在ModelGo使用Seedance 2.0：实测每条视频省$0.36

四组不同的生产需求、两个API平台，以及实际费用记录。本组测试中，ModelGo合计费用为$4.636556，OpenRouter为$6.073200；在实际交付规格一致的条件下，ModelGo费用低23.66%。

Seedance 2.0是什么？

字节跳动Seedance 2.0是支持文字与图像参考工作流的多模态音视频生成模型。镜头控制、运动连续性和参考主体保持能力，使其适合产品短片、建筑发布和游戏预告。

测试方法与实际费用

本次测试包含四组独立需求：两组纯文字生成、两组参考图生成；在双平台可访问时，使用相同Prompt、参考文件和请求设置。ModelGo费用来自逐条匹配的已结算账单，OpenRouter费用来自已完成响应中的用量费用；表格展示模型输出费用，保留六位小数，不包含单独生成参考图的费用。

费用口径：本文比较模型调用费用，不是为账户充值所需的实际支出；不包含充值或支付手续费，包括另一篇价格对比文章计入的5.5%充值费。本文费用差额及比例仅按调用费用计算。

四个真实生产案例

为准确复现真实调用，本节保留当时提交的英文Prompt原文；中文说明用于解释生产目标。

运动鞋拆解装配短片

应用行业：鞋履/产品广告。

以拆解装配的镜头展示运动鞋结构，重点检查鞋面、鞋底与其他部件在运动过程中的形状保持、对位和装配连贯性。

实际提交的英文Prompt

Create a 10-second premium product film for an original graphite running shoe with a translucent blue sole. Seconds 0–3: a slow lateral dolly reveals the intact shoe on a dark studio plinth. Seconds 3–7: the outsole, cushioning layer and textile upper separate vertically into a clean exploded view, with each component retaining its shape and aligned position. Seconds 7–10: all three components settle back into the original shoe, ending on a stable hero composition with empty space on the left for copy added later. Maintain one continuous camera path, plausible materials and controlled rim lighting. No text, logos, numerical performance claims, extra shoes or chaotic particles. Emphasize assembly clarity over fast editing. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

ModelGo费用低23.66%。

轨道无人机机库揭幕

应用行业：游戏工作室/预告片预演。

为游戏预告制作轨道无人机机库揭幕，检查空间尺度、主体运动与镜头路径，让环境揭示有明确节奏。

实际提交的英文Prompt

Create a 10-second cinematic reveal for an original science-fiction game. Begin just behind a maintenance rail inside a vast orbital hangar, with a compact triangular exploration drone in the foreground. Over seconds 0–4, the camera glides past two grounded maintenance stations and approaches the drone. During seconds 4–8, the heavy segmented hangar doors open to reveal a blue planet and a distant ring station; the drone slowly rises one meter from its platform. Finish with a two-second stable silhouette against space. Preserve the drone’s geometry, correct scale and consistent direction of light. Use restrained cyan instrument lights and warm mechanical highlights. No combat, explosions, licensed designs, text or abrupt camera cuts. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

ModelGo费用低23.66%。

机械腕表剖面演示

应用行业：奢侈品/产品说明。

基于原创腕表参考图制作机械结构剖面演示，重点检查表壳身份、内部部件关系和镜头过程中的连续性。

这张原创参考图用于ModelGo；对OpenRouter可访问的模型，对照请求也使用同一文件。

实际提交的英文Prompt

Animate the supplied original mechanical watch into a 10-second technical brand film. Preserve the case silhouette, brushed steel finish, dark dial and hand positions from the input. Seconds 0–3: a gentle macro push-in across the crystal. Seconds 3–7: the camera transitions through a clean semi-transparent case section to reveal three clearly connected gears turning at a slow controlled pace; the outer case remains stable. Seconds 7–10: the crystal becomes opaque again and the camera ends on the original three-quarter hero view. Use warm edge lighting, realistic metal reflections and black negative space for later editorial copy. Do not change the dial markings, add a logo, multiply gears randomly or suggest a functional engineering certification. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

