
Google 补发的 5 页评测报告把 Carbon 那张 17 行对照表的每一行都标上了运行方;本文用一张合并总表把这些口径落到每格,说明哪 12 行能直接横向比、哪 5 行只够比同族两代,以及 Carbon 让出的两行里只有一行是真输。
deepmind.google/models/evals-methodology/gemini-4-carbon。那个地址当时打开是空白页。1| Benchmark(口径) | Gemini 4 Carbon | Gemini 4 Argon | GPT-6 Astra | GPT-6.1 Terra | Claude Opus 5.5 | Claude Fable 5.1 | DeepSeek-V4.1-Flash | Kimi K3 | GLM-5.3 | 谁跑 |
|---|---|---|---|---|---|---|---|---|---|---|
| AA Index v5.0(带工具) | 59 | 50 | 57 | — | 58 | 49 | — | — | — | 中立 · Artificial Analysis 3 |
| Vals Index v2.2 | 70.4% | 69.2% ⚠️ 68.9% | 65.0% | — | 70.2% | — | — | — | — | 中立 · Vals AI;68.9% 出自 Argon 模型卡,版本未标 34 |
| SWE-sweep(按子任务加权奖励) | 9.6% | 4.3% | 9.5% | — | 3.6% | 2.9% | — | — | — | 混 · Carbon 与 Argon 用 mini-swe-agent 自跑,对手取 SWE-sweep 公开榜 3 |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | — | 62.3% | 56.3% | — | — | — | 中立 · Proximal 公开榜 3 |
| Terminal-Bench 4.0(AA 口径) | 65.2% | 57.1% ⚠️ 57.4% | 59.1% ⚠️ 57.9% | 50.3%(max) | 59.6% ⚠️ 66.4%(xhigh) | 52.0% | — | — | — | 中立 · Artificial Analysis;⚠️ 并列 Argon 模型卡、OpenAI 自报 Astra、Anthropic 自报 Opus 5.5 3456 |
| Agents' Last Exam v1.1(评分标准平均分) | 77.4% | 69.0% ⚠️ 39.5%(v1) | 73.1% | — | 71.6% | — | — | — | — | 混 · Carbon 与 Argon 自跑 122 题 v1.1,Astra 与 Opus 5.5 取公开榜;Fable 5.1 未进榜 34 |
| OSWorld-2.1(离线 82 题子集,部分分) | 81.5% | 74.7% | 80.9% | — | 87.2% | 76.8% | — | — | — | 混 · Carbon 与 Argon 自跑三次取最高,对手取各自公开材料 3 |
| AutomationBench(Zapier 私有集) | 53.9% | 51.3% | 41.4% | — | 42.5% ⚠️ 40.0% | 31.4% | — | — | — | 中立 · Zapier 公开榜;40.0% 由 Zapier 自跑、不设回退模型 35 |
| Vending-Bench 2(银行余额,美元) | $16,484 | $13,718 | $15,515 | — | $9,235 | $5,422 | — | — | — | 中立 · Andon Labs 公开榜 3 |
| Chartography(无工具) | 72.5% | 71.6% | 71.0% | — | 66.3% | 46.2% | — | — | — | 中立 · Surge 公开榜;Opus 5.5 自报的 89.0% 是带工具口径 3 |
| Blueprint-Bench 2(图纸相似度) | 58.6% | 54.4% | 49.7% | — | 51.2% | 41.9% | — | — | — | 中立 · Andon Labs 公开榜 3 |
| GraphWalks(256k–1M,BFS F1) | 96.2% | 84.2% | 71.8% | — | 66.8% | 65.0% | — | — | — | 统一 · Google 自跑全部模型,200 题 3 |
| RiemannBench | 80.0% | 76.0% | 72.0% | — | 69.6% | 65.6% | — | — | — | 中立 · Surge 公开榜 3 |
| Terminal-Bench-Science 0.1(AA 口径) | 64.8% | 52.9% ⚠️ 57.6% | 63.3% ⚠️ 64.6% | — | 59.0% ⚠️ 58.7% | 43.3% | — | — | — | 中立 · Artificial Analysis;⚠️ 并列 Argon 模型卡与 Anthropic、OpenAI 口径 345 |
| CWE-bench(programmatic pass@1) | 71.0% | 68.0% | 68.0% | — | 67.0% | 58.0% | 55.0% | — | — | 中立 · Collinear AI 替全部模型统一评测;同分按 pass@4 破平 3 |
| ExploitGym(缓解措施关闭) | 54.8%(10 小时) | 42.0%(10 小时) | 42.4%(不设上限) | — | 34.5%(6 小时) | — | — | — | 15.0%(6 小时) | 混 · Carbon 与 Argon 自跑,Antigravity 脚手架、无网络;对手取各自系统卡,时限不一 3 |
| SEC-bench Pro(183 题 JS 引擎拆分) | 88.2% | 83.4% | 85.4% | 83.3% | — | — | 62.8% | — | — | 混 · Carbon 与 Argon 自跑,每题 3 次取平均;对手取各自系统卡,拆分未标 36 |


model-cards/gemini-4-carbon 打开是 404。7blog.google
storage.googleapis.com
deepmind.google
anthropic.com
openai.com
deepmind.google
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.