搜「Gemini 4 Carbon benchmark」的人会发现：分数很多，而且互相打架——官方表说 17 项里 15 项第一，AA 指数说 60 分，LMArena 说 1526，还有人说它「输给了 Fable 5.2」。这些说法可以同时成立，关键是分清谁测的、跟谁比、什么版本。本站把全部可核查分数盘点如下，先说总判断：Carbon 是顶级模型，但不是榜单第一；它的真正优势在长上下文、安全与任务成本，而非通用智能的绝对峰值。
公告当天发布的官方对比表，Carbon 15/17 第一，只输两项：
| 测试 | Carbon | Argon | GPT-6 Astra | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1（离线） | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2（美元余额） | 16,484 | 13,718 | 15,515 | 9,235 | 5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks 256K–1M | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — |
最值得注意的两处：GraphWalks 256K–1M 的 96.2% 是全表最大领先（第二名 Argon 只有 84.2%），印证了 Carbon 在超长上下文检索上的定位；SEC-bench Pro 的 88.2% 与 ExploitGym 的 54.8% 则支撑了它的网络安全卖点。但这张表有三个口径问题：多项分数由 Google 自行计算（SWE-sweep、Agents' Last Exam、OSWorld、GraphWalks、ExploitGym、SEC-bench Pro）；安全测试在 Antigravity 内无网、防护关闭的环境下运行；竞品数字取自对方报告而非同场运行。AA 59/58/57、Vals 70.4/70.2 这类小差距基本等于平局。最重要的是：这张表对比的是上一代 Fable 5.1——发布时 Fable 5.2 还不存在。
Artificial Analysis 智能指数是目前最受关注的独立综合榜。时间线很关键：10 月 28 日公告时，Google 引用 v5.0 的 59 分称第一；11 月 7 日指数更新至 v5.0.1，Carbon（High）60.0 分登顶，Opus 5.5 为 59、Astra 为 57；11 月 9 日 Fable 5.2 发布，以 64.6 分（约 65）超过。截至 11 月 12 日，Carbon 60.0 分居 Fable 5.2 之后。分档看：Low 54、Medium 58.3、High 60。Carbon 在 9 项子评估中的 4 项拿最高分：HLE-Diamond、Terminal-Bench-Science、AA-Omniscience、AutomationBench-AA。关于排名的一个提醒：不同统计口径（按模型计 / 按 122 个配置计）给出的 Carbon 排名在 2–4 位之间，本站只写分数与前后名次。
LMArena Text：Carbon（High）1526 预览分第一（11 月 6–7 日，约 4,483 票），此前为 1528（11 月 3–5 日）；第二名 Argon（High）1523，两者在误差范围内重叠。WebDev Arena：第 4，1768 elo。AA-Omniscience v1.1（幻觉与知识可靠性）：Carbon（high）净分 54 第一，Opus 5.5 为 48；准确率 58%（Argon 48%），但幻觉率 11% 高于 Argon 的 4%——最低幻觉率属于 Argon。CursorBench 4.0：第一（high 档 60.4%）。Perplexity WANDR 与 Zapier AutomationBench：均第一，任务成本不到次佳外部模型的一半（Google 口径）。MKA 多模态知识 Agent 基准（Google 与 AA 合作）：Carbon 第二，平均任务成本不到第一名的六分之一。Coding Agent Index（AA）：Carbon 59，对 Claude Code + Opus 5.5 的 62，落后 3 分——这是 Carbon 少数几个明确落后的独立榜单之一。FrontierSWE v2：Fable 5.2 69.9%，Carbon 61.2%。
如果说通用智能榜单上 Carbon 只是「顶级之一」，安全与诚实指标上它领先得更实在（均为 Google 公告口径，尚未有独立复现）：模拟网络攻击测试拦截 99% 攻击轨迹，单次尝试突破率 0.8%（Argon 3.1%）；Gray Swan 自动网络攻击查询成功率 2.17%（Argon 9.86%）；Trajectory Labs 在 132 小时 / 13,000+ 请求中零新的通用越狱；Gray Swan IPI 提示注入基准攻击成功率 0%（首次）；Agent 诚实基准 92% 通过（Argon 70%）；Terminal-Bench 4.0 作弊审计中，作弊尝试 3.6%→0.4%、成功作弊 2.4%→0%。
参考来源：本文为本站原创改写与整理，事实与数据核对自 Quantslant 完整官方分数表与 Artificial Analysis 榜单（2026-11-11/12）
← 返回 Gemini 4 Carbon 专题搜「Gemini 4 Carbon benchmark」的人會發現：分數很多，而且互相打架——官方表說 17 項裏 15 項第一，AA 指數說 60 分，LMArena 說 1526，還有人說它「輸給了 Fable 5.2」。這些說法可以同時成立，關鍵是分清誰測的、跟誰比、什麼版本。本站把全部可覈查分數盤點如下，先說總判斷：Carbon 是頂級模型，但不是榜單第一；它的真正優勢在長上下文、安全與任務成本，而非通用智能的絕對峯值。
公告當天發佈的官方對比表，Carbon 15/17 第一，只輸兩項：
