Google DeepMind 在 Argon 发布仅两周后推出了 Gemini 4 Carbon，新模型在 Artificial Analysis 智能指数中登顶。
Carbon 在最近更新的 Artificial Analysis 智能指数 5.0 版本中得分为 59 分，而前代产品 Argon以 50 分紧随其后。除了代理任务（agentic tasks）和编程方面的普遍提升外，Google 还强调了其覆盖整个网络防御生命周期的能力，从检测和验证漏洞到修补漏洞。
Carbon 在 AA 总指数中领先，但在 Claude Code 的 Opus 5.5 面前，其在代理编码方面仍略逊一筹。| 图片来源：AA
专为多日编码会话设计
Google DeepMind 表示，Carbon 是为“更难、长周期工程问题”而设计的，这些问题需要持续的代理工作流。据 Google 称，内部团队连续多天运行该模型，在此期间，它自行处理了死胡同，并完成了大型代码库和低层软件中的根本原因分析和细微错误修复等复杂的多步骤任务。
Google 还报告称，根据用户反馈，Carbon 在科学推理、多模态理解和全栈 Web 开发方面有所改进。Web 开发中更高质量的输出和工件使 Carbon 在 WebDev Arena 上提升了 89 个 Elo 点，目前排名第四，得分为 1,769。
在 Google 自己的基准测试选择中，Gemini 4 Carbon 在 16 项测试中的 15 项中排名第一，尤其在长上下文和安全任务方面拉开了与竞争对手的差距。| 图片来源：Google DeepMind
在 Gray Swan 的间接提示注入基准测试中，Carbon 将攻击成功率从 Argon 的 0.7% 降至 0.0%，据 Google 称，它阻止了所有测试的操纵尝试。这并不意味着提示注入问题已解决，但该测试似乎已达到饱和状态。
独立评估支持 Google 的数据
根据 Artificial Analysis 的数据，Carbon 在其最高推理层级（“high”）下主要凭借在复杂推理、知识检索和科学任务中的高分领跑智能指数 v5，在该指数的九项单独评估中四项位居第一。
在 Antigravity 的代理编码方面，尽管相比 Argon 从 53 分跃升至 59 分，Carbon 仍以三分之差落后于使用 Opus 5.5 的 Claude Code。在 AA-Briefcase 的代理知识工作方面，Carbon (high) 相比 Argon (high, 1,483 Elo) 增加了 126 分，达到 1,609 Elo，这得益于呈现质量（+293 Elo 至 1,598）和分析质量（+120 Elo 至 1,671）的提升。但它仍然落后于 Kimi K3.1 Pro（1,678 Elo）、Muse Wonder（max, 1,659 Elo）和 Grok 4.7（xhigh, 1,638 Elo）。
与 Argon 相比，答案准确率从 48% 上升至 58%，而幻觉率从 4% 攀升至 11%。虽然更高，但这仍然是排名中第二低的水平。Claude Opus 5.5 (max) 和 GPT-6 Astra (max) 均位于 50%。
评估平台 Vals AI 将 Carbon 排在第二位，准确率为 70.36%，仅次于 Claude Sonnet 5.5（70.85%），高于 Claude Opus 5.5（70.18%）和排名第四的 Argon（69.24%），且成本更低。
在 Vals Index v2.2 中，Gemini 4 Carbon 的准确率为 70.36%，介于 Claude Sonnet 5.5（70.85%）和 Claude Opus 5.5（70.18%）之间。每次测试成本为 10.51 美元，运行时间不到 48 分钟，比这两家 Anthropic 模型更便宜、更快。| 图片来源：Vals AI
在 Vals.ai 上，Carbon 还在金融指数和法律研究基准测试中名列前茅。在法律研究中，Vals AI 记录显示，与 Argon 相比，Carbon 引用了约 45% 更多的来源，工具调用更频繁，响应也更长。
但在金融和税务代理基准测试中，任务耗时是 Argon 的两倍多；在法律研究中，Carbon 平均每个任务耗时超过 50 分钟。该模型生成的推理令牌数量是 Argon 的 2.3 倍，输出令牌数量多出 64%，但这仍不到 Sonnet 5.5 和 Opus 5.5 输出令牌数量的一半。
在 Arena.ai 上，Gemini 4 Carbon (High) 以 1,528 分领先文本竞技场，略微超过 Argon（1,520 分）。在代理竞技场中，该模型在超过 21,900 个真实会话中排名第二，优于 GPT-6 Astra 和 Fable-5.1，净改善幅度为 16.2%，并具有最高的可引导性得分（+18.6%）。其中位任务成本为 1.80 美元，比排名第一的 Claude Opus 5.5 (Max) 低 72%。
Arena.ai 将 Gemini 4 Carbon (High) 排在代理竞技场的第二位，净改善幅度为 +16.2%，仅次于 Claude Opus 5.5 (Max, +21.1%)，优于 GPT-6 Astra (Max, +13.4%)。| 图片来源：Arena
在代码竞技场的 WebDev 类别中，Carbon 达到 1,768 分（领先第十一名的 Argon 89 分），总排名第四，仅落后于 Claude Sonnet 5.5 (xHigh, 1,770 分)。这使其在所有三个学科中都处于帕累托曲线上。
Carbon 在无护栏情况下主导安全基准测试
对于几项安全测试，Google 在未启用安全护栏的情况下运行了该模型。在衡量现实世界软件中自主漏洞发现的 SEC-bench Pro 上，Carbon 得分为 88.2%，领先所有测试模型。在测试系统是否能将发现的漏洞转化为有效利用程序的 ExploitGym 上，该模型达到 54.8%。Carbon 还在 CWE-Bench 编写修复程序方面保持大幅领先优势。
在 SEC-bench Pro（88.2%）、CWE-bench（71.0%）和 ExploitGym（运行十小时达 54.8%）上，Gemini 4 Carbon 均领先于 GPT-6 Astra 和前代产品 Argon。| 图片来源：Google DeepMind
Google 还引用了无法独立验证的内部结果。据报告，在内部基准测试中，Carbon 生成安全代码的频率是 Argon 的两倍多；在 Google Chrome 团队的测试中，它在编写安全补丁方面的表现比一家未具名竞争对手的前沿模型高出十倍多。
安全专家率先获得访问权限
Google 正在逐步推出该模型，并参与美国政府的新预发布测试流程，理由是 Carbon 具备先进的安全能力。本周起，选定的网络安全专家将率先获得访问权限。随后，通过 Google DeepMind 的 Fairwind 计划，访问权限将扩展到更多受信任的合作伙伴。
该公司计划在收集更多反馈并调整安全护栏后，通过其产品界面向开发者、企业和消费者开放该模型。
Gemini 4 Carbon 以与 Argon 相同的入门价格发布：每百万输入令牌 2 美元，每百万输出令牌 10 美元，缓存输入令牌可享受常规输入价格 95% 的折扣。
