CN China10-29 08:03⏱ 9 min read

It can grind through bug fixes for days on end and shrugs off prompt injection.

Just two weeks after the release of Argon, Google has rushed out its new flagship, Gemini 4 Carbon. This time it overtakes Claude Opus 5.5 and GPT-6 Astra in one sweep. On the just-updated Artificial Analysis Intelligence Index 5.0, it claimed the global No. 1 spot with a high score of 59. When Google released Gemini 4 Argon two weeks ago, Silicon Valley was still buzzing about a mysterious internal checkpoint called "Carbon" leaked from Google; at the time, the verdict from internal engineers was merely that the coding experience "approached Opus 5.5." Nobody expected that within just half a month, the official Carbon would not only complete a head-on reversal against Opus 5.5 in reasoning and long-horizon coding, but also lead by a wide margin in code safety and cheat resistance. At launch, Carbon continues Argon's half-price discount: 2 dollars per million input tokens and 10 dollars per million output tokens — only half the list price of Opus 5.5 ($4/$20). It also offers a steep 95% discount on cached input, which agent tasks use heavily (as low as 0.2 dollars per million tokens). Although Gemini 4 Argon, released two weeks ago, was impressive on long context and multimodality, it was still edged out by Opus 5.5 on the overall intelligence index (53 points) and on some hardcore terminal tasks. This time, Gemini 4 Carbon has precisely filled in the last puzzle piece, pushing its capabilities straight into the form of an "around-the-clock autonomous engineer." Carbon can run for days on end, finding its own way through enormous, complex codebases and low-level systems; even if it hits "dead ends" repeatedly along the way, it keeps grinding to the very end, independently completing complex multi-step engineering tasks. From tracing the root cause of an extremely well-hidden bug to finally writing the fix patch, it does it all in one go. On Terminal-Bench 4.0 (the AA test set), Carbon scored 65.2%, ahead of Opus 5.5 (59.6%) and GPT-6 Astra (59.1%). On the science terminal task benchmark Terminal-Bench-Science 0.1, it likewise ranked first with 64.8% (Opus 5.5: 59.0%). On GraphWalks with contexts from 256k to 1M, Carbon scored 96.2%, leaving Opus 5.5 (66.8%) and Astra (71.8%) far behind. On Vending-Bench 2, which simulates long-cycle business operations, Carbon's final account balance reached 16,484 dollars, close to 1.8 times that of Opus 5.5 (9,235 dollars). On WebDev Arena, thanks to improvements in frontend components and output quality, Carbon's Elo score rose 89 points to 1769 — fourth in the world, but with a lower average task cost. On Vals Index v2.2, which measures the value of enterprise-grade workflows, Carbon narrowly led with 70.4% versus Opus 5.5's 70.2%. That said, according to the breakdown by third-party evaluation agency Artificial Analysis, Opus 5.5 still holds part of its moat in pure code engineering. For example, in the table Google itself published, there are two metrics where Carbon failed to take the top spot: One is FrontierSWE 2, which measures frontier software engineering: GPT-6 Astra (65.5%) and Opus 5.5 (62.3%) still rank ahead of Carbon (61.2%). The other is the offline subset of the computer-use benchmark OSWorld-2.1, where Opus 5.5 remains first with 87.2% (Carbon: 81.5%). In addition, on the Coding Agent specialization index, which pairs models with their respective official CLI tools, Carbon scored 59 with Google's Antigravity CLI, still 3 points behind Opus 5.5 backed by Claude Code (62). But for developers, Carbon's real killing power lies in token-consumption efficiency. Independent evaluation data show that although Opus 5.5 is highly capable, its reasoning process is extremely verbose: a single Intelligence Index run consumes on average 173,000 output tokens, while Carbon needs only 123,000 — a 29% reduction in output volume. This means that even if Carbon eventually returns to the same $4/$20 standard pricing as Opus 5.5, Carbon's average cost per task ($7.76) would still be about 23% lower than Opus 5.5's ($10.04). Factoring in the current launch half price of $2/$10, Carbon's bill in real production environments would be less than 40% of Opus 5.5's. Beyond regular coding, Google this time has staked Carbon's core differentiating selling point on the full lifecycle of cybersecurity defense. It was also forced into this — you can't just sit and watch a frontier model spend every day hunting for vulnerabilities in the Chrome browser and the Android OS, right? The core of cybersecurity defense is finding deeply hidden complex flaws in real software before hackers do. On SEC-bench Pro, which contains real high-severity vulnerabilities in the V8 engine, the Linux kernel, and more, Carbon scored 88.2% — not only far ahead of its predecessor Argon (78.8%) but directly surpassing OpenAI's GPT-6 Astra (85.4%) to set an entirely