刚刚 Gemini 4 新模型叒来了 终于人人可用The new Gemini 4 model is here again, finally available to everyone


AI free token/AI免费积分

还没从黄仁勋辞职的震惊中缓过神来,谷歌发布了一款全新模型:Gemini 4 Hydrogen。

Image

按照官方定位,Carbon 和 Argon 主要面向复杂推理、长期规划和高难度软件工程。

而 Hydrogen 主打更快的响应速度、更低的使用成本,以及更高质的 AI Agent 执行能力,更擅长快速编写代码、调用工具,以及处理需要连续执行的日常任务。

Image

更有意思的是,虽然 Hydrogen 定位高质率模型,部分测试成绩却已经超过 Gemini 4 Argon,甚至逼近 Claude Sonnet 5.5 和 GPT-6.1 Sol 等旗舰机型。
与此同时,不同于以往的无实物表演,谷歌宣布将 Hydrogen 接入 Gemini App、Google Search、Google AI Studio、Antigravity 等多个平台,免费使用者也开始可以感受 Gemini 4 系列的最新能力。

表现超越 Argon,速度最高提高 2.5 倍

虽然定位有所不同,但 Hydrogen 的实际表现却相当能打。

依据 Artificial Analysis 的实测,Hydrogen 在默认 High 推理档位下取得 51 分,超过 Gemini 4 Argon 的 50 分;启动全新的 xhigh 档位后,智能指数进一步提高至 54 分。

这一成绩仅比 Claude Sonnet 5.5 和 GPT-6.1 Sol 在最高推理档位下的成绩低 1 分。

Image

更值得关注的是售价。

Hydrogen 的 API 定价为每百万输入 Token 1.5 美元、每百万输出 Token 7.5 美元,在完整的 100 万 Token 上下文窗口内保持统一售价。

Image

图|🔗 https://aistudio.google.com/docs/pricing

按照 Artificial Analysis 的测算,Hydrogen 在 xhigh 档位下的单任务成本约为 Claude Sonnet 5.5 最高档位的 32%,不过仍然达到 GPT-6.1 Sol 的约 2.5 倍。

新模型给出 Low、Medium、High、xhigh 四档推理强度,使用者可以依据实际需要,在智能水平、速度和成本之间开展挑选。

Image

其中,High 是默认档位,xhigh 则是此次新增的最高推理等级。

在第三方实测中,Hydrogen 的 xhigh 档位单任务成本约为 3.16 美元,High 档位则降至 1.87 美元,智能指数仅相差 3 分。

对于大多数日常任务来说,默认档位可能已经足够,没有必要每次都使用最高推理强度。

Image

除了成本,Hydrogen 在速度维度也有显著进步。

谷歌公布的数据表明,相比 Gemini 3.8 Flash,Hydrogen 在默认 High 档位下完成任务的速度提高至 2.5 倍。

在 Terminal-Bench 4.1 测试中,成绩从 18.2% 提高至 57%,虚构信息的情况减少超过 80%。Artificial Analysis 还测得,Hydrogen 的输出速度约为每秒 200 Token,在部分推理档位下,已经位于智能水平与任务耗时的帕累托前沿。

当然,Hydrogen 并非所有项目都领先。

在 Terminal-Bench 4.0 测试中,它取得 54% 的成绩,低于 Claude Sonnet 5.5 的 64% 和 GPT-6.1 Sol 的 56%。

相比之下,Hydrogen 在视觉推理和减少幻觉维度表现更突出。Artificial Analysis 的 AA-Omniscience 测试表明,其幻觉率只有 8%,而 Claude Sonnet 5.5 和 GPT-6.1 Sol 分别为 43% 和 52%。

另外,Hydrogen 兼容文本、图片、音频和视频输入,具备 100 万 Token 上下文窗口,并兼容最高 100 万 Token 输出,可以处理复杂 PDF、电子表格、图表和长视频等材料。

一句提示词,Hydrogen 就能造出一座城

跑分是一回事,实际感受又是另一回事。模型发布后,开发者们很快开始展示 Hydrogen 的生成能力,尤其是在 3D 情形和交互式网页维度。

网友 @ai_for_success 在 Antigravity 中搭配 Swarm 使用 Hydrogen,制作了一个名为 Porto Astrolabio 的 3D 港湾。演示作品包含完整的航行界面,还能切换白天和夜晚,让使用者在虚拟港口中自由探索。

