刚刚 Gemini 4 发布!性价比「斩杀」GPT-6 和 Opus 5.5 重回御三家Gemini 4 has just been released! Cost-effective "kill" GPT-6 and Opus 5.5 return to the Imperial family


AI free token/AI免费积分

接连更新 Flash 3.5 至 3.8 系列后,迟迟未兑现旗舰 Pro 版本的谷歌,急需一款重磅机型稳住市面预期。两周前悄然启动小范围测试的 Gemini 4 Argon,便承担着这一角色。

Argon 意为氩气,无色无味且性质稳定,常在焊接与半导体制造等关键工艺中给出保护屏障。以此命名,恰好呼应了谷歌当下的需要:不要更多噱头,而是用一次扎实的发布稳住局面。

Image

图|官方博客地址:https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/

就在刚刚,潜心打磨六个月、历经超过 10 万名谷歌内部使用者实战测试的 Gemini 4 Argon 正式面向全球上线。

谷歌 AI 机型负责人 Logan Kilpatrick 在社交平台上表示,这仅仅是 Gemini 4 时代的开端,新模型的外观初衷就是为了帮人们攻克以前想都不敢想的棘手难题。

Image

被催了半年的谷歌,把新旗舰端上来了

从定位上看,Gemini 4 Argon 把重心完全放在了复杂、多步骤的真实世界深度任务上,覆盖编程、金融、法律、网络安全与前沿科学等领域。在衡量关键经济领域表现的 Vals Index 榜单中,Argon 直接拿下了第一名。

为了支撑高强度的长程任务,谷歌此次在 API 中给出了高达 100 万的统一输入输出 Token 上限,并配套推出了全新的续写协议。

相比于高昂的算力消耗预期,Argon 的首发优惠期定价显得相当激进,每百万输入 Token 仅售 2 美元,每百万输出 Token 为 10 美元,缓存输入的折扣更是高达 95%。

Image

当下,开发者已可借助 AI Studio、Android Studio 以及 Gemini API 接入该模型,Google AI Pro 与 Ultra 订阅使用者也能在 Gemini 应用、搜索 AI 模式以及智能体开发平台 Google Antigravity 中直接使用。

在推向市面之前,谷歌内部工程团队已经把 Argon 当作了主力生产力工具。

过去几周内,借助 Antigravity 平台上的 Argon 模型,谷歌工程师的智能体代码提交量增长了 35% 以上,代码回滚率则下降了近三分之一。

在代码与模型优化上,Argon 展现出极高的自主性。它协助 Chrome 团队在几周内重构了 4000 万行代码库中的 600 万行遗留代码,彻底消除了一类浏览器崩溃难题。

Image

在仅获知简短目标和规则的情况下,它夜间自主优化了谷歌智能体框架 ADK,将长会话内存准备时间缩短 40%;还在几天内将开源模型 Gemma 4 E2B 的端到端推理速度提高 2.85 倍,且保持零质量损失。

这种长程推理能力很快延伸至商业、科研与物理世界。

早期实测中,德勤测出其长周期跨工具任务完成率达 100%,Kepler 发现其生成的多年期财报与人类分析师模型吻合度超 99%,法律科技公司 MinutaIA 甚至用它揪出了人类专家遗漏的案情细节。

科研维度,它不仅大幅加速了量子化学计算与量子纠错解码,还在几小时内完成五种新型酶外观,并协助找出了核聚变「格拉德猜想」的新反例。

Image

Image

在具身智能测试中,它仅靠观看人类玩沙包的视频,便在几小时内自主算出关节轨迹并写出仿真代码,直接部署到 Apollo 人形机器人上达成了真机连中七投。谷歌同时宣布向全球万名科学家免费开放半年 Antigravity 最高使用权限。

伴随能力扩张,谷歌为其部署了最严密的安全栈。

借助引入推理步骤实时监测器,模型在编程测试中走未授权捷径的比例降至 2%,间接提示词注入攻击成功率压低至 0.7%,并已配合安全公司 Wiz 协助联邦机构与医疗平台提前修复了多个高危漏洞。

顶级 AI 多一个挑选,使用者就少一分将就

随着模型权限的全面放开,全球开发者与科技博主在第一时间涌入平台展开实测,3D 空间构建、物理仿真与运行效率成为了首批测试的焦点。

在官方演示中,Argon 能用纯 Python 从零编写 3D 光追引擎,或将办公桌照片变为微缩赛车游戏。这种空间推理与编码能力在网友实测中得到了直观印证。

博主 Bhavy 称其为一头猛兽,展示了一个无外部 3D 资产和游戏引擎、纯靠数学代码在单浏览器文件中构建的阳光书房,光影反射与窗外山脉细节逼真;

