[{"data":1,"prerenderedAt":676},["ShallowReactive",2],{"blog-post-/blogs/deepseek-price-drop-guide":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"image":11,"alt":12,"ogImage":11,"tags":13,"published":18,"body":19,"_type":669,"_id":670,"_source":671,"_file":672,"_stem":673,"_extension":674,"sitemap":675},"/blogs/deepseek-price-drop-guide","blogs",false,"","DeepSeek 降价 75%：作为开发者，我该什么时候用？","DeepSeek V4-Pro 降价 75%，缓存命中低至 0.025 元/百万 Token。本文分析作为开发者在不同场景下如何选择模型，以及降价带来的开发习惯改变。","2026-05-25","/blogs-img/deepseek-price-drop-guide.jpg","DeepSeek 降价与开发者选择指南封面图",[14,15,16,17],"AI编程","DeepSeek","开发者工具","模型选择",true,{"type":20,"children":21,"toc":658},"root",[22,30,41,46,57,64,69,79,95,100,106,111,119,124,129,137,142,150,155,163,183,189,199,204,217,225,230,237,255,261,266,274,279,287,292,300,313,319,329,334,352,357,363,368,376,381,389,394,402,407,412,567,572,584,632,644,649],{"type":23,"tag":24,"props":25,"children":26},"element","p",{},[27],{"type":28,"value":29},"text","这周最大的新闻，不是哪个模型又刷榜了，而是 DeepSeek 宣布 V4-Pro API 永久降价 75%。",{"type":23,"tag":24,"props":31,"children":32},{},[33,35],{"type":28,"value":34},"缓存命中 0.025 元/百万 Token。我算了一下，按 Claude Code 的典型使用量——96.6% 缓存命中率，30 天 18.7 亿 Token——换到 DeepSeek 的价格，一个月只要 46.8 元。\n",{"type":23,"tag":36,"props":37,"children":40},"img",{"alt":38,"src":39},"image.png","https://jjbiji-pic.oss-cn-beijing.aliyuncs.com/20260525103902363.png",[],{"type":23,"tag":24,"props":42,"children":43},{},[44],{"type":28,"value":45},"同一天，DeepSeek 完成了 700 亿融资，估值 450 亿美元。融完钱第一件事不是涨价回血，而是继续降价。",{"type":23,"tag":24,"props":47,"children":48},{},[49,51],{"type":28,"value":50},"商业逻辑我不展开，我只关心一件事：",{"type":23,"tag":52,"props":53,"children":54},"strong",{},[55],{"type":28,"value":56},"这对写代码的人来说，到底意味着什么？我该不该用？什么时候用？怎么用？",{"type":23,"tag":58,"props":59,"children":61},"h2",{"id":60},"先说结论降价不是让你换模型是让你多一个选择",[62],{"type":28,"value":63},"先说结论：降价不是让你换模型，是让你多一个选择",{"type":23,"tag":24,"props":65,"children":66},{},[67],{"type":28,"value":68},"很多开发者看到降价第一反应是\"要不要从 Claude/GPT 切到 DeepSeek\"。",{"type":23,"tag":24,"props":70,"children":71},{},[72,74],{"type":28,"value":73},"我的建议是：",{"type":23,"tag":52,"props":75,"children":76},{},[77],{"type":28,"value":78},"别急着换，先想清楚场景。",{"type":23,"tag":24,"props":80,"children":81},{},[82,84,89,91],{"type":28,"value":83},"模型不是越便宜越好，也不是越强越好，而是",{"type":23,"tag":52,"props":85,"children":86},{},[87],{"type":28,"value":88},"匹配场景才好",{"type":28,"value":90},"。DeepSeek V4-Pro 的特性决定了它在某些场景是最佳选择，在另一些场景你还是会选 Claude 或 GPT。\n",{"type":23,"tag":36,"props":92,"children":94},{"alt":38,"src":93},"https://jjbiji-pic.oss-cn-beijing.aliyuncs.com/20260525101824952.png",[],{"type":23,"tag":24,"props":96,"children":97},{},[98],{"type":28,"value":99},"下面我按实际开发场景拆一下。",{"type":23,"tag":58,"props":101,"children":103},{"id":102},"场景一agent-工作流deepseek-的甜蜜区",[104],{"type":28,"value":105},"场景一：Agent 工作流——DeepSeek 的甜蜜区",{"type":23,"tag":24,"props":107,"children":108},{},[109],{"type":28,"value":110},"如果你在做 AI Agent 开发，DeepSeek V4-Pro 是目前性价比最高的选择。原因有三个：",{"type":23,"tag":24,"props":112,"children":113},{},[114],{"type":23,"tag":52,"props":115,"children":116},{},[117],{"type":28,"value":118},"1. 