Google launches Gemini 4 Hydrogen, a faster, cheaper model for everyday coding

Google DeepMind introduced Gemini 4 Hydrogen on November 13, 2026, a smaller and cheaper addition to its Gemini 4 lineup built for everyday coding and multi-step agent tasks, according to the company’s announcement. The post appeared in near-identical form on both the Google DeepMind blog and Google’s own blog, so this account treats Google DeepMind, as the model’s maker, as the primary source on specifics. Google says Hydrogen is now available in every pricing tier of Google AI Studio and the Gemini API, inside Gemini for business and the consumer Gemini app, through Google Search, and for Pro and Ultra subscribers paying for Google Antigravity.

Hydrogen sits below Google’s two frontier models, Gemini 4 Carbon and Gemini 4 Argon, which the company reserves for slow, deep reasoning on hard problems. Hydrogen is meant for quick execution: writing code, running tests, and checking changes inside agent workflows that need to move fast rather than think long. On the Artificial Analysis Intelligence Index, a third-party benchmark of model capability, Hydrogen scores 51 at its default setting, edging past Argon’s 50, and climbs to 54 when pushed to its highest “xhigh” thinking mode, Google DeepMind reports.

On price, Google DeepMind set Hydrogen at $1.50 per million tokens of input and $7.50 per million tokens of output, with that rate holding steady across the model’s entire one-million-token context window.

Google Antigravity is also gaining a feature called Orchestrator, which swaps between models such as Carbon and Hydrogen within a single session rather than relying on one model throughout. On a benchmark called Terminal-Bench 4.0, Google DeepMind says Orchestrator matched Carbon’s success rate while using 24% less quota and finishing 19% faster. The company also pitches Hydrogen as a replacement for Gemini 3.8 Flash, claiming it completes tasks 2.5 times faster at default settings, more than triples scores on Terminal-Bench 4.1 (57% versus 18.2%), and cuts fabricated claims by more than five times.

Google DeepMind also published a cluster of benchmark results for Hydrogen: fifth place on Agent Arena, a live agent leaderboard; a class-leading 58.8% on HLE-Diamond; first place overall on the Vals Tax Agent Bench at 52.3%; a tie for first on the Vals Legal Research Bench at 55.8%; and 53.4% on a new Multimodal Knowledge Agent benchmark. The company illustrated the release with demos including a voxel-building game called Geminisland, a one-shot e-commerce storefront build, and an animated spreadsheet built from Google Sheets formulas.

What this means

The release points to Google competing on cost and speed rather than raw capability. Hydrogen trails Carbon and Argon on deep reasoning by design, but its flat, mid-tier pricing and claimed speed gains over Gemini 3.8 Flash target developers running high-volume agent workloads where token costs and latency matter more than peak intelligence. The Orchestrator feature suggests Google wants customers routing work between models automatically rather than picking one model for an entire task.

What is not yet independently verified is how Hydrogen performs outside Google’s own benchmarks and partner tests; the reports cite Google DeepMind’s internal results and benchmarks it selected, including figures from unnamed enterprise partners, without outside confirmation.

Hydrogen is available now across Google’s consumer and developer products, though neither report specifies a broader rollout timeline beyond today’s launch.