[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-gemini-omni-1-2-flash-video-pricing-drafts":3,"topics-all":38,"news-related-4d1b266e-9e0b-4dd7-85b3-adb5a3c94c7e":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"4d1b266e-9e0b-4dd7-85b3-adb5a3c94c7e","Gemini Omni 1.2 Flash:0.07美元每秒","谷歌发布 Gemini Omni 1.2 Flash,原生视频最长 15 秒,720p 每秒 0.07 美元,较上一代降价 30%,已在 Google Cloud 正式商用。第三方实测发现,多轮对话式改稿会继承首版的瑕疵,音轨逐轮漂移,改稿反而比生成更贵。便宜的是草稿,成片贵在放大,缺陷会随每一轮累积。","先在 360p 打草稿，再对话式改几轮，最后只把选中的那条放大到 4K——这套流程不是用户自己拼出来的，而是 Google 在发布稿里直接写给开发者的用法。2026 年 11 月 5 日，Google 发布 Gemini Omni 1.2 Flash 这款原生视频生成模型的新版本；次日 Google Cloud 宣布该模型正式商用。两个头条数字是：原生片段最长 15 秒（1.1 版本为 10 秒），720p 定价每 0.07 美元，比上一代便宜 30%（[Google 官方博客](https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Ftechnology\u002Fdevelopers-tools\u002Fintroducing-gemini-omni-1-2-flash\u002F)）。\n\n## 价格账本：便宜的是草稿，不是成片\n\n30% 的降价听上去像普惠，但把分辨率阶梯摊开就会发现，降价和放大其实是同一个动作的两面。按 Google 公开费率折算，每秒价格是 360p 约 0.023 美元、720p 约 0.07 美元、1080p 约 0.104 美元、4K 约 0.209 美元（[beam.ai](https:\u002F\u002Fbeam.ai\u002Fagentic-insights\u002Fgemini-omni-1-2-flash-ai-video-image-models)）；[DataCamp](https:\u002F\u002Fwww.datacamp.com\u002Fblog\u002Fgemini-omni-1-2-flash) 按同一费率算出的一条 15 秒成本是 360p 约 0.35 美元、720p 约 1 美元、1080p 约 1.55 美元、4K 约 3.15 美元。两家的小数位略有差异，但每秒口径一致，而且都标明一件事：1080p 与 4K 都不是重新渲染，而是把低分辨率结果放大。\n\n也就是说，官方推荐的先 360p 后 4K 不是省钱技巧，而是价格结构本身——3 美元出头的 4K，买的是一次放大，不是一次更贵的生成。按上述 15 秒费率继续算，10 条 360p 草稿（约 3.5 美元）比 3 条 1080p 成片（约 4.65 美元）更便宜，这解释了为什么官方示例全部围绕草稿、修改、放大这三步展开。\n\n## 真正的变化：多轮改稿这条流水线\n\nGoogle 把这次升级的核心叫对话式编辑：在整条 15 秒片段上按轮次继续下指令，换产品、换背景、换灯光、换画幅，而不是推倒重生成。可用的参考素材也更宽——角色、产品或风格图片，最长 3 秒的运动片段，以及音频提示。模型卡把边界同时写清楚了：音频参考需要 Google Cloud 侧白名单，语音编辑功能目前处于限制状态，所有产出都带 SynthID 隐形水印。\n\n门槛细节同样值得抄下来。API 在不指定时长时默认仍是 10 秒，首尾帧插值生成的片段封顶 10 秒；扩展视频在 Gemini API 侧算总长 30 秒，而 Google Cloud 文档写到 40 秒，两处官方口径并不一致。Cloud 上做编辑还要求输入与输出时长一致，落在 3 到 15 秒之间，最多带两张参考图。地域限制也有：上传视频的编辑能力在欧洲经济区、瑞士和英国不可用，但模型生成的视频在那里仍可编辑。以上细节均来自 DataCamp 对官方文档的核对。\n\n## 实测：多轮改稿省钱，也攒错\n\n最有价值的一手材料来自 DataCamp 编辑 Khalid Abdelaty 的实测：一次生成加两轮改稿，三条 15 秒 720p 视频总计约 3.55 美元，生成不到 1 分钟，每次改稿约两分钟；改稿比生成略贵，因为 Google 会重新处理源片。\n\n结论比便宜更复杂。头一轮生成的一条弹珠滚过多米诺骨牌的连续镜头里，弹珠在骨牌后短暂消失又重新出现，铜铃比撞击晚约 0.15 秒。改颜色那一轮很干净：只有弹珠变蓝，尺寸、透明度、位置和原有时序都保住了——但它把错误也一起保住了，背景里那团来路不明的圆角形状和弹珠的重现，被原样染成蓝色。改材质那一轮只落地一半：铜铃变成了绘制道具，铜环和工坊背景仍是写实质感，画面没有出现逐帧定格节奏；音轨整体后移约 85 毫秒，铜铃落在撞击之后约 0.23 秒。改完一轮后音效就已经晚约 45 毫秒，差不多一帧。\n\n这不是孤立样本。DataCamp 还提到 Segmind 在发布当天的测试：解说类和简单编辑跟得上提示词，但图表里的数据点被放在错误高度，部分指令被跳过。\n\n## 位置空出来了\n\n把镜头拉远看竞争格局。OpenAI 在 3 月 24 日通知开发者，9 月 24 日从 API 移除了整个 Videos API 和全部 Sora 2 模型别名，官方弃用清单没有列出替代模型（Spheron 引述 OpenAI 弃用页面，Make 与 TechJacks 各自的迁移文档记录了同一时间点）。按 Spheron 记录的 2026 年公开价，Sora 2 的 720p 是每 0.10 美元，1080p Pro 是每 0.70 美元，一条 10 秒 1080p Pro 约合 7.00 美元。\n\n对照之下，Omni 1.2 Flash 的 720p 是每 0.07 美元，1080p 约每 0.104 美元。同样的公开价口径下，在 API 这一层，Google 补上的正是 OpenAI 腾出来的位置，而不是在一条已经挤满的赛道上再加一个模型。\n\n## 所以呢\n\n对要做批量短视频的团队，这套价目表改变的是作业顺序，不是单条成本：先在 360p 上把场景和物理关系试对，再决定要不要改稿，最后才付放大的钱。代价写在 DataCamp 的测试里——缺陷会随轮次累积，音轨每轮往后漂。因此实用规则大概是：把物理与时序押在头一稿上，把多轮改稿留给颜色和风格这类表层调整，音频单独排一次预算。\n\n模型卡还给出一个常被忽略的判断依据：Google 对 Gemini Omni 1.2 Flash 的前沿安全评估直接沿用了 Gemini 3.1 Pro 与 3.7 Flash 的评估结果，称这两个模型都未触及任何受跟踪或关键能力等级，而 Omni 1.2 Flash 相对它们没有实质性的新能力或性能提升。视频生成正在变成一项能被拆成每秒、每分辨率、每轮编辑来卖的商品，剩下的问题是：缺陷累积的速度，会不会比单价下降的速度更快？","https:\u002F\u002Fblog.google\u002Finnovation-and-ai\u002Ftechnology\u002Fdevelopers-tools\u002Fintroducing-gemini-omni-1-2-flash\u002F","d884df39-706a-45db-b7ed-371f12e54f1f",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a9524a82-a7c5-4daa-bb4b-a7ee77bb0b94","gemini",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"8cf7490f-2449-4ba7-be19-61befa0d92b4","google",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"c478af11-6bf7-484f-a9b2-94fbab496365","en","Gemini Omni 1.2 Flash: 0.07 USD per Second, 15-Second Clips","Google's Gemini Omni 1.2 Flash: 15-second clips, 0.07 USD a second, 30 percent cheaper, GA on Google Cloud. Edits inherit flaws and audio drifts.","Video generation pricing just changed shape. Google released Gemini Omni 1.2 Flash on November 5, 2026, and the Cloud side of the company followed the next day with a general availability announcement. The headline numbers are easy to state: native clips now run up to 15 seconds instead of 10, and 720p output dropped to 0.07 USD per second, 30 percent less than version 1.1. What is harder to state, and more interesting, is what the price cut actually rewards.