缓存命中率天然高",{"type":23,"tag":24,"props":120,"children":121},{},[122],{"type":28,"value":123},"Agent 工作流的特点是：系统提示词固定、工具定义固定、上下文反复调用。这意味着缓存命中率极高——DeepSeek 自己的数据显示 Claude Code 的缓存命中率 96.6%，你的 Agent 工作流大概率也在 90% 以上。",{"type":23,"tag":24,"props":125,"children":126},{},[127],{"type":28,"value":128},"缓存命中 0.025 元/百万 Token，未命中 3 元——差了 120 倍。命中率越高，DeepSeek 的价格优势越夸张。",{"type":23,"tag":24,"props":130,"children":131},{},[132],{"type":23,"tag":52,"props":133,"children":134},{},[135],{"type":28,"value":136},"2. 100 万 Token 上下文",{"type":23,"tag":24,"props":138,"children":139},{},[140],{"type":28,"value":141},"Agent 需要维护大量上下文：对话历史、工具调用记录、中间结果。100 万 Token 意味着你几乎不用担心上下文溢出，尤其是多轮 Agent 协调的场景。",{"type":23,"tag":24,"props":143,"children":144},{},[145],{"type":23,"tag":52,"props":146,"children":147},{},[148],{"type":28,"value":149},"3. 大量工具调用的成本可控",{"type":23,"tag":24,"props":151,"children":152},{},[153],{"type":28,"value":154},"一个复杂的 Agent 任务可能涉及几十次甚至上百次 API 调用。如果用 Claude Opus，单次调用可能要几块钱；用 DeepSeek，因为缓存命中的存在，同样的任务可能只要几分钱。",{"type":23,"tag":24,"props":156,"children":157},{},[158],{"type":23,"tag":52,"props":159,"children":160},{},[161],{"type":28,"value":162},"实操建议：",{"type":23,"tag":164,"props":165,"children":166},"ul",{},[167,173,178],{"type":23,"tag":168,"props":169,"children":170},"li",{},[171],{"type":28,"value":172},"Agent 的系统提示词和工具定义尽量固定，最大化缓存命中",{"type":23,"tag":168,"props":174,"children":175},{},[176],{"type":28,"value":177},"中间结果用结构化格式存储，减少重复上下文",{"type":23,"tag":168,"props":179,"children":180},{},[181],{"type":28,"value":182},"搭配 DeepSeek-TUI（2.3 万 Star 的终端工具）可以做本地 Agent 编排",{"type":23,"tag":58,"props":184,"children":186},{"id":185},"场景二批量代码处理成本敏感型任务的最佳选择",[187],{"type":28,"value":188},"场景二：批量代码处理——成本敏感型任务的最佳选择",{"type":23,"tag":24,"props":190,"children":191},{},[192,194],{"type":28,"value":193},"什么是\"批量代码处理\"？代码审查、批量重构、注释生成、测试用例生成、文档翻译……这些任务的特点是：",{"type":23,"tag":52,"props":195,"children":196},{},[197],{"type":28,"value":198},"单次不需要最强推理，但量很大。",{"type":23,"tag":24,"props":200,"children":201},{},[202],{"type":28,"value":203},"比如你要给一个 50 个文件的项目批量生成单元测试。每个文件调用一次 API，50 次调用。",{"type":23,"tag":164,"props":205,"children":206},{},[207,212],{"type":23,"tag":168,"props":208,"children":209},{},[210],{"type":28,"value":211},"Claude Opus 4.7：50 × 约 3 元 = 150 元",{"type":23,"tag":168,"props":213,"children":214},{},[215],{"type":28,"value":216},"DeepSeek V4-Pro：50 × 约 0.5 元（缓存命中后更低）= 25 元",{"type":23,"tag":24,"props":218,"children":219},{},[220],{"type":23,"tag":52,"props":221,"children":222},{},[223],{"type":28,"value":224},"6 倍差距，而且代码质量差距不大。",{"type":23,"tag":24,"props":226,"children":227},{},[228],{"type":28,"value":229},"这种\"不需要天才级推理，但需要大量重复调用\"的场景，DeepSeek 就是最佳选择。",{"type":23,"tag":24,"props":231,"children":232},{},[233],{"type":23,"tag":52,"props":234,"children":235},{},[236],{"type":28,"value":162},{"type":23,"tag":164,"props":238,"children":239},{},[240,245,250],{"type":23,"tag":168,"props":241,"children":242},{},[243],{"type":28,"value":244},"把项目结构作为系统提示固定下来，每轮只传差异部分",{"type":23,"tag":168,"props":246,"children":247},{},[248],{"type":28,"value":249},"批量任务用脚本自动化，不要手动一个个跑",{"type":23,"tag":168,"props":251,"children":252},{},[253],{"type":28,"value":254},"输出格式在系统提示里定义好，减少后处理成本",{"type":23,"tag":58,"props":256,"children":258},{"id":257},"场景三长上下文对话需要看情况",[259],{"type":28,"value":260},"场景三：长上下文对话——需要看情况",{"type":23,"tag":24,"props":262,"children":263},{},[264],{"type":28,"value":265},"DeepSeek V4-Pro 有 100 万 Token 上下文，这在长文档分析、大项目代码理解上有天然优势。