Image

Image

在「3D 鲸鱼跃出水面」的四方盲测中,Argon 的照片级真实感也与 Claude Opus 5.5、GPT-6 Astra 与 Kimi K3 不分伯仲。

Image

交互生成同样细腻,使用者 Ashutosh 仅用几次提示词便做出了带逼真水纹与物理操控感的海岛帆船游戏;

Image

开发者 Adit_Yah 测试的微缩海岛树脂块,仅耗时 13 分钟、消耗 2% 周配额,就完成了包含焦散、波浪与果冻物理成效的可调节界面。

Image

在复杂综合任务拆解上,投资人 Justine Moore 在未给出任何图文数据的情况下,让 Argon 制作一个互动式苹果电脑博物馆。

Image

结果模型自主抓取了档案中的 Mac 照片,自己写工具清理图像,拉取跨越 50 年的硬件规格数据,最终构建出一个 UI 会随年代滑动自动变形的 3D 封面流网页。

Image

笔者也第一时间对 Argon 的复杂代码构建与中文创意写作开展了实测。

在代码测试中,要求它用 Three.js 和 TypeScript 制作一个包含逻辑分层、连招判定与完整 HUD 的东方纸本风古代战场「无双式」动作游戏原型,最终产出的项目完成度颇高,实战画面与交互成效相当不错。

Image

而在描写暴雨前菜市面感官氛围以及书店橘猫视角等中文写作测试里,新模型延续了前代 Flash 系列一如既往的稳定水准,且文风有了显著进步,过去喜欢堆砌宏大词汇的习惯收敛了不少。

Image

Image

左右滑动查看图片

与其他前沿大模型的正面交锋,更是将 Argon 的速度与售价长处展现得淋漓尽致。

AIHubmix 团队借助 API 使用完全相同的提示词,对 Gemini 4 Argon 和 GPT-6 Astra 开展了像素风古建筑夜景的生成测试。结果表明,GPT-6 Astra 耗时 25 分钟,花费高达 11 美元,而 Argon 仅仅用了 3 分钟就完成了任务,费用更是低至 0.50 美元,且画面的色彩层次与光影细节毫不逊色。

Image

Image

在与 Claude Opus 5.5 的多项比较中,Argon 也展现出了极高的性价比。

博主 Rubens Soto 让两款模型执行同一任务,即制作一个电影级的 3D 动画来解释特斯拉式电动机的工作原理。在这场测试中,Argon 的单次生成成本为 3.83 美元,远低于 Opus 5.5 的 16.36 美元。

Image

Image

另一位测评者 Pranav Reddy 则使用 Three.js 测试了 3D 人形机器人的格斗与动作仿真,从生成的交互界面完整度与关节动作流畅度来看,Argon 在多个关键指标上超越了 Opus 5.5。

Image

Image

不过,初代上线的 Argon 并非在所有情形下都完美无缺。

Image

测试机构 AI/ML API 在开展「哥斯拉大战金刚」的 3D 情形生成比较时指出,虽然 Argon 的调用成本(4.22 美元)远低于 GPT-6 Astra(23.82 美元),但在特定复杂生物模型的画质与形态还原上仍有显著的提高空间。

从官方披露的硬核工程数据到全网铺开的真实测评,可以看出谷歌在 Gemini 4 世代的打法已经发生了显著转变。

不再执着于单纯的聊天对话技巧,而是借助压低推理成本、拉长上下文窗口、强化代码与空间物理推理,将大模型真正塞进工业级代码库、科学实验室与生产管线之中。

考虑到当下 Claude 封号严重、ChatGPT 频繁降智流口水的背景下,能力全面且 API 售价亲民的 Gemini 4 Argon 无疑是一场及时雨,很大程度上缓解了日常生产力工具的焦虑,也实打实地给友商们上了强度。

只是面对高强度推理带来的算力消耗,谷歌这个「地主家」似乎也没有太多余粮了。

Image

图|🔗 https://support.google.com/gemini/answer/17004136

依据官方最新生效的访问规则,未付费使用者当下已默认交由平台自动分配模型,绝大多数请求由最轻量的 Flash-Lite 处理,一旦关闭自动挑选更是只能固定使用 Flash-Lite。

