Gemini 4 Hydrogen 对比 Gemini 4 Pro：其中一个模型确实存在，而 Google 正将另一个淘汰并转为 Flash 模型。

The most informative fact about the Pro tier is a routing change buried in G​oogle's deprecation page. Effective 1 December 2026, every API request to gemini-2.5-pro is served by gemini-3.8-flash and billed at Gemini 3.8 Flash rates. The last general-purpose model G​oogle ever shipped under the Pro label will, in three weeks, be answered by a throughput model that scores 31 on the Artificial Analysis Intelligence Index. Meanwhile G​emini 4 Hydrogen — generally available since 12 November 2026 at $1.50 and $7.50 per million tokens — scores 54 at xhigh and, at its own default, 51. G​emini 4 Pro does not exist and never has: there is no such identifier on the rate card, in the model list, or in the documentation. So the comparison is not between two models. It is between a model you can call this afternoon and a label that G​oogle is in the process of retiring.

每个带有 Pro 标签的通用模型都已退役、正在退役或无法评分。

G​oogle's deprecation table is short and it makes the pattern plain. Four identifiers have carried the Pro name for general text work in the past year, and none of them is a live, scored, recommended model today.

• gemini-3-pro-preview——于 2025 年 11 月 18 日发布，2026 年 3 月 9 日停止服务，由 gemini-3.1-pro-preview 取代。然而大量从未读过弃用公告页面的代码仍在调用它。

• gemini-3.1-pro-preview —— 于 2026 年 2 月 19 日发布，未公布下线日期，定价仍为：20 万 token 以内每百万 token 2.00 美元 / 12.00 美元，超出部分 4.00 美元 / 18.00 美元。它是一个没有有效期的预览版，而作为预览版却持续了九个月，这颇为蹊跷；Artificial Analysis 也完全没有为其公布任何智能指数数值——只给出了每秒 113 个输出 token 和 15.28 秒的首个回答 token 延迟。

• gemini-2.5-pro — deprecated, no longer selectable in G​oogle AI Studio, API access limited to accounts that have actively used it. On 1 December 2026 it stops being a Pro model in any practical sense, because the requests are answered by Gemini 3.8 Flash.

• Nano Banana Pro — gemini-3-pro-image. Real, current, and an image model. It is the only live Pro identifier G​oogle has, and it is not what anyone typing "G​emini 4 Pro" is looking for.

把那四行当作一个序列来读，Pro 层级就有了方向：预览、弃用、重新路由到 Flash 模型。这不是路线图，而是逐步收尾。

What replaced it, in G​oogle's own words

The G​emini 4 generation shipped three models and gave them element names instead of position names. G​oogle's own selection guidance, from the Hydrogen migration documentation, is unusually direct about the split.

• Gemini 4 Carbon——“我们最强的前沿模型”，每轮消耗更多思考 token，用于在困难架构、调试、科学和安全审计中检验假设并验证边界情况。11 月 30 日前为 $2.00 / $10.00，12 月 1 日起为 $4.00 / $20.00。无免费层级，无 Contributor 层级。这款模型占据的正是过去 Pro 模型所占据的位置。

• Gemini 4 Argon——高效前沿模型，适合需要深度但不需要顶配的工作。同样是 $2.00 / $10.00 的入门价目卡，同样是 12 月 1 日翻倍。它占据了 Pro 模型在状态好时曾占据的位置，而这正是让这个标签值得舍弃的那种模糊之处。

• G​emini 4 Hydrogen — "recommended upgrade from Gemini 3.8 Flash and earlier", $1.50 / $7.50 flat with no dated step, four thinking levels, a rate-limited free tier, and now the default model behind the Antigravity agent in G​emini Managed Agents.

价格结构就是线索。当一个产品家族的上限是 $4 / $20、主力款是 $1.50 / $7.50 时，阶梯中段的 $2 / $10 被两款在同一日期价格变得相同的型号占据。一个用来描述中间档位的层级名称，已经没有什么可描述的了。

你能真正运行的对比

Since there is no G​emini 4 Pro, the only Pro-labelled model a reader can put a request against is gemini-3.1-pro-preview, and the only honest comparison is on the handful of dimensions where both it and G​emini 4 Hydrogen publish a number.

• 每百万 token 价格 — Hydrogen 在整个窗口内统一为 $1.50 / $7.50，缓存读取 $0.075；Gemini 3.1 Pro Preview 200K 以内为 $2.00 / $12.00，超过 200K 为 $4.00 / $18.00，缓存读取 $0.20，享受 90% 折扣。低档区间输出成本高出 60%，超过 200K 后高出 140%。

• 输出速度——Hydrogen 在 xhigh 下为每秒 190 个 token，在其默认的 high 下为 179 个；Gemini 3.1 Pro Preview 为 113 个。Hydrogen 的生成速度快 58% 到 68%。

• 首个答案 token 时间——Gemini 3.1 Pro Preview 为 15.28 秒；Hydrogen 在 xhigh 下为 37.8 秒，在 high 下为 16.7 秒。这是旧版 Pro 模型唯一完胜的一项，而且即使对比 Hydrogen 的默认设置，它也领先 1.4 秒。如果让 Hydrogen 保持其发布报道所引用的设置，差距则为 22 秒。