| 測試 | Carbon | Argon | GPT-6 Astra | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1（離線） | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2（美元餘額） | 16,484 | 13,718 | 15,515 | 9,235 | 5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks 256K–1M | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — |
最值得注意的兩處：GraphWalks 256K–1M 的 96.2% 是全表最大領先（第二名 Argon 只有 84.2%），印證了 Carbon 在超長上下文檢索上的定位；SEC-bench Pro 的 88.2% 與 ExploitGym 的 54.8% 則支撐了它的網絡安全賣點。但這張表有三個口徑問題：多項分數由 Google 自行計算（SWE-sweep、Agents' Last Exam、OSWorld、GraphWalks、ExploitGym、SEC-bench Pro）；安全測試在 Antigravity 內無網、防護關閉的環境下運行；競品數字取自對方報告而非同場運行。AA 59/58/57、Vals 70.4/70.2 這類小差距基本等於平局。最重要的是：這張表對比的是上一代 Fable 5.1——發佈時 Fable 5.2 還不存在。
Artificial Analysis 智能指數是目前最受關注的獨立綜合榜。時間線很關鍵：10 月 28 日公告時，Google 引用 v5.0 的 59 分稱第一；11 月 7 日指數更新至 v5.0.1，Carbon（High）60.0 分登頂，Opus 5.5 爲 59、Astra 爲 57；11 月 9 日 Fable 5.2 發佈，以 64.6 分（約 65）超過。截至 11 月 12 日，Carbon 60.0 分居 Fable 5.2 之後。分檔看：Low 54、Medium 58.3、High 60。Carbon 在 9 項子評估中的 4 項拿最高分：HLE-Diamond、Terminal-Bench-Science、AA-Omniscience、AutomationBench-AA。關於排名的一個提醒：不同統計口徑（按模型計 / 按 122 個配置計）給出的 Carbon 排名在 2–4 位之間，本站只寫分數與前後名次。
LMArena Text：Carbon（High）1526 預覽分第一（11 月 6–7 日，約 4,483 票），此前爲 1528（11 月 3–5 日）；第二名 Argon（High）1523，兩者在誤差範圍內重疊。WebDev Arena：第 4，1768 elo。AA-Omniscience v1.1（幻覺與知識可靠性）：Carbon（high）淨分 54 第一，Opus 5.5 爲 48；準確率 58%（Argon 48%），但幻覺率 11% 高於 Argon 的 4%——最低幻覺率屬於 Argon。CursorBench 4.0：第一（high 檔 60.4%）。Perplexity WANDR 與 Zapier AutomationBench：均第一，任務成本不到次佳外部模型的一半（Google 口徑）。MKA 多模態知識 Agent 基準（Google 與 AA 合作）：Carbon 第二，平均任務成本不到第一名的六分之一。Coding Agent Index（AA）：Carbon 59，對 Claude Code + Opus 5.5 的 62，落後 3 分——這是 Carbon 少數幾個明確落後的獨立榜單之一。FrontierSWE v2：Fable 5.2 69.9%，Carbon 61.2%。
如果說通用智能榜單上 Carbon 只是「頂級之一」，安全與誠實指標上它領先得更實在（均爲 Google 公告口徑，尚未有獨立復現）：模擬網絡攻擊測試攔截 99% 攻擊軌跡，單次嘗試突破率 0.8%（Argon 3.1%）；Gray Swan 自動網絡攻擊查詢成功率 2.17%（Argon 9.86%）；Trajectory Labs 在 132 小時 / 13,000+ 請求中零新的通用越獄；Gray Swan IPI 提示注入基準攻擊成功率 0%（首次）；Agent 誠實基準 92% 通過（Argon 70%）；Terminal-Bench 4.0 作弊審計中，作弊嘗試 3.6%→0.4%、成功作弊 2.4%→0%。
參考來源：本文爲本站原創改寫與整理，事實與數據覈對自 Quantslant 完整官方分數表與 Artificial Analysis 榜單（2026-11-11/12）
← 返回 Gemini 4 Carbon 專題Search "Gemini 4 Carbon benchmark" and you will find plenty of scores that seem to contradict each other — first in 15 of 17 on Google's table, 60 on the AA index, 1526 on LMArena, and yet "beaten by Fable 5.2". All of these can be true at once; the trick is knowing who ran each test, against whom, and on which version. Here is every verifiable score, mapped. The summary up front: Carbon is a top-tier model, but not the leaderboard leader; its real advantages are long context, security and cost per task — not the absolute peak of general intelligence.