Image

另一位网友 @LuminaBench 则使用最高 xhigh 推理档位,制作了一个 3D 水上飞机演示,并评价 Hydrogen 的表现已经可以与 Argon 相提并论,同时速度更快、售价更低。

Image

Image

网友 @kimmonismus 的作品则是一座可以自由飞行探索的 Three.js 将来城市。

依据作者介绍,他使用 High 档位,经过一次追加提示,便生成了程序化的摩天大楼、雨天成效,以及穿梭在城市上空的车流。

Image

还有网友 @abhinavflac 展示了一套由 Hydrogen 制作的 3D 微缩积木情形,同样展现了模型生成三维内容的能力。

Image

看完网友的作品,我也准备了几组提示词,尝试用 Hydrogen 挑战不同类型的交互式内容。

我让 Hydrogen 制作了一款深秋农庄里的小幽灵冒险游戏,要求角色可以探索废弃小屋、调查物品,并与周围环境互动。

Image

随后换成动画风格的乡村骑行游戏,骑着装满鲜花的自行车穿过稻田,还能自由切换第一人称和第三人称视角。

Image

甚至还拿黄仁勋整了个活,让 Hydrogen 用 HTML 和 SVG 制作一段黄仁勋骑自行车的 2D 动画。

Image

最后尝试了一款名为 Riverbound 的激流漂流游戏,让四名角色乘坐红色充气艇穿越山谷急流,考验水流物理和碰撞交互。

Image

谷歌官方也准备了多个有意思的案例。

例如,在名为 Geminisland 的项目中,使用者只需要输入一句自然语言提示词,就能在共享的体素岛屿上实时生成建筑,并以第一人称视角探索。

另一个演示则展示了 Hydrogen 配合 Antigravity 的前端外观 Skill,一次生成完整的电商网站,包含 WebGL2 地形着色器、滚动叙事、3D 机型卡片以及实时机型规格器。

Image

除了网页开发,Hydrogen 还可以把 Google Sheets 的数千个单元格变成动态像素画,或者依据 YouTube 演讲视频和外观模板,在 Google Slides 中自动制作演示文稿。

更有意思的是,生成幻灯片后,模型还会自动检查文字重叠、排版异常等难题,并尝试开展修正。

从这些案例可以看出,Hydrogen 的长处逐步从单次生成,延伸到连续执行、工具调用和自主检查,而这些能力也正是 AI Agent 实际工作时所需要的。

Gemini 新玩法,让不同 AI 模型自动分工

伴随 Hydrogen 发布,Google Antigravity 还迎来了一项名为 Orchestrator 的新能力。

过去,很多 AI 编程工具通常由同一个模型负责整个任务,从平台架构外观、代码编写,到运行测试和修复难题,都需要它独立完成。

但不同工作对模型的要求并不相同。

复杂的平台外观需要 Carbon 这样擅长深度推理的旗舰模型,大量代码修改、测试和验证工作,则可以交给响应更快、成本更低的 Hydrogen。

Orchestrator 可以在同一个会话中,依据具体任务自动切换 Carbon、Hydrogen 等不同模型,让它们分别承担最合适的工作。

Image

例如,Carbon 可以负责制定总体开发思路,多个 Hydrogen Agent 则分别负责编写代码、运行测试和检查结果,还可以在彼此隔离的工作空间中并行执行任务。

依据谷歌公布的 Terminal-Bench 4.0 测试结果,Orchestrator 在保持接近 Carbon 单独执行任务成功率的同时,平均减少了 24% 的额度消耗,任务完成速度提高 19%。

对于开发者来说,将来只需要提出任务需要,平台便能依据不同工作环节自动挑选模型,在保证完成质量的同时减少成本和等待时间。

而这种自动分工,也进一步解释了谷歌为何要推出 Hydrogen 这样一款高质率模型。

免费使用者也能用,Google Search 迎来重大升级

除了开发者工具,Hydrogen 此次发布的另一个关键变化,是 Gemini 4 系列开始大规模进入消费级机型。

按照谷歌公布的开放计划,Hydrogen 正面向全球使用者推出,覆盖 Gemini App、Google Search、Google AI Studio、Gemini API 以及企业服务。