在与 Argon 几乎相同的单次任务成本下（7.76 美元 vs. 7.68 美元），Gemini 4 Carbon 提升了 9 分至 59 分，接近帕累托曲线的顶部，同时也超过了更昂贵的 Claude Opus 5.5（10.04 美元）。| 图片来源：Artificial Analysis
在 Artificial Analysis 智能指数中，Carbon 的单次任务成本为 7.76 美元，比 Claude Opus 5.5 (max, 10.04 美元) 低约 23%，大约是 GPT-6 Astra (max, 6.25 美元) 的 1.2 倍。它平均每个任务生成 123,000 个输出令牌，比 Anthropic 模型少 29%（173,000 个令牌），价格几乎与 Argon 持平（7.68 美元）。
Artificial Analysis 的计算基于 4 美元/百万输入令牌和 20 美元/百万输出令牌（缓存输入令牌为 0.20 美元）的常规列表价格，忽略了截至 2026 年底仍在进行的当前 50% 的入门折扣。
查看原文
Google Deepmind shipped Gemini 4 Carbon just two weeks after Argon, and the new model takes the top spot on the Artificial Analysis Intelligence Index.
Carbon scores 59 points on the recently updated version 5.0 of the Artificial Analysis Intelligence Index, while predecessor Argon trails at 50. Beyond general gains in agentic tasks and programming, Google highlights capabilities spanning the full cyber defense lifecycle, from detecting and validating vulnerabilities to patching them.
Carbon leads the overall AA Index but still falls slightly short of Opus 5.5 in Claude Code for agentic coding. | Image: AA
Built for multi-day coding sessions
Google Deepmind says it designed Carbon for "much harder, long-horizon engineering problems" that demand persistent agentic workflows. Internal teams ran the model for multiple days straight, during which it worked through dead ends on its own and handled complex, multi-step tasks like root-cause analysis and subtle bug fixes in large codebases and low-level software, Google claims.
Google also reports improvements in scientific reasoning, multimodal understanding, and full-stack web development based on user feedback. Higher-quality outputs and artifacts in web development pushed Carbon up 89 Elo points on the WebDev Arena, where it sits in fourth place at 1,769.
In Google's own benchmark selection, Gemini 4 Carbon takes first place in 15 of 16 tests, pulling away from the field especially in long-context and security tasks. | Image: Google Deepmind
On Gray Swan's indirect prompt injection benchmark, Carbon pushed the attack success rate from 0.7 percent with Argon down to 0.0 percent, blocking every tested manipulation attempt according to Google. That doesn't mean prompt injection is solved, but the test appears to be saturated.
Independent evaluations back up Google's numbers
According to Artificial Analysis, Carbon at its highest reasoning tier ("high") leads the Intelligence Index v5 mainly because of top scores in complex reasoning, knowledge retrieval, and scientific tasks, taking first place in four of the index's nine individual evaluations.
In agentic coding on Antigravity, Carbon still trails Claude Code with Opus 5.5 by three points, despite jumping from 53 to 59 compared to Argon. In agentic knowledge work on AA-Briefcase, Carbon (high) gains 126 points over Argon (high, 1,483 Elo) to reach 1,609 Elo, driven by presentation quality (+293 Elo to 1,598) and analytical quality (+120 Elo to 1,671). It still falls behind Kimi K3.1 Pro (1,678 Elo), Muse Wonder (max, 1,659 Elo), and Grok 4.7 (xhigh, 1,638 Elo).