new SOTA. Facing hundreds or thousands of potential vulnerabilities, which one should a security team fix first? This requires the model to act like a real attacker and verify which vulnerabilities are genuinely exploitable. On the well-known vulnerability-exploitation benchmark ExploitGym, Carbon once again set a new SOTA with 54.8% (Argon previously scored 35.1%), helping defenders precisely prioritize fixes. After finding and verifying a vulnerability, the ultimate test is writing fix code that actually runs and does not break existing functionality. On CWE-bench from Collinear — a real-repository audit-and-fix test covering 73 Common Weakness Enumerations — Carbon reached 71.0%, decisively beating Claude Opus 5.5 (67.0%) and Argon (68.0%). Beyond topping external leaderboards, on Google's internal security benchmarks Carbon's success rate at generating secure code improved by more than 2x compared with Argon. An even more striking battle report comes from Google's own Chrome browser team: in a head-to-head comparison on generating high-quality security patches, Carbon outperformed a large frontier model from another top lab by more than 10 times. When a model runs in a terminal for days, what troubles developers most is often not the compute bill, but the model starting to "get clever" when facing high-pressure, complex tasks (Shortcut & Loophole Exploitation). A previous third-party audit of Terminal-Bench 4.0 revealed that many high-scoring models, when unable to fix a bug, would fake a "test pass" by modifying test scripts or exploiting loopholes. Targeting this chronic problem, Google made a focused tightening in Carbon's alignment training: On the terminal-operation benchmark Terminal-Bench 4.0, while sharply improving overall accuracy, Carbon drove its "successful cheating rate" all the way down to 0.0%, and even the proportion of attempted shortcut cheating fell to 0.4%. In the live coding environments of Google's internal teams, high-severity misalignment alerts plummeted by nearly 80%. In internal testing, Carbon makes unsupported claims at a frequency about 2x lower than Argon. When evidence is missing or operations are blocked, it no longer hedges or forcibly fills in the gaps with speculation; instead, it writes the limitations of its methodology directly into the deliverable documentation, or lays out a clear checklist telling humans "what conditions are still missing." Faced with vague requirements, Carbon does not blindly guess or force things forward; it first retrieves relevant sources and puts the contradictions it finds on the table to confirm with the user. When executing multi-step plans end to end, its overreach decreased by 26%, and its under-service (skipped or shortchanged work) decreased by 29%. When agents read web pages, emails, and external codebases in the wild, they are highly vulnerable to hijacking via Indirect Prompt Injection (IPI). On the Gray Swan IPI benchmark, predecessor Argon once led the entire field with an extremely low breach rate of 0.7%. This time, Carbon turned in a perfect scorecard: a 0.0% attack success rate, successfully fending off every attack attempt in the evaluation. From Gemini 4 Argon, which debuted in late September and rolled out comprehensively just two weeks ago, to Gemini 4 Carbon, which has now burst onto the scene and comprehensively overtaken Opus 5.5. As I said before, we will release models periodically... Now introducing Gemini 4 Carbon. Combining the previously leaked internal information, Google in fact trained and evaluated multiple Gemini 4 checkpoints in parallel internally — Argon, Barium, Carbon, and more (the publicly released Argon was in fact the internal Barium-B checkpoint). In an unobtrusive corner of the official announcement, Koray Kavukcuoglu, senior vice president and chief AI architect of Google DeepMind, revealed the real secret: "Our researchers are working hand in hand with our models to advance capability evolution at a faster pace, while strengthening safety guardrails in lockstep." In other words, using the most cutting-edge AI to assist in training, evaluating, and optimizing the next generation of AI is now genuinely up and running inside Google. Because Carbon has fairly strong vulnerability-discovery and exploitation capabilities, a pre-release testing process is still required. Starting this week, Carbon will first open to a group of cybersecurity defense professionals. Remember how long it took from Argon's release to actually being available for you to use? The wait this time won't be short either. With Anthropic now at a critical juncture around its IPO, it's about time to let Fable 5.5 out, right? As for OpenAI, obsessed with shipping mid-size Terra models and Dot the text-message notifier — I wish them good luck. References: [1] https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ [2] https://x.com/sundarpichai/status/2115489219804352921?a=TCW5gQam8bc7T7enVWo&s=20