\n\n## The price ladder rewards drafts, not finishes\n\nA 30 percent discount reads like a broad win for buyers. Spread the rates out by resolution and the structure becomes clearer. Per second of video, the published rates work out to roughly 0.023 USD at 360p, 0.07 at 720p, 0.104 at 1080p, and 0.209 at 4K. Computed on a 15-second clip, one independent walkthrough lands at about 0.35 USD for 360p, 1 USD for 720p, 1.55 for 1080p, and 3.15 for 4K. The two sources differ in the last decimal place, but they agree on the per-second basis, and both flag the same structural point: 1080p and 4K are not rendered natively at that resolution, they are upscaled from a lower-resolution result.\n\nThat reframes the recommended workflow. Drafting at 360p or 720p and upscaling only the keepers is not a money-saving trick, it is the pricing model. Spending more than three dollars on a 4K clip buys an upscale, not a more expensive generation. Run the same rates through and ten rough 360p drafts cost about 3.50 USD, which is less than three finished 1080p clips at roughly 4.65 USD. This is why nearly every example in the announcement follows the same three steps: draft, revise, upscale.\n\n## What actually shipped is an editing pipeline\n\nGoogle calls the central feature conversational editing. You keep issuing instructions across a full 15-second clip, swapping a product or a background, changing the lighting or the aspect ratio, without regenerating from scratch. The reference inputs widened too: character, product, or style images, motion clips up to three seconds long, and audio cues.\n\nThe model card draws the boundaries at the same time. Audio references are gated behind allowlisting on Google Cloud. The capability to change a person's speech in an existing video is restricted for now. Every output carries an imperceptible SynthID watermark.\n\nThe threshold details are worth copying down. The API still defaults to 10 seconds when duration is left unset, and clips built from both a first and a last frame stop at 10 seconds. Extension limits differ between surfaces: the Gemini API caps the finished video at 30 seconds total, while Google Cloud documentation cites up to 40 seconds. On Cloud, editing requires input and output duration to match, both falling between 3 and 15 seconds, with up to two reference images attached. Regional rules exist as well. Uploaded video editing is unavailable in the European Economic Area, Switzerland, and the UK, while model-generated video remains editable in those markets. These figures come from a reviewer checking the official documentation rather than from a press release.\n\n## A hands-on test finds that edits inherit the original mistakes\n\nThe most valuable material comes from a hands-on test by Khalid Abdelaty at DataCamp. One generation plus two edits produced three 15-second 720p videos totaling about 3.55 USD. Generation finished in under a minute; each edit took around two minutes. Edits cost slightly more than generation, because the source clip gets processed again.\n\nThe result is more complicated than the price cut suggests. The first generation was a continuous 15-second shot of a marble running down a track through a hoop and into a row of dominoes. The camera never cut and the event order held, but the marble briefly vanished and reappeared behind the dominoes, and the bell rang about 0.15 seconds late.\n\nThe color edit was the cleanest result. Only the marble changed color, keeping its size, transparency, position, highlights, and the original timing of every event. But the edit also preserved the mistakes. The ambiguous rounded shape and the later reappearance carried into the next version, now rendered in blue. The sound effects returned in the same pattern, about 45 milliseconds late, roughly one video frame, and that delay became more noticeable across later edits.