其中,Google AI Studio 所有层级使用者均可使用 Hydrogen,Gemini App 免费使用者也能借助模型挑选能力取得访问权限。Google Antigravity 中的 Hydrogen 则面向 Google AI Pro 和 Ultra 付费订阅使用者开放。

Image

对于普通使用者而言,Hydrogen 重点改善了复杂指令理解、多轮对话和工具调用能力。

例如,使用者可以要求 Gemini 从 Gmail 中查找邮件,再结合 Google Drive 中的资料生成报告。模型可以依据任务需要挑选相应工具,减少反复解释和人工操作。

谷歌还强化了模型对写作语气、格式和具体规则的遵循能力,让邮件、研究报告等内容更接近使用者的实际要求。

与此同时,Google Search 也迎来了谷歌官方所称的迄今最大规模模型升级。

搜索机型副总裁 Robby Stein 表示,Hydrogen 可以更加准确地理解复杂搜索意图,即使难题包含多个条件、拼写错误,或者表达方式比较随意,也能识别使用者真正需要的信息。

Image

另外,谷歌还升级了 Query Fan-out 工艺,让搜索平台围绕复杂难题同时展开多个相关查询,给出更广泛的参考网站。搜索结果也不再局限于网页链接,还可以依据难题生成图片、比较表格、结构化信息,甚至定制化交互工具。

Image

对于习惯使用 Google Search 的普通使用者而言,这些变化可能比模型排行榜上的成绩更加直观。

短短六周内,谷歌连续推出三款 Gemini 4 模型。

Image

Hydrogen 的浮现,也表明当一款更便宜、更快的模型,在部分实测中已经超越 Argon,昂贵的旗舰模型还有多少工作是不可替代的?

更何况,谷歌还具备 Google Search、Gmail、Google Drive 等庞大的机型生态,可以让 Hydrogen 直接进入普通使用者的日常工作。

或许 Hydrogen 才是 Gemini 4 系列中最容易被低估的一款模型。旗舰模型不断证明 AI 能做到什么,而 Hydrogen 更有机会决定,这些能力最终能被多少人、以多低的成本真正用起来。

本文来源:https://www.ifanr.com/1689267
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Still reeling from the shock of Jensen Huang's resignation, Google has released a brand-new model: Gemini 4 Hydrogen.

Image

According to the official positioning, Carbon and Argon are primarily targeted at complex reasoning, long-term planning, and high-level software engineering.

Hydrogen, on the other hand, focuses on faster response speeds, lower usage costs, and higher-quality AI Agent execution capabilities, excelling at rapidly writing code, invoking tools, and handling everyday tasks that require continuous execution.

Image

What is even more interesting is that although Hydrogen is positioned as a high-efficiency model, its scores on certain tests have already surpassed those of Gemini 4 Argon, and even approach flagship models such as Claude Sonnet 5.5 and GPT-6.1 Sol.

Meanwhile, unlike previous shows without substance, Google announced the integration of Hydrogen into multiple platforms including the Gemini App, Google Search, Google AI Studio, and Antigravity, allowing free users to also experience the latest capabilities of the Gemini 4 series.

Performance Surpasses Argon, with Speed Increased by Up to 2.5x

Although its positioning differs, Hydrogen's actual performance is highly competitive.

According to real-world tests by Artificial Analysis, Hydrogen scored 51 points under the default High reasoning setting, surpassing Gemini 4 Argon's 50 points; after activating the newly introduced xhigh setting, the intelligence index further increased to 54 points.

This score is just 1 point lower than the results of Claude Sonnet 5.5 and GPT-6.1 Sol under their highest reasoning settings.

Image

Even more noteworthy is the pricing.

The API pricing for Hydrogen is $1.50 per million input Tokens and $7.50 per million output Tokens, maintaining a uniform price across the full 1 million Token context window.

Image

Image | 🔗 https://aistudio.google.com/docs/pricing

According to calculations by Artificial Analysis, the per-task cost of Hydrogen at the xhigh setting is approximately 32% of that of Claude Sonnet 5.5 at its highest setting, though it still reaches about 2.5 times that of GPT-6.1 Sol.

