Google introduced Gemini 4 Hydrogen last night, the fast model in the Gemini 4 line. It is built for everyday coding and for agents that run many steps in a row. And it goes everywhere: all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity.
The number Google leads with is awkward for its own frontier models. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default high thinking level, ahead of Argon at 50, and reaches 54 on xhigh. Gemini 4 Carbon, only cleared for all API customers two days ago, stays the model for the hard, long problems.
The price is the real announcement
$1.50 per million input tokens, $7.50 per million output tokens, flat across the entire one-million-token context window. No surcharge past a certain length. Claude Fable 5.2 sits at $8 and $40, though it plays in a different class.
Orchestrator mixes two models in one session
To keep that pairing from being a manual job, Google is adding Orchestrator in Google Antigravity. Instead of running a session on one model start to finish, Orchestrator switches between Carbon and Hydrogen inside the same session. On Terminal-Bench 4.1 the pair reaches near-Carbon task success rates while using 24 percent less quota on average and finishing 19 percent faster, according to Google’s figures. For parallel agents, Google suggests Carbon as the coordinator with Hydrogen powering the workers and verifiers.
Against its own predecessor the jumps are large
For teams coming from Gemini 3.8 Flash, Google lists: 2.5 times faster at the high default setting, Terminal-Bench 4.1 from 18.2 to 57 percent, and more than five times fewer fabricated claims. Exploring large repositories, Hydrogen batches file reads and searches in parallel rather than inspecting files one by one. Partners saw runs 2x to 2.8x faster at 12 to 23 percent fewer tokens in testing.
Then the benchmark list: #5 on Agent Arena, arena.ai’s live benchmark, 58.8 percent on HLE-Diamond, #1 on Vals Tax Agent Bench at 52.3 percent, and tied for #1 on the Vals Legal Research Bench at 55.8 percent. For multimodal work, one million input tokens pair with up to one million output tokens, turning dense PDFs, spreadsheets, charts, audio, and recorded video into cited reports.
Search feels it first
Google is upgrading Google Search with Hydrogen at the same time, and expanding its query fan-out technique there: a broader spread of links in the answer, including ones you might not have found before. If you run a website, that is the line in this announcement to read first.
A model in this class was the gap in the Gemini 4 lineup
Carbon and Argon got the headlines, Flash did the work, and a wide gap sat between them. Hydrogen fills it and still lands ahead of Argon on the index, which for a model at $1.50 says one thing above all: a benchmark number keeps getting weaker as a justification for a frontier model’s price. The interesting part is what Anthropic and OpenAI answer with in this class, because that is where the bulk of agent runs actually happens.