• 上下文与模态——两者均具有 100 万 token 的上下文窗口；Gemini 3.1 Pro Preview 接受文本、图像、语音和视频输入，输出文本，而 Hydrogen 接受文本、图像、视频、音频和 PDF。在 2026 年并不构成差异化优势。

• 独立评分——Hydrogen 在 xhigh 下为 54，在 high 下为 51，在 medium 下为 41；Gemini 3.1 Pro Preview 未公布任何评分。这个空白正是仅凭信念维持 Pro 部署的全部问题所在：该层级最后一个仍在线的模型有延迟测量结果，却没有质量测量结果，因此没人能说清 60% 的输出溢价到底买到了什么。

这四条标价产品线合起来说明：标有 Pro 的选项如今在除最初一秒半延迟之外的每一个维度上，都是更慢、更贵、且未经实测的那一个。这个优势窄到足以被精确命名：它是一个模型能够立即开始回答的优势；这对自动补全很重要，而对任何无人值守运行的任务则远没那么重要。

“Pro”买的是什么，以及为什么没有谁把它买回去

This part is our reading rather than G​oogle's, and it is worth labelling as such. Tier names describe a position: Flash is throughput, Pro is general capability, Ultra is the ceiling. Position names work when the ladder is monotonic — more money, more capability — because a buyer can pick a rung and trust the ordering.

The G​emini 4 family broke the monotonicity. G​emini 4 Hydrogen, the cheapest member, gives up nine index points to Gemini 4 Carbon and none to Gemini 4 Argon, yet it costs less to finish an index task than either ($1.87 against $7.76 and $4.97 in Artificial Analysis's runs) and runs at two and a half times their throughput. Once a cheap model is the best economics in the family and a dear model is the best capability, the middle rung stops being a decision and starts being a compromise. Element names are what you use when the ladder is not a ladder.

There is a second reason, less flattering. A preview identifier with no shutdown date and no independent score is a way of not committing. G​oogle has kept gemini-3.1-pro-preview priced and callable for nine months without either promoting it or retiring it, while shipping three G​emini 4 models around it. Whatever the internal reason, the effect on a customer is that the Pro string is the one part of the stack nobody has promised anything about.

如果你今天拿的是 Pro 球线

四种情况，四种不同的行动，其中只有一种是紧急的。

• 生产环境中的 gemini-2.5-pro——这是紧急事项。请在 12 月 1 日前针对 Gemini 3.8 Flash 进行测试，因为此后无论你是否做了更改，响应你请求的都将是它，价格为 $0.75 / $3.75，而不是 Pro 费率。静默的模型替换更便宜、更快，而且也是另一个模型；如果你的提示词是针对 2.5 Pro 的习惯调校的，那么出问题的是调校，而不是端点。

• 生产环境中的 gemini-3.1-pro-preview——应重新测量，而不是迁移。它保持延迟优势已有九个月，并且是本次比较中唯一无需好言相求就能在十六秒内开始作答的模型。但它在当前指数上未评分，每百万输出 token 收费 12.00 美元，而 Hydrogen 为 7.50 美元，且没有公布未来规划。在认定 15 秒值得一个未经测量的模型之前，请自行以 high 设置评测 Hydrogen，并以 medium 设置评测 Gemini 4 Argon。

• A Pro string in a document rather than a config — replace it with the decision it stood for. Planning and architecture turns go to gemini-4-carbon; execution, tool calls and parallel subagents go to gemini-4-hydrogen; anything already deployed on Argon stays on gemini-4-argon until the eval says otherwise. G​oogle documents the Carbon-plans-Hydrogen-executes pattern itself, including switching models between turns of one session with previous_interaction_id.

• Nano Banana Pro — 无需处理。它是一个图像模型，目前仍是现行型号，而命名冲突是它出现在本文中的唯一原因。

That third action is the one that benefits from a routing layer rather than a model choice, because "Carbon plans, Hydrogen executes" is a composition, not a purchase. OrcaRouter carries Gemini 4 Carbon and Gemini 4 Argon at $2.00 / $10.00 with provider list price passed through at 0% markup, so the 1 December doubling of both cards appears here the same day it appears on G​oogle's, and it carries gemini-3.1-pro-preview at its two-band rate for as long as that preview stays callable — which means a migration off a Pro string needs no second contract. The routing DSL is what lets a planner-and-reviewer split stay in one call instead of in application code, and model fusion lets Carbon and Argon answer the same hard question as a panel rather than as a choice. G​emini 4 Hydrogen is not in our catalogue; it comes from G​oogle's own API and AI Studio, and calling it direct alongside the routed models is the honest arrangement.

能一锤定音的测试

If a G​emini 4 Pro is ever real, three mechanical things will be true within a day of each other: a gemini-4-pro identifier appears under the G​emini 4 heading in G​oogle's model list, the rate card gains a row with a published figure, and a model card appears carrying a context length and a knowledge cutoff. Any one alone means nothing — an identifier with no price is a gated programme, and a price with no card is a quota reservation. This is checkable in a browser in under a minute, which is more than can be said for any benchmark rumour.

Until then the useful date is 1 December 2026, and it is useful twice over. Carbon and Argon double on it. On the same day, the last consumer Pro model stops being served by a Pro model. If both happen as published, the family G​oogle is selling in 2027 has a $4 / $20 ceiling, a $1.50 / $7.50 workhorse, and no middle — and the correct response to a question about G​emini 4 Pro is not "wait for it" but "you are asking about a tier G​oogle has already replaced with a name."