Gemini 4 Hydrogen deep dive: 8% hallucination rate, beats Argon on AA index, 1M-token I/O

新智元 · wechat · 2026-11-13

Xinzhiyuan's deep dive on Gemini 4 Hydrogen frames it as breaking the "smart, fast, cheap" trilemma.

Benchmarks

51 on AA Intelligence Index at high, 54 at xhigh — beating Google's own reasoning flagship Argon (50).
Terminal-Bench 4.1: 57%; hallucination-guessing rate of just 8% on AA-Omniscience (vs 43% for Sonnet 5.5 and 52% for GPT-6.1 Sol).
Native 1M-token input and up to 1M-token output; $1.5/$7.5 per million input/output tokens.
Vertical agent leaderboards: #5 on Agent Arena live blind test, first in class on HLE-Diamond (58.8%), #1 on Vals Tax Agent Bench (52.3%) and tied #1 on Legal Research Bench (55.8%).

Showcase demos: turning Google Sheets into an animated pixel canvas (Python + LET formulas + conditional formatting); one-prompt walkable voxel worlds; an interactive 3D globe driven by live global lightning data with WebAudio; a WebGL2 water-and-glass renderer hand-built from formulas and SDF textures; one-shot luxury e-commerce site; self-checked slide decks from a YouTube video.

Orchestrator: Carbon architects, Hydrogen executes and verifies — matching Carbon's Terminal-Bench 4.0 success rate with 24% less quota and 19% faster. Enterprise tests show 2-2.8x speedups over 3.8 Flash with 12-23% fewer tokens.

Original post →

Gemini 4 Hydrogen deep dive: 8% hallucination rate, beats Argon on AA index, 1M-token I/O

More from Models