科技 · Gemini 4 Hydrogen
谷歌的 Gemini 4 Hydrogen 更便宜。真正的考验在于能否完成工作。
原语言视频
译自英文原文。
谷歌推出的新款、更便宜的 AI 模型在一项智能指数上略胜其更大的同系列模型。这是颇具吸引力的发布宣传,但还没有回答开发者最关心的问题:要以更低成本正确完成工作吗?
Token 是用于计量输入模型的内容和模型输出内容的单位。因此,这些是按使用量计费的价格,并非完成一项任务的固定费用。如果模型需要更多次尝试,或每次尝试消耗更多 token,较低的单位价格优势就可能被抵消。比较相同数量的输入和输出很简单;比较获得可用结果的成本则更难。
附带条件的基准测试领先优势
最吸引眼球的比较结果差距很小。谷歌称,在默认的“high”思考设置下,Hydrogen 在 Artificial Analysis Intelligence Index 上得分为 51,Argon 为 50。新增的第四档思考级别“xhigh”将 Hydrogen 的据报得分提高到 54。更高设置属于不同配置,不能替代默认设置下的比较。1
这一区别很重要,因为“更便宜、更聪明”这个口号把多项选择压缩成了一句话。开发者必须选择思考设置、衡量 token 用量,并测试结果是否可用。通用指数上的微弱领先,是评估 Hydrogen 的理由,而不是假定它能胜任所有工作、取代 Argon 的理由。
谷歌还报告称,在测试命令行操作的 Terminal-Bench 4.1 上,Hydrogen 的表现较 Flash 大幅提升。Hydrogen 得分为 57%,Flash 为 18.2%——超过旧款模型的三倍。这是该测试上的显著进步,但不能据此认定客户的软件或工作流程也会有同等幅度的改善。1
切换模型,减少支出
根据谷歌报告的测试结果,Orchestrator 在平均少用 24% 配额的情况下,成功率接近 Carbon。“接近”在这一说法中不可或缺:谷歌并未将结果描述为完全相同。配额减少 24%,也不代表用户账单会相应减少 24%。这表明在谷歌的测试条件下,该方法可以节省资源,并非对普遍节省的承诺。1
早期用户的评价,让这些数字背后出现了一个更容易理解的问题。前 OpenAI 科学家、现任 3D 动画初创公司 Cartwheel 首席科学家的 Andrew Carr 表示,团队在一种方法上卡了 10 天,Hydrogen 找出了其中两个问题。由前谷歌首席科学家 Jeff Dean 联合创立的 Discovery Loop,也是被点名的早期测试者之一。这些是谷歌挑选的发布案例,并非受控证据,不能说明其他团队能多常获得同样的好处。1
发布之初即可广泛使用
这与其更大的同系列模型推出时的安排形成对比。Argon 最初仅向谷歌 Fairwind Program 中经过审核的网络安全防御人员开放,近两周的受限访问期结束后,于 10 月 13 日向付费开发者开放。Carbon 同样先面向 Fairwind 成员,随后于 11 月 10 日向付费开发者和 Google AI Ultra 订阅用户开放。两天后,Hydrogen 发布,使该系列在短短六周多一点的时间里推出了三款模型。推出方式不同,并不能证明它们的能力或安全措施完全相同。1
下一个有明确日期的商业节点是 11 月 30 日,届时 Argon 的入门定价预计结束。SiliconANGLE 报道称,之后的费率为每百万输入 token 4 美元、每百万输出 token 20 美元。如果届时确如报道所述调整,Hydrogen 标价上的优势将显著扩大。与当前的价格比较一样,未来费率仍需依据谷歌的定价记录确认。1
Hydrogen 让更多人现在就能尝试一款更便宜的模型。它能否成为更值得依赖的低价模型,取决于一些不如发布日排名吸引眼球的指标:结果是否正确、是否需要重试、思考设置以及资源总用量。真正有用的比较,不只是每百万 token 的价格,而是无需返工即可完成工作的成本。
来源19
Vercel
AI/TLDR
ArtificialWatch
Google
Aiholics
TechCrunch
The Decoder
The Verge
- 另有+5家媒体
- HeyDay News
按媒体查看所有来源
来自 19 家媒体的 21 个来源· 引用最多的优先
9to5google.com1
AI Early Editionaiearlyedition.com1
AI/TLDRai-tldr.dev1
Aiholicsaiholics.com1
ArtificialWatchartificialwatch.com1
eesel AIeesel.ai1
Google Antigravityantigravity.google1
Google News Canada (CA:en)siliconangle.com1
MakeUseOfmakeuseof.com1
播出原貌
- Google released Gemini 4 Hydrogen to developers and free app users at once.
