Google Launches Gemini 4 Hydrogen: Scores 51 on the Artificial Analysis Intelligence Index, Above Argon, Runs 2.5x Faster Than 3.8 Flash and Lifts Terminal-Bench 4.1 From 18.2% to 57%; $1.50/$7.50 per Million Tokens, Free Tier Included

Background and the problem it solves

Until now the Gemini 4 line had two flagships: Argon, opened to everyone on October 13, and Carbon, opened on November 10. Both excel at long-horizon reasoning but are slow and costly for the bulk of real development work: editing code, running tests and checking results. Hydrogen is the fast execution model for that work, positioned much like earlier Flash models. It was tested under its codename on Arena and AI Studio, a Google Cloud SKU for it was spotted yesterday, and today Google released it straight to GA.

Architecture and mechanisms

Google has not disclosed parameter count or architecture. What it has said: a 1M-token context window with output of up to 1M tokens; four thinking levels (low, medium, default high, and a new xhigh) so you pay only for the reasoning depth a task needs; and native parallel batching of file-read and search calls when exploring large repositories. Antigravity also gets Orchestrator, which switches between Carbon and Hydrogen within one session; in parallel agent setups Carbon can coordinate while Hydrogen runs worker and verifier agents.

Benchmarks

All figures are Google's own: Artificial Analysis Intelligence Index 51 at default (Argon 50), 54 at xhigh; Terminal-Bench 4.1 57% vs 18.2% for 3.8 Flash; HLE-Diamond 58.8%, best in class; Vals Tax Agent Bench 52.3%, #1 overall; Vals Legal Research Bench 55.8%, tied #1; #5 on Agent Arena; 53.4% on the Multimodal Knowledge Agent benchmark; 71.6% on Agents' Last Exam v1.1. Overreach in agentic tasks is down by up to 81%. Artificial Analysis has since published its independent evaluation: 54 on its Intelligence Index at xhigh, 1 point behind Claude Sonnet 5.5 (max) and GPT-6.1 Sol (max), and 51 at default high; an 8% hallucination rate on AA-Omniscience, among the lowest recorded; but only 54% on Terminal-Bench 4.0, below Sonnet 5.5 at 64%, with no gain from xhigh; cost per task of $1.87 at high and $3.16 at xhigh, at about 200 output tokens per second. Cursor then officially said Hydrogen is in the editor and updated its public CursorBench 4.0 leaderboard: 54.3% at default High (about $5.59 per task), up from 39.6% for Gemini 3.8 Flash; Medium is 46.7% (about $1.76) and Low is 27.9% (about $0.28). Cognition reports 38.2% on FrontierCode 2.0, just behind GPT-6.1 Sol.

Getting started

Switch the model to gemini-4-hydrogen in the Gemini API or AI Studio; both GenerateContent and the Interactions API are supported, including on the free tier. Before migrating from 3.8 Flash, read the migration checklist on the official What's new page. Use the default level as a baseline, test whether low or medium meets your latency target, and save xhigh for hard tasks. If you already use Carbon, turn on Orchestrator in Antigravity to hand repetitive execution steps to Hydrogen and save quota. On cost, the official price list puts Hydrogen input 62.5% below the $4 Argon rate that starts December 1, but at double the $0.75 promotional 3.8 Flash rate for this year and equal to the new 3.8 Flash price from January 1, 2027, so check costs before moving batch workloads. Hydrogen is also live in Cursor. Cursor's docs say it has four effort levels (xhigh, high as the default, medium, and low): keep high for everyday agentic coding, because quality drops noticeably at medium and low, and raise it to xhigh only for hard debugging. As a third-party model it is not included in Cursor plan usage, and is billed from prepaid credit on The X Subscription or the Other Models pool on legacy plans.