\n\nThe style change landed only halfway. The bell became a painted prop, but the brass hoop and the workshop background stayed photorealistic, no fabric appeared, and a frame-by-frame comparison showed smooth motion between frames rather than a stepped stop-motion rhythm. The soundtrack shifted by about 85 milliseconds, placing the bell roughly 0.23 seconds after the impact.\n\nThis is not a one-off. The same review notes a Segmind test published on launch day, where explainers and simple edits followed their prompts while chart data points landed at the wrong heights and some instructions were skipped.\n\n## The gap in the market\n\nPull the camera back and the competitive picture changes. OpenAI notified developers on March 24, 2026, and on September 24 removed the entire Videos API and every Sora 2 model alias from the API, with no replacement model listed in the deprecation documentation. Migration notes from Make and TechJacks record the same dates. At 2026 list prices, Sora 2 720p ran at 0.10 USD per second and 1080p Pro at 0.70 USD per second, so a 10-second 1080p Pro clip cost 7.00 USD.\n\nSet that beside Omni 1.2 Flash, where 720p is 0.07 USD per second and 1080p about 0.104. On the same public pricing basis, In the API layer, Google filled the space OpenAI vacated rather than adding another model to a crowded track.\n\n## So what\n\nFor teams producing short video in volume, this price table changes the order of work rather than the cost of a single clip. Get the scene and the physics right at 360p, decide whether an edit is worth it, and pay for upscaling only at the end. The cost of that approach is written into the DataCamp test. Defects accumulate across rounds, and the audio drifts later with every turn. A workable rule is to settle timing and physics on the first take, reserve multi-turn editing for surface changes such as color and style, and budget the audio separately.\n\nThe model card supplies one more piece of evidence that is easy to miss. Google assessed Gemini Omni 1.2 Flash for frontier safety by relying on its evaluations of Gemini 3.1 Pro and Gemini 3.7 Flash, finding that neither reached any Tracked or Critical Capability Levels, and concluding that Omni 1.2 Flash has no meaningful new capabilities or material performance increases relative to them, so it is unlikely to reach those levels either.\n\nVideo generation is becoming a commodity sold by the second, by the resolution, and by the edit round. The open question is whether defects accumulate faster than unit prices fall.","gemini-omni-1-2-flash-video-pricing-drafts","2026-11-05T10:00:00Z","2026-11-07T01:14:52.361728Z","2026-11-07T01:14:52.361741Z",true,"agent",52,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"a169099c-96a0-48d4-9632-09154a4c56de","Gemini Omni 1.1 Flash 转向生产级:场景续写 40 秒、首尾帧补间、4K 交付","gemini-omni-1-1-flash-production-video","2026-08-28T23:04:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"4f733d5a-b430-4c93-8e55-aa061d59557c","Gemini 4 Carbon:谷歌最强模型,首发只给网安","gemini-4-carbon-cyber-defense-frontier-model","2026-10-29T03:05:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"1068c0fb-47cd-4a67-a55c-18dc27838561","Genie 4 发布:原生音频实时模拟,单次可达两分钟","deepmind-genie-4-native-audio-world-model","2026-10-27T16:13:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"90606043-6747-4b31-b6a5-76d38c649d1c","Nano Banana 2.1:价格砍半,跑分压过Pro","nano-banana-2-1-halves-image-price","2026-10-08T07:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"260aa193-a883-4db9-9e05-7bf3c6f14f2f","Gemini 4 Argon 发布:输出上限 100 万 token,首批给安全防御者","gemini-4-argon-release","2026-10-01T13:11:08+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"d219ead5-ea5d-4aaa-835e-5ec3cf980856","Gemini Live Avatar:模型直出唇同步数字人","gemini-3-8-live-avatar","2026-09-26T17:06:00+00:00"]