- The new model opens a cheaper route into Google's latest AI family, without the restricted launch its larger models had. Google also says Hydrogen edges past Argon on one intelligence index. The practical question is whether cheaper tokens will mean cheaper, successfully completed work.
- Hydrogen reportedly costs one dollar fifty per million input tokens and seven dollars fifty per million output tokens. Argon's introductory rates are two dollars for input and ten dollars for output. Hydrogen is cheaper on both sides of that calculation.
- Tokens are the units used to meter what you send the model and what it produces. A task can consume different amounts of each. Those rates apply across a reported roughly one-million-token context window, the material a model can work with at once. Listings differ on the exact maximum, describing it as one million or 1.05 million tokens.
- Google says Hydrogen scores 51 at its default high thinking setting, against Argon's 50. A new fourth thinking level, called xhigh, takes Hydrogen to 54. That higher result comes from changing the setting, so the default comparison remains 51 against 50.
- The benchmark operator also lists Hydrogen at xhigh with a score of 54. Google's default-setting comparison is its own reported result. An index score doesn't tell a developer how often their particular software will work correctly. The same model still has to handle the actual job, at the thinking setting the developer chooses.
- Google DeepMind senior vice president Koray Kavukcuoglu calls Hydrogen “built for rapid execution.” Google says it finishes tasks 2.5 times faster at its default setting than Gemini 3.8 Flash, the model released in September.
- Google launched Hydrogen on November 12, 2026, through the Gemini API, Google AI Studio and Google Search, with access also reaching the free Gemini app. Search availability is global for all users. Google AI Pro and Ultra subscribers can use Hydrogen in Antigravity, Google's coding tool. Free-app access comes without a specified numerical allowance.
- Subscribe to this channel as we follow whether Hydrogen's lower rates translate into cheaper completed work.
- Google says Hydrogen scored 57 percent on Terminal-Bench 4.1, which tests work done at a command line. Gemini 3.8 Flash scored 18.2 percent. Hydrogen's result is more than triple the older model's score on that test.
- Flash is also the comparison model for that 2.5-times speed claim. Google hasn't demonstrated that speedup across every user's workload. Its separate coding-tool launch tackles resource use another way. Antigravity now has Orchestrator, a feature that chooses between Hydrogen and the more capable Gemini 4 Carbon as a coding session progresses.
- Google says Orchestrator used an average of 24 percent less quota while approaching Carbon's success rate on Terminal-Bench 4.0. That's a separate test version from the Flash comparison. The combination got close to Carbon's result while using less of the resource allowance.
- Orchestrator's proposition is selective use of the larger model, saving its resources for the parts of a session that need them. A 24 percent quota reduction doesn't guarantee a matching reduction in the bill. Google also offered an early example of Hydrogen helping a team work through a problem it had been struggling to solve.
- Andrew Carr, a former OpenAI scientist and now chief scientist at animation startup Cartwheel, said Hydrogen caught two problems in an approach his team had spent 10 days stuck on. Another early tester was Discovery Loop, co-founded by former Google chief scientist Jeff Dean.
- Those experiences are launch testimonials. They don't measure how frequently another team will get the same benefit. Hydrogen's immediate broad release does give more people the chance to try it. Argon and Carbon both began with Google's Fairwind Program, a group of vetted cybersecurity defenders, before Google expanded their access.
- Argon reached paying developers on October 13 after nearly two weeks of restricted access. Carbon opened to paying developers and Google AI Ultra subscribers on November 10. Hydrogen followed on November 12. Google shipped three Gemini 4 models in just over six weeks.
- Hydrogen's lower token rates are available now. Whether they buy cheaper successful work depends on actual task accuracy, thinking settings and resource use. Argon's introductory pricing is scheduled to last through November 30. Its quoted rates then rise to four dollars per million input tokens and twenty dollars per million output tokens, widening the price gap if that change takes effect.
译文更新于





