Gemini 4 Hydrogen benchmarks: beats Claude Sonnet 5.5 on most agentic tasks, lags on terminal use
GoogleDeepMind · x · 2026-11-13
Google DeepMind officially launched Gemini 4 Hydrogen, built for rapid coding, multimodal synthesis and everyday workflows — 2.5x faster than 3.8 Flash and surpassing Argon on the Artificial Analysis Intelligence Index.
Key benchmark results vs Claude Sonnet 5.5 and GPT-6.1 Terra:
- Agentic coding: SWE-sweep 13.5% (Claude 3.1%, GPT 5.8%); Terminal-Bench 4.1 cheating-adjusted 57.0%, behind Claude's 63.9%
- Knowledge work: Agents' Last Exam v1.1 71.6%, AutomationBench-AA 76.1%, Vending-Bench 2 bank balance $14,549 — all leading
- Computer use: OSWorld-2.1 offline 77.6%, behind Claude's 83.9%
- Multimodal: MKA 53.4%, Chartography 66.6%, both best in class
- Reasoning: HLE-Diamond 58.8% vs Claude 53.3% and GPT 47.3%; Terminal-Bench-Science 51.4% trails Claude's 59.9%
Overall, Hydrogen leads on knowledge work, reasoning and agentic coding, but still trails Claude Sonnet 5.5 on terminal use and science tasks.
Related event: Google Launches Gemini 4 Hydrogen Across Apps, Search and API(41 posts)→
More from Models
- Gemini 4 Hydrogen simulates real-time light refraction and dispersion in WebGL2 with pure math — GeminiApp · 2026-11-13
- In a 10-model AI house share, Claude Opus 5.5 spends the night reading The Pillow Book in the library — liminal_bardo · 2026-11-13
- Agent slop moves up the stack: models rewrite existing library code and build worse replacements — a1zhang · 2026-11-13
- Gemini 4 Hydrogen posts 13.5% on SWE-sweep, an standout result for a small-class model — marktenenholtz · 2026-11-13
- Alleged Gemini 4 Hydrogen benchmark table leaks, pitted against Claude Sonnet 5.5 — ColbyHawker · 2026-11-13
- Hands-on with Google's MKA multimodal benchmark across six AI workflows — Emergency_Abies_6508 · 2026-11-13