但要注意两点：",{"type":23,"tag":24,"props":267,"children":268},{},[269],{"type":23,"tag":52,"props":270,"children":271},{},[272],{"type":28,"value":273},"1. 上下文长 ≠ 推理强",{"type":23,"tag":24,"props":275,"children":276},{},[277],{"type":28,"value":278},"100 万 Token 意味着你\"放得进去\"，不代表它\"想得明白\"。对于需要深度推理的任务（比如复杂 bug 定位、架构设计决策），Claude Opus 的推理能力仍然更强。",{"type":23,"tag":24,"props":280,"children":281},{},[282],{"type":23,"tag":52,"props":283,"children":284},{},[285],{"type":28,"value":286},"2. 价格只在缓存命中时香",{"type":23,"tag":24,"props":288,"children":289},{},[290],{"type":28,"value":291},"如果你每次都在传全新的 50 万 Token 上下文，走的都是\"缓存未命中\"的 3 元/百万 Token，价格优势就缩水了。",{"type":23,"tag":24,"props":293,"children":294},{},[295],{"type":23,"tag":52,"props":296,"children":297},{},[298],{"type":28,"value":299},"我的建议：",{"type":23,"tag":164,"props":301,"children":302},{},[303,308],{"type":23,"tag":168,"props":304,"children":305},{},[306],{"type":28,"value":307},"长上下文读操作（理解代码、总结文档）→ DeepSeek，性价比高",{"type":23,"tag":168,"props":309,"children":310},{},[311],{"type":28,"value":312},"长上下文写操作（复杂重构、架构修改）→ 优先 Claude/GPT，质量更稳",{"type":23,"tag":58,"props":314,"children":316},{"id":315},"场景四快速原型验证deepseek-的隐藏优势",[317],{"type":28,"value":318},"场景四：快速原型验证——DeepSeek 的隐藏优势",{"type":23,"tag":24,"props":320,"children":321},{},[322,324],{"type":28,"value":323},"很多开发者忽略了一点：",{"type":23,"tag":52,"props":325,"children":326},{},[327],{"type":28,"value":328},"价格低到一定程度，会改变你的开发习惯。",{"type":23,"tag":24,"props":330,"children":331},{},[332],{"type":28,"value":333},"以前用 Claude 或 GPT，每次调用我都要犹豫一下——\"这个问题值不值得花几毛钱问 AI？\"现在用 DeepSeek，成本几乎可以忽略，我会更自由地：",{"type":23,"tag":164,"props":335,"children":336},{},[337,342,347],{"type":23,"tag":168,"props":338,"children":339},{},[340],{"type":28,"value":341},"让 AI 生成多个方案快速比较",{"type":23,"tag":168,"props":343,"children":344},{},[345],{"type":28,"value":346},"大胆尝试各种 prompt 策略",{"type":23,"tag":168,"props":348,"children":349},{},[350],{"type":28,"value":351},"用 AI 做快速 POC 验证，不行就扔",{"type":23,"tag":24,"props":353,"children":354},{},[355],{"type":28,"value":356},"这种\"试错成本趋近于零\"的状态，反而能帮你更快找到正确方向。",{"type":23,"tag":58,"props":358,"children":360},{"id":359},"那什么时候不该用-deepseek",[361],{"type":28,"value":362},"那什么时候不该用 DeepSeek？",{"type":23,"tag":24,"props":364,"children":365},{},[366],{"type":28,"value":367},"说完了适合的场景，也说不适合的：",{"type":23,"tag":24,"props":369,"children":370},{},[371],{"type":23,"tag":52,"props":372,"children":373},{},[374],{"type":28,"value":375},"1. 需要顶级推理的任务",{"type":23,"tag":24,"props":377,"children":378},{},[379],{"type":28,"value":380},"SWE-Bench 这类复杂编程基准测试上，Claude Opus 和 GPT-5.5 仍然领先。