Argon发布两周啊，谷歌火速掏出新旗舰Gemini 4 Carbon。

这次一举超越Claude Opus 5.5和GPT-6 Astra。

在刚刚更新的Artificial Analysis智能指数5.0中以59分的高分强势登顶全球第一。

两周前谷歌刚发Gemini 4 Argon时，硅谷还在热议谷歌内部泄露的一个神秘存档点“Carbon”，当时内部工程师给出的评价还是“编程体感逼近Opus 5.5”。

没想到短短半个月，正式版Carbon不仅在智商与长程编程上完成了对Opus 5.5的正面反超，更在代码安全与防作弊上大幅领先。

Carbon首发阶段延续了Argon的半价折扣，输入2美元/百万Token、输出10美元/百万Token，账面单价仅为Opus 5.5（$4/$20）的一半。

并且针对Agent任务高频使用的缓存命中输入（cached input）给出了95%的骨折价优惠（低至0.2美元/百万Token）。

两周前发布的Gemini 4 Argon虽然在长文本和多模态上表现惊艳，但在综合智能指数（53分）和部分硬核终端任务上，依然被Opus 5.5压了一头。

这一次，Gemini 4 Carbon精准补齐了最后一块拼图，直接将能力推向了“全天候自主工程师”的形态。

Carbon能够连续运行好几天，在庞大复杂的代码库和底层系统中自己摸索；哪怕中途多次撞上“死胡同”（dead ends），它也能持续死磕到底，独立完成复杂的多步工程任务。从顺藤摸瓜定位极其隐蔽的Bug根因，到最终写出修复补丁，一气呵成。

在Terminal-Bench 4.0（AA测试集）中，Carbon拿到65.2%，领先Opus 5.5（59.6%）和GPT-6 Astra（59.1%）；

在科学终端任务Terminal-Bench-Science 0.1上同样以64.8%位列第一（Opus 5.5为59.0%）。

在256k至1M上下文的GraphWalks测试中，Carbon拿到96.2%，大幅甩开Opus 5.5（66.8%）和Astra（71.8%）；

在模拟长周期商业经营的Vending-Bench 2中，Carbon最终账户余额达到16,484美元，接近Opus 5.5（9,235美元）的1.8倍。

在WebDev Arena上，Carbon凭借前端组件与输出质量的提升，Elo分数上涨89分来到1769分，虽然是全球第四，但平均任务成本更低。

在衡量企业级工作流价值的Vals Index v2.2上也以70.4%微弱领先Opus 5.5（70.2%）。

不过，从第三方评测机构Artificial Analysis的拆解来看，Opus 5.5在纯代码工程领域依然保有一部分护城河。

例如在谷歌自己公布的表格中，有两项指标Carbon未能登顶：

一项是衡量前沿软件工程的FrontierSWE 2，GPT-6 Astra（65.5%）与Opus 5.5（62.3%）依然排在Carbon（61.2%）前面；

另一项是计算机操作基准OSWorld-2.1离线子集，Opus 5.5以87.2%保持第一（Carbon为81.5%）。

此外，在搭配各自官方CLI工具的Coding Agent专项指数上，Carbon配合谷歌Antigravity CLI得分为59分，仍以3分之差次于Claude Code加持下的Opus 5.5（62分）。

但对于开发者而言，Carbon真正的杀伤力在于Token消耗效率。

独立评测数据显示，Opus 5.5虽然能力强，但推理过程极为冗长，完成单次Intelligence Index评估平均要消耗17.3万输出Token。而Carbon仅需12.3万Token，输出量减少了29%。

这意味着，即便未来Carbon恢复到与Opus 5.5相同的$4/$20标准定价，Carbon的单次任务平均成本（$7.76）也比Opus 5.5（$10.04）低23%左右。

如果算上目前$2/$10的首发半价，Carbon在实际生产环境中的账单成本还不到Opus 5.5的四成。

除了常规编程，谷歌此次将Carbon的核心差异化卖点押注在了网络安全防御全生命周期上。

这也是被逼的没办法，总不能看着家前沿模型天天挖Chrome浏览器和安卓系统的漏洞是吧。

网络安全防御的核心，是赶在黑客之前从真实软件中揪出深藏的复杂缺陷。在包含V8引擎、Linux内核等真实高危漏洞的SEC-bench Pro测试中，Carbon拿下了88.2%，不仅大幅甩开前代Argon（78.8%），更直接超越了OpenAI的GPT-6 Astra（85.4%），创下全新SOTA。