Answer accuracy rises from 48 to 58 percent compared to Argon, while the hallucination rate climbs from four to 11 percent. That's higher but still the second-lowest in the ranking. Claude Opus 5.5 (max) and GPT-6 Astra (max) each sit at 50 percent.
The evaluation platform Vals AI puts Carbon in second place at 70.36 percent accuracy, behind Claude Sonnet 5.5 (70.85 percent) and ahead of Claude Opus 5.5 (70.18 percent) and Argon in fourth (69.24 percent), all at lower cost.
In the Vals Index v2.2, Gemini 4 Carbon slots between Claude Sonnet 5.5 (70.85 percent) and Claude Opus 5.5 (70.18 percent) at 70.36 percent accuracy. At $10.51 per test and just under 48 minutes of runtime, it runs cheaper and faster than both Anthropic models. | Image: Vals AI
On Vals.ai, Carbon also leads both the Finance Index and the Legal Research benchmark. In legal research, Vals AI recorded roughly 45 percent more cited sources, more frequent tool calls, and longer responses than Argon.
But tasks in the Finance and Tax Agent benchmark take more than twice as long as with Argon, and in Legal Research, Carbon averages more than 50 minutes per task. The model generates 2.3 times as many reasoning tokens and 64 percent more output tokens than Argon, though that's still less than half the output token count of Sonnet 5.5 and Opus 5.5.
On Arena.ai, Gemini 4 Carbon (High) takes the Text Arena lead with 1,528 points, edging past Argon (1,520). In the Agent Arena, the model places second across more than 21,900 real sessions, ahead of GPT-6 Astra and Fable-5.1, with a net improvement of 16.2 percent and the highest steerability score (+18.6 percent). Its median task cost of $1.80 is 72 percent below first-place Claude Opus 5.5 (Max).
Arena.ai places Gemini 4 Carbon (High) second in the Agent Arena at +16.2 percent net improvement, behind Claude Opus 5.5 (Max, +21.1 percent) and ahead of GPT-6 Astra (Max, +13.4 percent). | Image: Arena
In the WebDev category of the Code Arena, Carbon reaches 1,768 points (89 ahead of eleventh-place Argon) for fourth overall, just two behind Claude Sonnet 5.5 (xHigh, 1,770). That places it on the Pareto curve across all three disciplines.
Carbon leads security benchmarks without guardrails
For several security tests, Google ran the model without safety guardrails. On SEC-bench Pro, which measures autonomous vulnerability discovery in real-world software, Carbon scores 88.2 percent and leads all tested models. On ExploitGym, which tests whether a system can turn discovered vulnerabilities into working exploits, the model reaches 54.8 percent. Carbon also holds a wide first-place lead in writing fixes on CWE-Bench.
On SEC-bench Pro (88.2 percent), CWE-bench (71.0 percent), and ExploitGym (54.8 percent at ten hours of runtime), Gemini 4 Carbon leads both GPT-6 Astra and predecessor Argon. | Image: Google Deepmind
Google also cites internal results that can't be independently verified. On in-house benchmarks, Carbon reportedly produces secure code more than twice as often as Argon, and in tests by Google's Chrome team, it performed more than ten times better at writing security patches than the frontier model of an unnamed competitor.
Security experts get access first
Google is rolling out the model gradually and participating in the U.S. government's pre-release testing process, citing Carbon's advanced security capabilities. Selected cybersecurity experts get access first starting this week. Through Google Deepmind's Fairwind Program, access will expand to more trusted partners after that.
The company plans to open the model to developers, businesses, and consumers through its product interfaces after collecting more feedback and adjusting the safety guardrails.
Gemini 4 Carbon launches at the same introductory price as Argon: $2 per million input tokens and $10 per million output tokens, with cached input tokens getting a 95 percent discount off the regular input price.
At nearly the same per-task cost as Argon ($7.76 vs. $7.68), Gemini 4 Carbon climbs nine points to 59 and approaches the top of the Pareto curve, also surpassing the more expensive Claude Opus 5.5 ($10.04). | Image: Artificial Analysis
Carbon costs $7.76 per task in the Artificial Analysis Intelligence Index, about 23 percent less than Claude Opus 5.5 (max, $10.04) and roughly 1.2 times more than GPT-6 Astra (max, $6.25). It averages 123,000 output tokens per task, 29 percent fewer than the Anthropic model (173,000 tokens), and lands almost exactly at Argon's price point ($7.68).
Artificial Analysis bases its calculations on the regular list price of $4 per million input tokens and $20 per million output tokens (plus $0.20 for cached input tokens), ignoring the current 50 percent introductory discount that runs through the end of 2026.
原文版权归 The Decoder 及原作者所有，本站转载仅供阅读。如需删除，请联系 support@aisa.one。
Studio