Published with the announcement: Carbon first in 15 of 17, losing only two:
| Test | Carbon | Argon | GPT-6 Astra | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|---|
| AA Index v5.0 | 59 | 50 | 57 | 58 | 49 |
| Vals Index v2.2 | 70.4% | 69.2% | 65.0% | 70.2% | — |
| SWE-sweep | 9.6% | 4.3% | 9.5% | 3.6% | 2.9% |
| FrontierSWE 2 | 61.2% | 55.0% | 65.5% | 62.3% | 56.3% |
| Terminal-Bench 4.0 (AA) | 65.2% | 57.1% | 59.1% | 59.6% | 52.0% |
| Agents' Last Exam v1.1 | 77.4% | 69.0% | 73.1% | 71.6% | — |
| OSWorld-2.1 (offline) | 81.5% | 74.7% | 80.9% | 87.2% | 76.8% |
| AutomationBench (Zapier) | 53.9% | 51.3% | 41.4% | 42.5% | 31.4% |
| Vending-Bench 2 (USD balance) | 16,484 | 13,718 | 15,515 | 9,235 | 5,422 |
| Chartography | 72.5% | 71.6% | 71.0% | 66.3% | 46.2% |
| Blueprint-Bench 2 | 58.6% | 54.4% | 49.7% | 51.2% | 41.9% |
| GraphWalks 256K–1M | 96.2% | 84.2% | 71.8% | 66.8% | 65.0% |
| RiemannBench | 80.0% | 76.0% | 72.0% | 69.6% | 65.6% |
| Terminal-Bench-Science 0.1 (AA) | 64.8% | 52.9% | 63.3% | 59.0% | 43.3% |
| CWE-bench | 71.0% | 68.0% | 68.0% | 67.0% | 58.0% |
| ExploitGym | 54.8% | 42.0% | 42.4% | 34.5% | — |
| SEC-bench Pro | 88.2% | 83.4% | 85.4% | — | — |
Two results stand out: GraphWalks 256K–1M at 96.2% is the largest lead in the whole table (next best Argon at 84.2%), backing Carbon's ultra-long-context positioning; SEC-bench Pro at 88.2% and ExploitGym at 54.8% support its cybersecurity pitch. But the table has three methodology problems: Google calculated several scores itself (SWE-sweep, Agents' Last Exam, OSWorld, GraphWalks, ExploitGym, SEC-bench Pro); security tests ran inside Antigravity offline with safeguards disabled; and rival figures come from their own reports, not the same run. Small gaps like AA 59/58/57 or Vals 70.4/70.2 are effectively ties. Most importantly, the table compares against the previous-generation Fable 5.1 — Fable 5.2 didn't exist yet when it was published.
The Artificial Analysis Intelligence Index is the most-watched independent composite, and the timeline matters: at the Oct 28 announcement Google cited the v5.0 score of 59 as #1; on Nov 7 the index updated to v5.0.1 and Carbon (High) topped it at 60.0, with Opus 5.5 at 59 and Astra at 57; on Nov 9 Fable 5.2 shipped and passed it at 64.6 (rounded to 65). As of Nov 12, Carbon at 60.0 sits behind Fable 5.2. By level: Low 54, Medium 58.3, High 60. Carbon tops 4 of the 9 sub-evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA. One caution on rank: different counting methods (by model vs by 122 configurations) place Carbon anywhere from 2nd to 4th, so this site prints scores and the models ahead and behind, not a single rank.
LMArena Text: Carbon (High) 1526 preliminary, #1 (Nov 6–7, ~4,483 votes), previously 1528 (Nov 3–5); #2 Argon (High) at 1523, overlapping within margins. WebDev Arena: #4 at 1768 elo. AA-Omniscience v1.1 (hallucination and knowledge reliability): Carbon (high) net score 54, #1, with Opus 5.5 at 48; accuracy 58% (Argon 48%), but a hallucination rate of 11% versus Argon's 4% — the lowest hallucination rate belongs to Argon. CursorBench 4.0: #1 (60.4% at high effort). Perplexity WANDR and Zapier AutomationBench: both #1, at less than half the cost per task of the next-best external model (Google's claim). MKA, the Multimodal Knowledge Agent benchmark developed by Google with AA: Carbon #2, at under one-sixth the average cost per task of the top scorer. Coding Agent Index (AA): Carbon 59 vs Claude Code + Opus 5.5 at 62 — one of the few independent boards where Carbon clearly trails. FrontierSWE v2: Fable 5.2 69.9%, Carbon 61.2%.
If Carbon is merely "one of the top tier" on general intelligence boards, its safety and honesty metrics lead more convincingly (all Google-reported, not yet independently replicated): 99% of attack trajectories blocked in simulated cyberattack testing, with a 0.8% per-attempt breach rate (Argon 3.1%); Gray Swan automated cyberattack query success 2.17% (Argon 9.86%); zero new universal jailbreaks in 132 hours / 13,000+ requests at Trajectory Labs; 0% prompt-injection attack success on the Gray Swan IPI benchmark (a first); 92% pass on the agentic honesty benchmark (Argon 70%); and in the Terminal-Bench 4.0 cheating audit, cheating attempts fell from 3.6% to 0.4% and successful cheating from 2.4% to 0%.
Reference source: this article is an original rewrite produced by this site; facts and figures were checked against Quantslant's full official score table and the Artificial Analysis leaderboard (Nov 11–12, 2026)
← Back to the Gemini 4 Carbon topic page