Gemini 4 Hydrogen: benchmark results

Google DeepMind model adapted from Gemini 4 Carbon, released publicly through the Gemini API and apps on November 12, 2026. Low, medium, high (default) and xhigh thinking levels; text, image, audio and video inputs. Provider: Google. Released 2026-11-12. Access: API.

Aggregate Rating 1942 ± 21, rank #11 of 10871 rated models, from 80 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
AA GDP.pdf32.8All Criteria Pass Rate (%)100
AA GDP.pdf - Insurance - Criterion Pass Rate91.46Criterion Pass Rate (%)100
AutomationBench-AA - Operations - Raw Objective Completion97.6Raw Objective Completion (%)100
Google Hydrogen Release - Agents Last Exam v1.1 - Oct17 agy 1.4.871.6Mean rubric score (%)100
Google Hydrogen Release - Chartography - no tools66.6Accuracy (%)100
Google Hydrogen Release - MKA - Climb WSR53.4Climb weighted satisfaction rate (%)100
Google Hydrogen Release - OSWorld 2.1 - offline82 max377.6Offline partial score, best of three runs (%)100
Google Hydrogen Release - SWE-sweep - agy 1.4.813.5Subtask-weighted reward (%)100
Google Hydrogen Release - Terminal-Bench 4.1 - mini-swe vision57Cheating-adjusted average reward (%)100
Google Hydrogen Release - Terminal-Bench Science 0.1 - mini-swe vision51.4Average reward (%)100
Vals AI Tax Agent Bench52.31Accuracy (%)100
AutomationBench-AA - Finance - Raw Objective Completion95.3Raw Objective Completion (%)99.4

Interactive version: theaggregate.ai/model?slug=gemini-4-hydrogen · How It Works · Data refreshed daily, snapshot 2026-11-12.