如果你在做一个需要深度推理的 bug——涉及多模块交互、异步时序、底层系统调用——选最强的模型，别省这个钱。",{"type":23,"tag":24,"props":382,"children":383},{},[384],{"type":23,"tag":52,"props":385,"children":386},{},[387],{"type":28,"value":388},"2. 需要特定生态集成",{"type":23,"tag":24,"props":390,"children":391},{},[392],{"type":28,"value":393},"Cursor、Windsurf、Copilot 这些工具有自己的模型适配优化。如果你重度依赖某个 IDE 的 AI 功能，它的默认模型通常调得最好，换 DeepSeek 可能体验下降。",{"type":23,"tag":24,"props":395,"children":396},{},[397],{"type":23,"tag":52,"props":398,"children":399},{},[400],{"type":28,"value":401},"3. 对输出稳定性要求极高的生产环境",{"type":23,"tag":24,"props":403,"children":404},{},[405],{"type":28,"value":406},"DeepSeek 的输出质量在大部分场景够用，但在边界 case 上，Claude/GPT 的稳定性更好。生产环境的关键路径，建议还是用最稳的。",{"type":23,"tag":58,"props":408,"children":410},{"id":409},"一张表总结",[411],{"type":28,"value":409},{"type":23,"tag":413,"props":414,"children":415},"table",{},[416,440],{"type":23,"tag":417,"props":418,"children":419},"thead",{},[420],{"type":23,"tag":421,"props":422,"children":423},"tr",{},[424,430,435],{"type":23,"tag":425,"props":426,"children":427},"th",{},[428],{"type":28,"value":429},"场景",{"type":23,"tag":425,"props":431,"children":432},{},[433],{"type":28,"value":434},"推荐",{"type":23,"tag":425,"props":436,"children":437},{},[438],{"type":28,"value":439},"理由",{"type":23,"tag":441,"props":442,"children":443},"tbody",{},[444,463,480,497,514,532,550],{"type":23,"tag":421,"props":445,"children":446},{},[447,453,458],{"type":23,"tag":448,"props":449,"children":450},"td",{},[451],{"type":28,"value":452},"Agent 工作流",{"type":23,"tag":448,"props":454,"children":455},{},[456],{"type":28,"value":457},"DeepSeek ✅",{"type":23,"tag":448,"props":459,"children":460},{},[461],{"type":28,"value":462},"缓存命中高，上下文长，成本极低",{"type":23,"tag":421,"props":464,"children":465},{},[466,471,475],{"type":23,"tag":448,"props":467,"children":468},{},[469],{"type":28,"value":470},"批量代码处理",{"type":23,"tag":448,"props":472,"children":473},{},[474],{"type":28,"value":457},{"type":23,"tag":448,"props":476,"children":477},{},[478],{"type":28,"value":479},"量大成本低，质量够用",{"type":23,"tag":421,"props":481,"children":482},{},[483,488,492],{"type":23,"tag":448,"props":484,"children":485},{},[486],{"type":28,"value":487},"长上下文阅读",{"type":23,"tag":448,"props":489,"children":490},{},[491],{"type":28,"value":457},{"type":23,"tag":448,"props":493,"children":494},{},[495],{"type":28,"value":496},"100万Token，性价比高",{"type":23,"tag":421,"props":498,"children":499},{},[500,505,509],{"type":23,"tag":448,"props":501,"children":502},{},[503],{"type":28,"value":504},"快速原型验证",{"type":23,"tag":448,"props":506,"children":507},{},[508],{"type":28,"value":457},{"type":23,"tag":448,"props":510,"children":511},{},[512],{"type":28,"value":513},"试错成本趋零，开发习惯升级",{"type":23,"tag":421,"props":515,"children":516},{},[517,522,527],{"type":23,"tag":448,"props":518,"children":519},{},[520],{"type":28,"value":521},"复杂bug定位",{"type":23,"tag":448,"props":523,"children":524},{},[525],{"type":28,"value":526},"Claude/GPT",{"type":23,"tag":448,"props":528,"children":529},{},[530],{"type":28,"value":531},"推理深度更重要",{"type":23,"tag":421,"props":533,"children":534},{},[535,540,545],{"type":23,"tag":