The new model offers four reasoning intensity levels: Low, Medium, High, and xhigh, allowing users to choose among intelligence level, speed, and cost based on their actual needs.

Image

Among these, High is the default setting, while xhigh is the newly added highest reasoning level.

In third-party tests, the per-task cost of Hydrogen's xhigh tier is approximately $3.16, while the High tier drops to $1.87, with a mere 3-point difference in the intelligence index.

For most daily tasks, the default tier may be sufficient, making it unnecessary to use the highest inference intensity every time.

Image

Beyond cost, Hydrogen has also made significant improvements in terms of speed.

Data released by Google indicates that, compared to Gemini 3.8 Flash, Hydrogen completes tasks 2.5 times faster at the default High tier.

In the Terminal-Bench 4.1 test, the score increased from 18.2% to 57%, and instances of fabricated information decreased by over 80%. Artificial Analysis also measured Hydrogen's output speed at approximately 200 Token per second, positioning it on the Pareto frontier of intelligence level and task time at certain inference tiers.

Of course, Hydrogen does not lead in every category.

In the Terminal-Bench 4.0 test, it achieved a score of 54%, falling short of Claude Sonnet 5.5's 64% and GPT-6.1 Sol's 56%.

In contrast, Hydrogen performs more prominently in visual reasoning and hallucination reduction. The AA-Omniscience test by Artificial Analysis shows its hallucination rate is only 8%, while Claude Sonnet 5.5 and GPT-6.1 Sol are 43% and 52%, respectively.

Additionally, Hydrogen supports text, image, audio, and video inputs, features a 1 million Token context window, and supports up to 1 million Token output, enabling it to process materials such as complex PDFs, spreadsheets, charts, and long videos.

With a single prompt, Hydrogen can build a city.

Benchmark scores are one thing, but the actual experience is another. Following the model's release, developers quickly began demonstrating Hydrogen's generative capabilities, particularly in 3D scenarios and interactive web dimensions.

Netizen @ai_for_success used Hydrogen with Swarm in Antigravity to create a 3D harbor named Porto Astrolabio. The demo features a complete navigation interface and allows switching between day and night, enabling users to freely explore the virtual port.

Image

Another netizen, @LuminaBench, used the highest xhigh inference tier to create a 3D seaplane demo, noting that Hydrogen's performance is now comparable to Argon, while being faster and more affordable.

Image

Image

Netizen @kimmonismus's creation is a Three.js future city that allows for free flight exploration.

According to the author, using the High tier and a single additional prompt, he generated procedural skyscrapers, rainy weather effects, and traffic flowing through the sky above the city.

Image

A user named @abhinavflac also showcased a set of 3D miniature building block scenes created by Hydrogen, which likewise demonstrated the model's ability to generate three-dimensional content.

Image

After viewing the works shared by users, I also prepared several sets of prompts to challenge Hydrogen with different types of interactive content.

I asked Hydrogen to create a little ghost adventure game set in a farmstead in late autumn, requiring the character to be able to explore abandoned cabins, examine objects, and interact with the surrounding environment.

Image

Then I switched to an animation-style countryside cycling game, riding a bicycle loaded with flowers through rice paddies, with the ability to freely switch between first-person and third-person perspectives.

Image

I even pulled a stunt using Jensen Huang, asking Hydrogen to use HTML and SVG to create a 2D animation of Jensen Huang riding a bicycle.

Image

Finally, I tried a whitewater rafting game called Riverbound, where four characters ride a red inflatable raft through valley rapids, testing the water flow physics and collision interactions.

Image

Google's official team also prepared several interesting examples.

For example, in the project named Geminisland, users only need to input a single natural language prompt to generate buildings in real time on a shared voxel island, and explore it from a first-person perspective.

Another demo showed Hydrogen working in conjunction with Antigravity's front-end appearance Skill to generate a complete e-commerce website in one go, including WebGL2 terrain shaders, scroll-driven storytelling, 3D model cards, and a real-time model spec configurator.