与此同时,Gemini 应用已全面改用每 5 小时刷新一次、并与每周上限挂钩的动态算力限额机制,调高思考等级、使用 Deep Research 或调用旗舰模型都会加速消耗额度,想要痛快感受满血版能力,终究还是绕不开按阶梯付费的门槛。

但无论如何,正如氩气在工业制造中扮演的角色那样,Gemini 4 Argon 的落地,终于让谷歌在顶尖模型的牌桌上重新筑起了一道稳固的防线。

本文来源:https://www.ifanr.com/1684241
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

After successively updating the Flash 3.5 to 3.8 series, Google has delayed the realization of the flagship Pro version and urgently needs a heavy model to stabilize market expectations. Gemini 4 Argon, who quietly launched a small-scale test two weeks ago, took on this role.

Argon means argon, which is colorless, odorless and stable, and often gives a protective barrier in key processes such as welding and semiconductor manufacturing. Named after this, it just echoes Google's current need: not more gimmicks, but a solid release to stabilize the situation.

Image

Just now, Gemini 4 Argon, which has been honed for six months and has been field-tested by more than 100,000 Google internal users, was officially launched globally.

Logan Kilpatrick, head of Google AI models, said on the social platform that this is only the beginning of the Gemini 4 era, and the new model is designed to help people solve difficult problems that they could not imagine before.

Image

From the perspective of positioning, Gemini 4 Argon focuses entirely on complex, multi-step real-world deep tasks, covering programming, finance, law, cybersecurity and cutting-edge science. Argon took first place in the Vals Index, which measures performance in key economic sectors.

In order to support high-intensity long-range tasks, Google this time gave a unified input/output token limit of up to 1 million in the API, and launched a new renewal protocol.

Image

Compared to the high computing power consumption expectations, Argon's launch offer period pricing is quite aggressive, only $2 per million input tokens, $10 per million output tokens, and up to 95% discounts on cache inputs.

At present, developers can access the model through AI Studio, Android Studio, and the Gemini API, and Google AI Pro and Ultra subscribers can also use it directly in the Gemini app, search AI mode, and the agent development platform Google Antigravity.

Image

Prior to its launch, Google's in-house engineering team had used Argon as their primary productivity tool.

Thanks to the Argon model on the Antigravity platform, Google engineers have seen a more than 35% increase in code submissions and a nearly one-third decrease in code rollbacks over the past few weeks.

Image

Argon demonstrates a high degree of autonomy in code and model optimization. It helped the Chrome team refactor 6 million legacy lines of code in a repository of 40 million lines of code in a matter of weeks, eliminating a class of browser crashes altogether.

Image

With only brief goals and rules, it autonomously optimized the Google Agent Framework ADK at night, reducing long session memory preparation time by 40%, and improved the end-to-end inference speed of the open source model Gemma 4 E2B by 2.85 times in a few days, with zero quality loss.

This long-range reasoning ability soon extended to the worlds of business, research, and physics.

Image

In early testing, Deloitte measured that its long-cycle cross-tool task completion rate was 100%. Kepler found that its generated multi-year financial reports were more than 99% consistent with the human analyst model. Legal technology company MinutaIA even used it to find out the details of the case missed by human experts.

In the scientific research dimension, it not only greatly accelerated quantum chemical calculations and quantum error correction decoding, but also completed the appearance of five new enzymes in a few hours, and helped to find new counterexamples of the "Glad conjecture" of nuclear fusion.

Image

In the intelligent test, just by watching the video of the human playing with the sandbag, it independently calculates the joint trajectory and writes the simulation code within a few hours, and directly deploys it to the Apollo humanoid robot to achieve a real machine. Google also announced that it is opening up Antigravity to 10,000 scientists around the world for free for six months.

Along with the expansion of capabilities, Google has deployed the most tight security stack for it.

Image

With the introduction of the inference step real-time monitor, the proportion of unauthorized shortcuts taken by the model in programming tests has been reduced to 2%, the success rate of indirect prompt word injection attacks has been reduced to 0.7%, and Wiz, a security company, has assisted federal agencies and medical platforms to fix multiple high-risk vulnerabilities in advance.

With the full liberalization of model permissions, global developers and technology bloggers poured into the platform to carry out practical tests in the first time, and 3D space construction, physical simulation and operational efficiency became the focus of the first batch of tests.

Image

In the official demo, Argon can write a 3D light chase engine from scratch in pure Python, or turn a desk photo into a miniature racing game. This spatial reasoning and coding ability has been intuitively confirmed in the actual measurement of netizens.