448,"props":536,"children":537},{},[538],{"type":28,"value":539},"IDE深度集成",{"type":23,"tag":448,"props":541,"children":542},{},[543],{"type":28,"value":544},"跟IDE走",{"type":23,"tag":448,"props":546,"children":547},{},[548],{"type":28,"value":549},"适配优化更成熟",{"type":23,"tag":421,"props":551,"children":552},{},[553,558,562],{"type":23,"tag":448,"props":554,"children":555},{},[556],{"type":28,"value":557},"生产关键路径",{"type":23,"tag":448,"props":559,"children":560},{},[561],{"type":28,"value":526},{"type":23,"tag":448,"props":563,"children":564},{},[565],{"type":28,"value":566},"稳定性优先",{"type":23,"tag":58,"props":568,"children":570},{"id":569},"我的行动清单",[571],{"type":28,"value":569},{"type":23,"tag":24,"props":573,"children":574},{},[575,577,582],{"type":28,"value":576},"降价 75% 不是让你\"全切 DeepSeek\"，而是让你",{"type":23,"tag":52,"props":578,"children":579},{},[580],{"type":28,"value":581},"在合适的场景用合适的模型",{"type":28,"value":583},"。我的做法：",{"type":23,"tag":585,"props":586,"children":587},"ol",{},[588,598,608,618],{"type":23,"tag":168,"props":589,"children":590},{},[591,596],{"type":23,"tag":52,"props":592,"children":593},{},[594],{"type":28,"value":595},"Agent 和批量任务切到 DeepSeek",{"type":28,"value":597},"——这是最直接的降本",{"type":23,"tag":168,"props":599,"children":600},{},[601,606],{"type":23,"tag":52,"props":602,"children":603},{},[604],{"type":28,"value":605},"复杂推理任务继续用 Claude",{"type":28,"value":607},"——质量优先，别省错地方",{"type":23,"tag":168,"props":609,"children":610},{},[611,616],{"type":23,"tag":52,"props":612,"children":613},{},[614],{"type":28,"value":615},"搭建一个轻量的路由层",{"type":28,"value":617},"——根据任务类型自动选择模型，开发者无感切换",{"type":23,"tag":168,"props":619,"children":620},{},[621,626,628],{"type":23,"tag":52,"props":622,"children":623},{},[624],{"type":28,"value":625},"关注 DeepSeek-TUI 生态",{"type":28,"value":627},"——终端 Agent 工具链正在快速成熟，2.3 万 Star 不是偶然\n",{"type":23,"tag":36,"props":629,"children":631},{"alt":38,"src":630},"https://jjbiji-pic.oss-cn-beijing.aliyuncs.com/20260525103934398.png",[],{"type":23,"tag":24,"props":633,"children":634},{},[635,637,642],{"type":28,"value":636},"最后一个判断：DeepSeek 降价 75% 的真正意义，不是\"便宜了\"，而是",{"type":23,"tag":52,"props":638,"children":639},{},[640],{"type":28,"value":641},"让开发者可以更自由地选择用 AI 做什么",{"type":28,"value":643},"。当成本不再是限制因素，你会发现很多以前觉得\"不值得用 AI\"的场景，现在都值得了。",{"type":23,"tag":24,"props":645,"children":646},{},[647],{"type":28,"value":648},"这才是最大的利好。",{"type":23,"tag":24,"props":650,"children":651},{},[652],{"type":23,"tag":653,"props":654,"children":655},"em",{},[656],{"type":28,"value":657},"加加笔记 | 程序员视角的 AI 编程观察",{"title":7,"searchDepth":659,"depth":659,"links":660},2,[661,662,663,664,665,666,667,668],{"id":60,"depth":659,"text":63},{"id":102,"depth":659,"text":105},{"id":185,"depth":659,"text":188},{"id":257,"depth":659,"text":260},{"id":315,"depth":659,"text":318},{"id":359,"depth":659,"text":362},{"id":409,"depth":659,"text":409},{"id":569,"depth":659,"text":569},"markdown","content:blogs:deepseek-price-drop-guide.md","content","blogs/deepseek-price-drop-guide.md","blogs/deepseek-price-drop-guide","md",{"loc":4},1779683822879]