ModelGo费用低23.66%。

零售展亭昼夜发布短片

应用行业：建筑/地产营销。

保留零售展亭参考建筑的体量与空间关系，制作白天到夜间的发布短片，检查光照变化和建筑连续性。

这张原创参考图用于ModelGo；对OpenRouter可访问的模型，对照请求也使用同一文件。

实际提交的英文Prompt

Create a 10-second architectural launch film from the supplied retail pavilion image. Preserve the pavilion footprint, timber canopy, three glazed bays and surrounding planted courtyard. Begin with late-afternoon sunlight and a slow eye-level dolly along the front facade. From seconds 3–8, transition naturally through blue hour as interior lights illuminate sequentially behind the same windows and the sky darkens. Finish with two seconds of a stable warmly lit exterior, with two adult visitors walking slowly toward the entrance. Keep the architecture rigid and all reflections consistent. No additional storeys, changing window counts, signage, invented tenant logos or dramatic weather. The visual objective is a credible occupancy and lighting story. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.

ModelGo费用低23.66%。

如何计算费用差额

对于实际交付规格一致的输出，费用降低比例的计算方式为(OpenRouter 实际费用 − ModelGo 已结算费用) / OpenRouter 实际费用 × 100。判断两次输出是否可比时，应检查实际返回的尺寸、时长和音轨。这四组需求的平均单次费用差为$0.359161；如果按同样工作负载重复100或1,000次，费用差约为$35.916088或$359.160882。这只是按本次实测推算的用量示例，不是新增实际账单。

开始使用ModelGo API

先安装requests并设置环境变量MODELGO_API_KEY。以下代码使用首个真实案例的端点、模型ID和参数，已在代码中直接写入实际Prompt。此示例提交异步视频任务，后续轮询及下载方式见代码下方。

import os, requests

prompt = 'Create a 10-second premium product film for an original graphite running shoe with a translucent blue sole. Seconds 0–3: a slow lateral dolly reveals the intact shoe on a dark studio plinth. Seconds 3–7: the outsole, cushioning layer and textile upper separate vertically into a clean exploded view, with each component retaining its shape and aligned position. Seconds 7–10: all three components settle back into the original shoe, ending on a stable hero composition with empty space on the left for copy added later. Maintain one continuous camera path, plausible materials and controlled rim lighting. No text, logos, numerical performance claims, extra shoes or chaotic particles. Emphasize assembly clarity over fast editing. Deliver one coherent ten-second clip. Avoid speed ramps, unrelated cutaways and abrupt changes in camera direction; preserve the specified subject throughout.'
headers = {"Authorization": "Bearer " + os.environ["MODELGO_API_KEY"]}
payload = {'model': 'doubao-seedance-2.0',
 'prompt': prompt,
 'resolution': '720p',
 'duration': 10,
 'aspect_ratio': '16:9',
 'generate_audio': False}
response = requests.post('https://api.modelgo.ai/v1/video/generations', headers=headers, json=payload, timeout=420)
response.raise_for_status()
result = response.json()
print(result)  # Save the returned task ID for polling.

提交后保存响应中的id / task_id，通过GET https://api.modelgo.ai/v1/videos/{task_id}轮询至成功或失败状态；成功后使用响应中的metadata.url下载视频。

原生content端点的首帧生成请求使用content数组，其中包含文字部分，以及带有role: "first_frame"的image_url部分。

ModelGo与OpenRouter的模型ID和媒体请求格式存在差异。可以复用创作需求，接入时请根据各平台文档调整端点、参考图结构、任务轮询和下载处理。

ModelGo API文档

在ModelGo使用Seedance 2.0

使用本页介绍的Seedance 2.0模型，请前往ModelGo的Models页面。

前往ModelGo使用Seedance 2.0

参考资料

以下资料用于核对模型能力、接口与用量记录；案例需求由本次测试独立编写。

• 模型官方文档或系列展示：用于了解能力与创意方向；相关系列资料不代表其他变体的规格保证。
• OpenRouter API参考：用于核对请求约定及平台差异。
• OpenRouter用量与费用记录：用于理解用量信息；本文费用来自实际测试记录。
• OpenRouter模型目录API：在新接入前核对模型ID与目录属性。
• OpenRouter视频生成指南：用于了解媒体请求结构和工作流。