Image

Blogger Bhavy calls it a beast, showing a sunshine study built in single-browser files based purely on mathematical code, with no external 3D assets and a game engine, light and shadow reflections and realistic details of the mountains outside the window;

In the "3D Whale Leaping Out of the Water" blind test, Argon's photo-realism is also inseparable from Claude Opus 5.5, GPT-6 Astra and Kimi K3.

Image

The interactive generation is equally delicate, and the user Ashutosh has made an island sailing game with realistic water patterns and physical control in just a few prompt words;

The miniature island resin block tested by developer Adit_Yah took only 13 minutes and consumed 2% of the week's quota to complete an adjustable interface that included the physical effects of caustics, waves, and jelly.

Image

On the complex task of dismantling, investor Justine Moore asked Argon to create an interactive Apple Computer Museum without giving any graphic data.

As a result, the model independently grabbed the Mac photos in the archive, wrote its own tools to clean up the images, pulled the hardware specification data spanning 50 years, and finally built a 3D cover stream page that would automatically deform with the age of the UI.

Image

The author also tested Argon's complex code construction and Chinese creative writing for the first time.

In the code test, it was required to use Three.js and TypeScript to make an Oriental paper-style "twinless" action game prototype with logical layering, continuous stroke judgment and complete HUD. The final project was completed to a high degree, and the actual combat screen and interaction effect were quite good.

Image

In Chinese writing tests such as describing the sensory atmosphere of the stormy appetizer market and the perspective of the bookstore Orange Cat, the new model continued the stable level of the previous Flash series as before, and the style of writing has made significant progress. In the past, the habit of piling up grand vocabulary has converged a lot.

Image

Swipe left and right to see images

Frontal confrontation with other leading-edge models showcases Argon's speed and price advantages.

Image

The AIHubmix team tested the generation of pixelated ancient architectural night scenes for Gemini 4 Argon and GPT-6 Astra using identical prompts via an API. The results showed that the GPT-6 Astra took 25 minutes and cost up to $11, while Argon completed the task in only 3 minutes, the cost was as low as $0.50, and the color level and light and shadow details of the picture were not inferior.

In several comparisons to Claude Opus 5.5, Argon has also shown an excellent price/performance ratio.

Image

Blogger Rubens Soto lets two models perform the same task of creating a cinematic 3D animation that explains how a Tesla electric motor works. In this test, Argon's one-time generation cost was $3.83, much lower than Opus 5.5's $16.36.

Another evaluator, Pranav Reddy, used Three.js to test the fighting and motion simulation of 3D humanoid robots. From the interaction interface integrity and joint movement fluency generated, Argon surpassed Opus 5.5 in several key indicators.

Image

However, the original Argon is not perfect in all cases.

The AI/ML API of the testing organization in conducting the 3D scenario generation comparison of "Godzilla War Kong" pointed out that although the call cost of Argon ($4.22) is much lower than that of GPT-6 Astra ($23.82), there is still significant room for improvement in the image quality and morphological restoration of specific complex biological models.

Image

From the official disclosure of hardcore engineering data to the actual measurement of the whole network, it can be seen that Google's play in the Gemini 4 generation has undergone a significant change.

Image

Instead of clinging to simple chat conversations, big models are really crammed into industrial-grade code repositories, science labs, and production pipelines by reducing inference costs, stretching context windows, and strengthening code and spatial physics reasoning.

Considering the current background of Claude's serious title and ChatGPT's frequent drooling, the comprehensive ability and API price of Gemini 4 Argon is undoubtedly a timely rain, which greatly alleviates the anxiety of daily productivity tools and gives strength to businessmen.

Image

Just in the face of the computing power consumption caused by high-intensity reasoning, Google's "landlord family" does not seem to have much surplus food.

According to the latest official access rules in effect, unpaid users have now been automatically assigned the model by the platform by default, and the vast majority of requests are handled by the lightest Flash-Lite. Once the automatic selection is turned off, Flash-Lite can only be used fixedly.

Image

At the same time, the Gemini application has fully switched to a dynamic computing power limit mechanism that is refreshed every 5 hours and linked to the weekly upper limit. Raising the thinking level, using Deep Research, or calling the flagship model will accelerate the consumption quota. If you want to feel the full blood ability, you still can't bypass the threshold of paying by the step.

But in any case, just as argon played a role in industrial manufacturing, the Gemini 4 Argon's landing finally allowed Google to re-establish a solid line of defense on the top model's table.

Image

This article is from: https://www.ifanr.com/1684241 For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry