Box reported that its agent moved from 74% accuracy with Flash to 79% with Hydrogen on its Complex Work Eval. Vals AI calls Hydrogen the cheapest and fastest model in its top six.

Pricing and access

Developers pay $1.50 per million input tokens and $7.50 per million output tokens, rates that Ars Technica says are unchanged from Gemini 3.8 Flash. Hydrogen accepts up to 1 million input tokens at once. Batch, Flex and an invite-only Contributor tier each cost half the standard rate. Batch works through groups of requests in the background with a target turnaround of 24 hours, while Flex handles single requests at lower priority with a target response time of 1 to 15 minutes. The Contributor tier offers standard speed and reliability, but in return Google may use data from those requests to improve its products and train its models.

Access is where the reports diverge. Europa Press lists availability from Thursday across all plans of Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search, plus AI Pro and Ultra subscribers in Google Antigravity. Ars Technica reports that AI Plus and free users remain on the older 3.5 Flash-Lite model and that access is more restricted in the Gemini app and AI Studio than in Search. The Next Web reports that free users of the Gemini app and Google Search can use Hydrogen.

Search rollout

Google calls this the biggest model improvement it has made to Search. Robby Stein, Google's VP of product for Search, said Hydrogen is now the default model in AI Mode and handles the hardest questions in AI Overviews. The model is built to parse multi-part questions and typos, and its fan-out technique now runs simultaneous queries across subtopics so users find sites they might have missed. Responses also gain images, comparison tables, graphs and maps. Google's example shows an interactive animated driving simulation explaining how phones locate a car's lane where there is no cell service. Google says availability is global and covers all languages supported by AI Mode, and Ars Technica notes that Google has made similar fan-out claims before.

Limits and rivals

Google still presents Carbon and Argon as the stronger choices for long-horizon tasks and security work. Artificial Analysis places Hydrogen's top setting one point behind Claude Sonnet 5.5. Google says Hydrogen makes up facts five times less often than Gemini 3.8 Flash, while Ars Technica describes the reduction in fabricated claims as more than five-fold. The launch comes as OpenAI has recently cut GPT-6 prices with cheaper models of its own, according to The Next Web.