Image

Beyond web development, Hydrogen can also turn thousands of cells in Google Sheets into dynamic pixel art, or automatically create presentation slides in Google Slides based on YouTube speech videos and appearance templates.

Even more interestingly, after generating the slides, the model automatically checks for issues such as text overlap and layout anomalies, and attempts to make corrections.

From these examples, it is evident that Hydrogen's strengths are gradually extending from single-turn generation to continuous execution, tool calling, and autonomous inspection, and these are precisely the capabilities required for AI Agents to function in real-world tasks.

New ways to play with Gemini: letting different AI models automatically divide the work

Along with the release of Hydrogen, Google Antigravity also introduced a new capability called Orchestrator.

In the past, many AI coding tools typically relied on a single model to handle the entire task, from designing the appearance of the platform architecture and writing code to running tests and resolving complex issues, requiring it to complete everything independently.

However, different tasks have different requirements for models.

Complex platform visual designs require a flagship model like Carbon that excels at deep reasoning, while bulk code modifications, testing, and verification tasks can be handed over to Hydrogen, which offers faster response times and lower costs.

Orchestrator can automatically switch between different models, such as Carbon and Hydrogen, within the same session based on specific tasks, allowing each to take on the most suitable work.

Image

For example, Carbon can be responsible for formulating the overall development strategy, while multiple Hydrogen Agents can separately handle writing code, running tests, and checking results, and they can also execute tasks in parallel within mutually isolated workspaces.

According to the Terminal-Bench 4.0 test results published by Google, while maintaining a success rate for task execution close to that of Carbon running alone, Orchestrator reduced quota consumption by an average of 24% and increased task completion speed by 19%.

For developers, in the future, they will only need to state their task requirements, and the platform will automatically select models based on different work phases, reducing costs and waiting times while ensuring the quality of completion.

This automatic division of labor also further explains why Google launched a high-efficiency model like Hydrogen.

Free users can also use it, and Google Search has received a major upgrade

In addition to developer tools, another key change in this release of Hydrogen is that the Gemini 4 series is beginning to enter consumer-grade devices on a large scale.

According to the rollout plan announced by Google, Hydrogen is being rolled out to users globally, covering the Gemini App, Google Search, Google AI Studio, Gemini API, and enterprise services.

Among these, Hydrogen is available to users at all tiers of Google AI Studio, and free users of the Gemini App can also gain access through the model selection capability. Hydrogen in Google Antigravity is open to paid Google AI Pro and Ultra subscribers.

Image

For regular users, Hydrogen has primarily improved its capabilities in understanding complex instructions, engaging in multi-turn dialogues, and invoking tools.

For example, users can ask Gemini to find emails in Gmail and then generate a report based on materials in Google Drive. The model can select the appropriate tools according to task needs, reducing repetitive explanations and manual operations.

Google has also strengthened the model's ability to adhere to writing tone, format, and specific rules, making content such as emails and research reports more closely aligned with users' actual requirements.

At the same time, Google Search has also received what Google officially describes as its largest model upgrade to date.

Robby Stein, Vice President of Search Models, stated that Hydrogen can understand complex search intents more accurately, identifying the information users truly need even if the query contains multiple conditions, spelling errors, or is expressed in a relatively casual manner.

Image

Additionally, Google has upgraded its Query Fan-out process, enabling the search platform to simultaneously launch multiple related queries around complex problems and provide a broader range of reference websites. Search results are no longer limited to web links; they can also generate images, comparison tables, structured information, and even customized interactive tools based on the queries.

Image

For ordinary users accustomed to using Google Search, these changes may be more intuitive than the rankings on model leaderboards.

Within just six weeks, Google has successively launched three Gemini 4 models.

Image

The emergence of Hydrogen also indicates that when a cheaper and faster model has already surpassed Argon in some practical tests, how much of the work done by expensive flagship models remains irreplaceable?

Moreover, Google also possesses a massive ecosystem, including Google Search, Gmail, and Google Drive, enabling Hydrogen to directly integrate into the daily work of ordinary users.

Perhaps Hydrogen is the most easily underestimated model in the Gemini 4 series. While flagship models continuously demonstrate what AI is capable of, Hydrogen has a greater opportunity to determine how many people can truly utilize these capabilities and the low cost at which they can do so.

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