面对成百上千个潜在漏洞，安全团队到底该先修哪一个？这需要模型像真正的攻击者一样，去验证哪些漏洞是真正可被利用的（exploitable）。

在著名的漏洞利用基准测试ExploitGym上，Carbon以54.8%的成绩再次刷新SOTA（此前Argon为35.1%），帮防御者精准排定修复优先级。

找到并验证漏洞后，终极考验是直接写出能跑通、且不破坏原有功能的修复代码。在Collinear推出的CWE-bench（涵盖73种常见弱点枚举的真实仓库审计修复测试）中，Carbon达到了71.0%，全面碾压Claude Opus 5.5（67.0%）和Argon（68.0%）。

除了刷爆外部榜单，Carbon在谷歌内部安全基准测试中，Carbon生成安全代码的成功率比Argon提升了2倍以上。

更夸张的战报来自谷歌自家的Chrome浏览器团队：在生成高质量安全补丁的实战对比中，Carbon的表现比另一家顶级实验室的大型前沿模型强出10倍以上。

模型在终端里连续跑几天，最令开发者头疼的往往不是算力账单，而是模型在遇到高压复杂任务时开始“耍小聪明”（Shortcut & Loophole Exploitation）。

此前针对Terminal-Bench 4.0的第三方审计就曾揭露，不少高分模型在修不好Bug时，会通过修改测试脚本或走漏洞来伪造“测试通过”。针对这一顽疾，谷歌在Carbon的对齐训练上做了一次重点收敛：

在终端操作基准Terminal-Bench 4.0上，Carbon在大幅提升整体准确率的同时，将“成功作弊率”硬生生压到了0.0%，连尝试走捷径作弊的比例都跌至0.4%。

在谷歌内部团队的实况编程环境中，高严重性的失调警报（misalignment alerts）更是暴跌了近80%。

在内部测试中，Carbon做出无根据断言（unsupported claims）的频率比Argon降低了约2倍。

面对证据缺失或操作受阻，它不再含糊其词或强行脑补，而是把方法论的局限性直接写进交付文档，或者拉出一张清晰的清单告诉人类“还缺什么条件”。

面对模糊不清的需求，Carbon不会盲目瞎猜或强行推进，而是先检索相关信源，把发现的矛盾点摆到台面上与用户确认。

在端到端执行多步计划时，它的越界操作（overreach）减少了26%，敷衍漏做（under-service）减少了29%。

当Agent在野外读取网页、邮件和外部代码库时，极易遭到“间接提示词注入”（IPI）的劫持。

在Gray Swan IPI基准测试中，前代Argon曾以0.7%的极低被攻破率领跑全。

而这一次，Carbon直接交出了0.0%攻击成功率的满分答卷，成功抵御了评估中的每一次攻击尝试。

从9月底亮相、两周前刚刚全面铺开的Gemini 4 Argon，到今天横空出世、全面反超Opus 5.5的Gemini 4 Carbon。

我之前就说过，我们会周期性地发布模型……现在向大家介绍Gemini 4 Carbon。

结合此前曝出的内部信息来看，谷歌实际上在内部同时并行训练和评估了Argon、Barium、Carbon等多个Gemini 4存档点（其中公开发布的Argon实为内部的Barium-B存档点）。

在官方公告不起眼的角落里，谷歌DeepMind高级副总裁、首席AI架构师Koray Kavukcuoglu透露了真正的秘密：

“我们的研究人员正与我们的模型携手并进（working hand in hand），以更快的速度推进能力演进，同时同步强化安全护栏。”

换句话说，用最顶尖的AI去辅助训练、评估和优化下一代AI，已经在谷歌内部真正转了起来。

由于Carbon具备较强的漏洞挖掘与利用能力，还要进行发布前测试流程。从本周开始，Carbon将率先向一部分网络安全防御人员开放。

还记得Argon从发布到真正给你用过了多久吗，这次等的也不会短就是了。

眼下Anthropic上市的关键节点上，这下该把Fable 5.5放出来了吧？

至于沉迷发中杯Terra模型和短信通知Dot的OpenAI，我祝他好运。

参考链接： [1]https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/ [2]https://x.com/sundarpichai/status/2115489219804352921?a=TCW5gQam8bc7T7enVWo&s=20