在早期的独立测试中，谷歌的 Gemini 4 Argon 与 OpenAI 的 GPT-6 Astra 旗鼓相当。据 Artificial Analysis 称，该模型在人工智能指数（Artificial Analysis Intelligence Index）上得分为 53 分，与 GPT-6 Astra 持平，比 GPT-6.1 Sol 高出 1 分。Anthropic 的 Claude Opus 5.5 仍以 58 分领跑该指数。
Google’s Gemini 4 Argon is level with OpenAI’s GPT-6 Astra in early independent tests. It scored 53 on the Artificial Analysis Intelligence Index, the same as GPT-6 Astra and one point ahead of GPT-6.1 Sol, Artificial Analysis said. Anthropic’s Claude Opus 5.5 still leads the index, with 58.
这一得分比谷歌上一款高于 Flash 级别的模型 Gemini 3.1 Pro Preview 高出 23 分。这家基准测试公司表示，这使谷歌重新回到了三大 AI 实验室之列。谷歌于周三发布了 Argon，目前仅向一小部分网络防御者开放。
The score is 23 points above Gemini 3.1 Pro Preview, Google’s last model above the Flash class. The benchmarking firm said it puts Google back among the top three AI labs. Google unveiled Argon on Wednesday and has so far given it only to a small group of cyber defenders.
幻觉更少，Token 使用更多。在测试事实知识的 AA-Omniscience 指数中，Argon 的幻觉率为 15%。这是该指数中得分 45 分及以上的模型中最低的。GPT-6 Astra 的幻觉率为 51%，GPT-6.1 Sol 为 54%。该公司表示，Argon 在遇到不知道的问题时，更多选择承认而非胡乱猜测。不过，它的正确率也较低：为 50%，而 GPT-6 Astra 为 63%。
Fewer hallucinations, more tokens Argon had a hallucination rate of 15% on AA-Omniscience, a test of factual knowledge. That is the lowest of any model scoring 45 or more on the index. GPT-6 Astra came in at 51% and GPT-6.1 Sol at 54%. Argon more often admits it does not know an answer instead of guessing, the firm said. It also got fewer answers right: 50%, against 63% for GPT-6 Astra.
在测试商业软件工作流的 AutomationBench-AA 中，Argon 以 78% 的得分排名第一，比 Claude Sonnet 5.5 高出 7 个百分点。在编码测试 Terminal Bench 4 中，它得分为 57%。Claude Sonnet 5.5（64%）、Claude Opus 5.5（60%）和 GPT-6 Astra（59%）的得分均高于该模型。
Argon ranked first on AutomationBench-AA, a test of business software workflows, with 78%. That is seven points ahead of Claude Sonnet 5.5. On Terminal Bench 4, a coding test, it scored 57%. Claude Sonnet 5.5 (64%), Claude Opus 5.5 (60%) and GPT-6 Astra (59%) all scored higher.
在发布折扣期间，使用 Argon 完成每项指数任务的成本为 1.99 美元，是 GPT-6 Astra 3.26 美元成本的 60%。这种节省源于较低的 Token 价格，而非使用了更少的 Token。Argon 每项任务使用了约 62,000 个输出 Token，而 GPT-6 Astra 为 27,000 个。其每项任务的成本仍是 GPT-6.1 Sol 的 2.7 倍。
Each index task cost $1.99 with Argon at its launch discount, 60% of GPT-6 Astra’s $3.26. The saving comes from lower token prices, not from using fewer tokens. Argon used about 62,000 output tokens per task, against 27,000 for GPT-6 Astra. It still costs 2.7 times as much per task as GPT-6.1 Sol.
谷歌至少在一个月内提供半价优惠：每百万输入 Token 2 美元，每百万输出 Token 10 美元。据 Artificial Analysis 称，按 4 美元和 20 美元的标准价格计算，每项任务的成本将升至 3.98 美元，约为 GPT-6 Astra 的 1.2 倍。谷歌尚未说明折扣何时结束。
Google is charging half price for at least a month: $2 per million input tokens and $10 per million output tokens. At the standard $4 and $20, each task rises to $3.98, about 1.2 times GPT-6 Astra, according to Artificial Analysis. Google has not said when the discount ends.
在Arena文本评测榜单上，Argon同样位居榜首。该榜单由用户对各模型生成的答案进行投票评选。据The Decoder报道，Argon获得了1,525分，比Claude Opus 4.6高出20分。在Arena的Web开发能力评测中，它则排在第8位。此外，它还是首个在Vals Index评分中名列第一的Gemini系列模型，得分达68.9%。
First on the Arena text ranking Argon also leads Arena’s Text Arena, where people vote between model answers. It has 1,525 points, 20 ahead of Claude Opus 4.6, The Decoder reported. It placed eighth in Arena’s WebDev ranking. It is also the first Gemini model to top the Vals Index, at 68.9%.
谷歌内部存在疑虑：部分谷歌员工对该模型信心不足。据知情人士透露，该模型的编程能力表现参差不齐，在前端设计方面也并不出色——而前端设计直接决定了应用程序与网站的外观效果。彭博社记者Julia Love与Davey Alba援引内部评估信息报道了上述情况。
Doubts inside Google Some Google staff are less sure of the model. Its coding abilities are uneven, and it is not particularly good at front-end design, one person said. Front-end design shapes how apps and websites look. Julia Love and Davey Alba reported this for Bloomberg, citing people familiar with the internal evaluations.
据彭博社称，Argon属于超大规模模型，而这类模型的运行成本通常极高。有两位知情人士指出，该模型似乎受到了“过度追求测试分数”现象的影响——即工程师们一味追求高分而忽视了产品实际功能的完善。
Argon is also a very large model, according to Bloomberg, and large models are typically expensive to run. Two people said it appears affected by “benchmaxxing”, when engineers chase a test score over a product that does the job well.
部分员工认为Anthropic的Fable模型与OpenAI的Astra模型发展速度更快；也有人认为Gemini 4已经迎头赶上。针对外界关于该模型在编程等领域表现不佳的质疑，谷歌向彭博社回应称这种说法并不准确。谷歌还引用了谷歌DeepMind负责人Koray Kavukcuoglu在The Information主办的会议上发表的言论：“在我看来，我们始终处于技术前沿，这是毋庸置疑的。”
Some employees believe Anthropic’s Fable and OpenAI’s Astra models are improving faster than Gemini. Others believe Gemini 4 has caught up. Google told Bloomberg it would be inaccurate to say the model underperforms in areas such as coding. It pointed to remarks by Koray Kavukcuoglu, who runs Google DeepMind, at a conference hosted by The Information. “In my mind, it’s a certainty that we are always gonna be at the frontier,” Kavukcuoglu said.
Kavukcuoglu还透露，谷歌原本计划让Gemini 4在年底之前发布。目前，付费API用户以及Google AI Ultra服务的使用者将率先体验到Argon模型，但具体发布日期尚未公布。
Kavukcuoglu had said Google wanted Gemini 4 out much earlier than year-end. Paid API customers and Google AI Ultrars are next in line for Argon. Google has not given a date.