Atlas

Models

Google

Gemini 4 Hydrogen

166.8
Capability · 95% interval 164.8–168.9
Rank 9 of 895 (interval 5–12)
Relative cost 1.8× · $0.18 per short-answer task · $1.27 per agentic task

Gemini 4 Hydrogen (`gemini-4-hydrogen`) is Google's fast Gemini 4 model for everyday coding and agentic workflows, adapted from Gemini 4 Carbon and built for rapid execution on its own or as a sub-agent orchestrated alongside the frontier Carbon and Argon models. Google made it generally available on November 12, 2026 through the Gemini API, Google AI Studio, Google Antigravity, Gemini for business, Google Sheets, the Gemini app and Google Search, at $1.50 per million input and $7.50 per million output tokens across the full context window. It accepts text, image, audio, video, and PDF input and returns text, with a unified 1,048,576-token input-plus-output context window, 65,536 default synchronous output tokens or up to 1,048,576 with long decode, and low, medium, high (the default) and a new xhigh thinking level.

Scored on 156 benchmarks · 207 total scores · 201 independently reported

Score providers

Organizations reporting headline scores for 156 benchmarks

In fitin the capability model (18 of 156)
Legal · 68 benchmarks · first place on 5
Legal Research Bench Administrative/Regulatoryx-high
1 / 85———
Legal Research Bench Constitutional/Civil Rightsx-high
1 / 85———
Legal Research Bench Healthx-high
1 / 85———
Legal Research Bench Weighted Pass Constitutional/Civil Rightsx-high
1 / 85———
Legal Research Bench Weighted Pass Healthx-high
1 / 85———
Finance & Accounting · 35 benchmarks · first place on 4
Tax Agent Bench — Forms & Filingsx-high
1 / 78———
Finance Agent v2.1 - Financial Modelingx-high
1 / 37———
Tax Agent Bench — Citation-Adjusted All-Passx-high
1 / 78———
Tax Agent Bench — Citation-Adjusted Forms & Filings All-Passx-high
1 / 78———
Tax Agent BenchIn fitx-high
2 / 78~2 mo~Jul 2027—
Code · 17 benchmarks · first place on 3
Terminal-Bench 4.1high
1 / 1——2026-11-12
Code Migration v1.1 (Vals Index v2.2 subset)x-high
1 / 33———
SWE-sweepIn fitx-high
1 / 12~3 mo2028–20302026-11-12
Terminal-Bench 4.0 — Hardwarex-high
2 / 56———
Terminal-Bench 4.0 — Softwarex-high
3 / 56———
Indexes · 11 benchmarks
Vals Finance Indexx-high
2 / 37———
AA-Omniscience v1.1 Indexx-high
2 / 71———
Artificial Analysis Finance & Accounting Indexx-high
2 / 25———
Artificial Analysis Legal Indexx-high
2 / 25———
Artificial Analysis Engineering Indexx-high
3 / 25———
Agents · 9 benchmarks
AutomationBench 1.0.6 Financehigh
2 / 2——2026-11-12
AutomationBench-AA Financex-high
2 / 25———
AutomationBench-AA Support and Operationsx-high
2 / 25———
AutomationBench-AA Operations, HR, Marketing, Sales and Supportx-high
2 / 25———
AutomationBench-AA 1.0.6In fitx-high1 of 2
3 / 160~2 mo~Apr 2027—
General QA · 6 benchmarks
Humanity's Last Exam Diamondx-high
4 / 140——2026-11-12
Humanity's Last Exam Diamond (Knowledge Partition)x-high
4 / 96———
AA-Omniscience Law Accuracyx-high
4 / 25———
Humanity's Last Exam Diamond (Reasoning Partition)x-high
5 / 96———
AA-Omniscience v1.1 Accuracyx-high
8 / 71———
Safety · 3 benchmarks · first place on 2
AA-Omniscience Business Non-Hallucinationx-high
1 / 25———
AA-Omniscience Law Non-Hallucinationx-high
1 / 25———
AA-Omniscience v1.1 Hallucination Ratex-high
3 / 71———
Multimodal · 2 benchmarks
ChartographyIn fitx-high
5 / 47~2 mo~May 20272026-11-12
BuseyBench SVGIn fitnot reported
12 / 286~6 mo~Oct 2027—
Professional Work · 2 benchmarks · first place on 1
GDP.pdf (Artificial Analysis, All-pass)In fitmedium
1 / 158~2 mo~Dec 2028—
GDP.pdf (Artificial Analysis, Mean Criteria)high
4 / 158———
Science · 2 benchmarks
CritPt v2 (agentic, avg@5)In fitx-high
8 / 72~2 mo~Feb 2027—
Terminal-Bench-Science 0.1In fitx-high
9 / 74~1 mo~Mar 20272026-11-12
Games · 1 benchmark
Vending-Bench 2not reported
3 / 67——2026-11-12
Loading Atlas data…