Google 发布 Gemini 3.7 Flash

2026 年 8 月 13 日，Google 发布 Gemini 3.7 Flash。距离 3.6 Flash 发布只有三周。三周前那场发布的主角是「三款分工明确的 Flash」；这一次只有一款，但信息量并不少：官方称它是「迄今在编程与智能体场景下最智能的主力模型」，DeepSWE 从 49.0% 升到 65.3%，WebDev Arena Elo 从 1538 升到 1588，AutomationBench 从 17.0% 升到 30.4%。

这些数字需要带着厂商自述的边界读，但迭代方向是清楚的：编程与智能体。3.6 Flash 发布时，Google 还在用「工作马」的定位解释为什么旗舰缺席；三周后，这匹工作马自己先跑了一轮。把 Flash 系列的迭代周期压缩到三周，说明 Google 已经接受一个现实：在 3.5 Pro 缺席、Gemini 4 预训练刚启动的窗口期，Flash 就是主武器。

比基准更直接的是价格。截至 2026 年 12 月 31 日，3.7 Flash 限时五折：输入 0.75 美元、输出 3.75 美元每百万 token，2027 年 1 月 1 日起恢复 1.50/7.50，3.6 Flash 同步享受优惠。第一财经的标题把这件事说得很直白：「DeepSeek 涨价，谷歌降价」。DeepSeek 在 8 月预告 API 大幅调价，Google 在同一天用五折回应——价格战从「谁更便宜」变成「谁先出牌」。

发布渠道同样值得读。3.7 Flash 同步接入 Google Antigravity、Gemini API、Gemini Spark 与 Enterprise Agent Platform。Gemini Spark 是面向 160 多个国家 Google AI Pro/Ultra 订阅用户的智能体产品，发布当天即接入新模型；Enterprise Agent Platform 覆盖企业侧。个人与企业入口一起更新，说明 Google 把 Flash 当成全渠道的主力，而不是开发者的试验品。

三周迭代、限时五折、全渠道接入——这三件事放在一起，读出的不是「Google 又发了个模型」，而是「Google 选择了效率与价格作为旗舰缺席期的竞争方式」。对开发者，主力模型的迭代节奏从「半年一次」变成「三周一次」，选型决策的窗口被压缩；对行业，Flash 系列成为 Google 与对手竞争的主武器，价格战蔓延到主力档。至于 3.5 Pro 何时到来，发布页上没有答案——但至少这一周，Google 用行动回答了「没有旗舰怎么办」。

On August 13, 2026, Google released Gemini 3.7 Flash—just three weeks after 3.6 Flash. That earlier launch was about "three clearly divided Flash models"; this one is a single model, but it carries plenty of signal: Google calls it the most intelligent workhorse model for coding and agents, with DeepSWE up from 49.0% to 65.3%, WebDev Arena Elo from 1538 to 1588, and AutomationBench from 17.0% to 30.4%.

Read with the vendor-self-reported caveat, but the direction is clear: coding and agents. When 3.6 Flash launched, Google was still explaining the flagship's absence with the workhorse framing; three weeks later, the workhorse ran a lap itself. Compressing the Flash iteration cycle to three weeks means Google has accepted a reality: with 3.5 Pro absent and Gemini 4 pretraining just started, Flash is the main weapon.

More direct than benchmarks is price. Through December 31, 2026, 3.7 Flash is 50% off: $0.75 input and $3.75 output per million tokens, reverting to $1.50/$7.50 on January 1, 2027, with 3.6 Flash included in the discount. Yicai's headline said it plainly: "DeepSeek raises prices, Google cuts them." DeepSeek teased big API price changes in August; Google answered with a 50% cut the same week—the price war shifted from "who is cheaper" to "who moves first."

The distribution channels deserve equal attention. 3.7 Flash launched on Google Antigravity, the Gemini API, Gemini Spark, and the Enterprise Agent Platform. Gemini Spark, the agent product for Google AI Pro/Ultra subscribers in 160+ countries, picked up the new model on launch day; the Enterprise Agent Platform covers the corporate side. Personal and enterprise doors updating together shows Google treats Flash as an all-channel workhorse, not a developer experiment.

Three-week iteration, limited-time 50% cut, all-channel rollout—read together, the signal is not "Google released another model" but "Google chose efficiency and price as its way to compete while the flagship is absent." For developers, the workhorse cadence went from "twice a year" to "every three weeks," compressing selection windows; for the industry, Flash became Google's main weapon, spreading the price war to the workhorse tier. As for when 3.5 Pro arrives, the release page has no answer—but this week, Google answered "what do we do without a flagship" with action.