Gemini 4 Hydrogen 跑分
下表每一行都是已公布的数字，不是本站跑出来的。Google 模型卡（结果截至 2026年11月）是一张表；2026年11月12日的介绍博客另给出 Artificial Analysis 51 / 54、Agent Arena 第 5，以及点名的 Vals 分数；Vals AI 和 OpenRouter 公布的是他们自己的数字。请不要混用不同版本的基准。
不要混用不同版本的基准。Google 的 Hydrogen 表用的是 Terminal-Bench 4.1（57.0%）。Vals AI 模型页在 2026年11月13日查阅时写的是 Terminal-Bench 4.0（54.04%）。Anthropic 的 Sonnet 5.5 发布稿用 Terminal-Bench 4.0 衡量的是另一个模型。它们不是同一份排行。Vals Legal Research Bench 在 Google 博客中是 55.8%（并列第一），我们查阅 Vals 自己的页面时是 53.85%（94 个模型中排第 6）；我们并列保留两种出处，不擅自取舍。
出处: Google DeepMind, 12 November 2026 · 出处: Google Blog, 12 November 2026 · 出处: Vals AI (independent), Retrieved 13 November 2026 · 出处: OpenRouter (independent), Retrieved 13 November 2026
| 基准 | 版本 / 说明 | Hydrogen 分数 | 设置 | 测量方 | 日期 | 链接 |
|---|---|---|---|---|---|---|
| SWE-sweep | as published | 13.5% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| Terminal-Bench 4.1 | cheating-adjusted | 57.0% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| Terminal-Bench 4.1 | vs Gemini 3.8 Flash (18.2%) | 57% | high default (blog wording) | Google announcement blog | 12 November 2026 | 出处 |
| Agents’ Last Exam v1.1 | score | 71.6% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| AutomationBench-AA | as published | 76.1% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| Vending-Bench 2 | bank balance ($) | $14,549 | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| OSWorld-2.1 | offline subset | 77.6% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| MKA | Climb WSR accuracy | 53.4% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| Chartography | as published | 66.6% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| Terminal-Bench-Science 0.1 | as published | 51.4% | — | Google DeepMind model card | Results as of November 2026 | 出处 |
| HLE-Diamond | as published | 58.8% | — | Google DeepMind model card; also named on the announcement blog | Results as of November 2026 / 12 November 2026 | 出处 |
| Artificial Analysis Intelligence Index | as published by Google | 51 | default high thinking | Google announcement blog | 12 November 2026 | 出处 |
| Artificial Analysis Intelligence Index | as published by Google | 54 | xhigh thinking | Google announcement blog | 12 November 2026 | 出处 |
| Agent Arena | live benchmark, arena.ai (Google’s report) | #5 | — | Google announcement blog | 12 November 2026 | 出处 |
| Vals Tax Agent Bench | Google’s report of a Vals figure | 52.3% (#1 overall) | — | Google announcement blog | 12 November 2026 | 出处 |
| Vals Legal Research Bench | Google’s report of a Vals figure | 55.8% (tied #1) | — | Google announcement blog | 12 November 2026 | 出处 |
| Vals Index | Vals Index (Vals’s page lists v2.2 in its write-up) | 68.49% ± 0.99 (#6 of 37) | default (Vals page) | Vals AI (independent) | Retrieved 13 November 2026 | 出处 |
| Tax Agent Bench | Vals’s own page | 52.30% ± 3.60 (#1 of 85) | default (Vals page) | Vals AI (independent) | Retrieved 13 November 2026 | 出处 |
| Terminal-Bench 4.0 | Vals’s own page; not the same edition as Google’s 4.1 | 54.04% ± 3.31 (#11 of 64) | default (Vals page) | Vals AI (independent) | Retrieved 13 November 2026 | 出处 |
| Vibe Code Bench v1.1 | Vals’s own page | 87.78% ± 2.29 (#13 of 128) | default (Vals page) | Vals AI (independent) | Retrieved 13 November 2026 | 出处 |
| Models Arena Game Development Elo | OpenRouter summary table, source column “Design Arena” | 1357 | — | OpenRouter (independent) | Retrieved 13 November 2026 | 出处 |
Google 模型卡同一张表里也列出了 Claude Sonnet 5.5 和 GPT-6.1 Terra 的对照格，那些对比列在对比页。我们不会编一张 Hydrogen 对 Sol 的官方表。Anthropic 另有一条 Sonnet 5.5 的 Terminal-Bench 4.0 数字，也写在对比页，不抄进这张 Hydrogen 分数表。