[{"data":1,"prerenderedAt":1374},["ShallowReactive",2],{"header-counts":3,"footer-counts":6,"compare-litellm-vs-openrouter":9,"compare-a-litellm":10,"compare-b-openrouter":850},{"tools":4,"reviews":5},78,26,{"tools":4,"reviews":5,"playbooks":7,"news":8},22,21,null,{"id":11,"title":12,"alternatives":13,"api_compatible":9,"body":18,"category":782,"chinese_friendly":290,"cover":783,"description":784,"domestic":785,"extension":786,"faq":787,"free":785,"github":9,"languages":800,"lastVerified":9,"meta":802,"models":9,"navigation":309,"notSuitable":9,"opensource":309,"path":803,"pillar":804,"platforms":805,"priceTable":809,"pricing":819,"published":820,"relatedPlaybooks":821,"relatedReviews":823,"score":825,"self_host":309,"seo":826,"seoTitle":9,"slug":827,"sources":828,"stem":838,"suitable":9,"tagline":839,"tags":840,"updated":831,"verdict":847,"website":848,"__hash__":849},"tools\u002Ftools\u002Fcoding\u002Fapi\u002Flitellm.md","LiteLLM",[14,15,16,17],"coding\u002Fapi\u002Fopenrouter","coding\u002Fapi\u002Fone-api","coding\u002Fapi\u002Fportkey","coding\u002Fapi\u002Fhelicone",{"type":19,"value":20,"toc":770},"minimark",[21,26,35,38,41,141,144,161,167,171,176,199,204,229,233,438,441,604,607,671,675,701,704,731,734,766],[22,23,25],"h2",{"id":24},"tldr","TL;DR",[27,28,29,30,34],"p",{},"LiteLLM 是开源 LLM 网关里事实标准：MIT 协议，BerriAI 维护，双形态——SDK（",[31,32,33],"code",{},"pip install","，单进程嵌入）+ Proxy（Docker Compose + Postgres，团队级网关）。100+ 厂商统一 OpenAI 接口，虚拟 Key + 团队预算 + 三类 fallback（错误\u002F政策\u002F上下文）+ 成本追踪 + 内置 admin UI。2026-03 供应链事件后 v1.83+ 强化镜像验证，生产固定签名版本。",[27,36,37],{},"适合：中大型团队 \u002F 合规 \u002F 想完全控数据 + BYOK；微服务架构需要语言无关的 OpenAI 网关；要把成本治理 + 观测做到自建。不适合：不想运维（SaaS 走 OpenRouter \u002F Portkey）；中文支付 \u002F 业务运营（走 one-api \u002F new-api）；纯个人项目（pip 装 SDK 已够，不需要 Proxy）。",[22,39,40],{"id":40},"核心能力",[42,43,44,52,61,71,81,90,96,102,111,117,123,129,135],"ul",{},[45,46,47,51],"li",{},[48,49,50],"strong",{},"100+ providers","：OpenAI \u002F Anthropic \u002F Google \u002F AWS Bedrock \u002F Azure \u002F vLLM \u002F Ollama \u002F Together \u002F HuggingFace 等",[45,53,54,57,58],{},[48,55,56],{},"统一 OpenAI 接口","：所有模型走 ",[31,59,60],{},"\u002Fv1\u002Fchat\u002Fcompletions",[45,62,63,66,67,70],{},[48,64,65],{},"SDK 模式","：",[31,68,69],{},"from litellm import completion","，适合嵌入",[45,72,73,76,77,80],{},[48,74,75],{},"Proxy 模式","：HTTP 服务 + Postgres + admin UI（",[31,78,79],{},"\u002Fui","）",[45,82,83,66,86,89],{},[48,84,85],{},"虚拟 Key",[31,87,88],{},"\u002Fkey\u002Fgenerate"," 给团队 \u002F 服务签发独立 Key + 预算 + 模型白名单",[45,91,92,95],{},[48,93,94],{},"三类 fallback","：错误 \u002F 内容政策 \u002F context window",[45,97,98,101],{},[48,99,100],{},"重试 + 超时 + cooldown","：完整 SRE 配套",[45,103,104,66,107,110],{},[48,105,106],{},"Cost tracking",[31,108,109],{},"\u002Fglobal\u002Fspend\u002Freport"," + admin UI 看每团队 \u002F Key \u002F 模型成本",[45,112,113,116],{},[48,114,115],{},"回调","：Langfuse \u002F Prometheus \u002F Slack 一键挂载",[45,118,119,122],{},[48,120,121],{},"Guardrails","：Presidio PII masking \u002F 自定义内容检查",[45,124,125,128],{},[48,126,127],{},"缓存","：Redis 内置 + 语义缓存",[45,130,131,134],{},[48,132,133],{},"Routing strategy","：cost-based \u002F latency-based \u002F round-robin",[45,136,137,140],{},[48,138,139],{},"MIT 协议","：完全自由商用 + 修改",[22,142,143],{"id":143},"价格",[42,145,146,152,158],{},[45,147,148,151],{},[48,149,150],{},"OSS","：$0；自托管成本 = 1 台 Postgres + 1 台 LiteLLM container（~2GB RAM）",[45,153,154,157],{},[48,155,156],{},"Enterprise","：Custom；SSO \u002F SAML \u002F 审计 \u002F SLA \u002F on-prem 部署支持",[45,159,160],{},"模型成本走各厂商直接结算（BYOK）",[162,163,164],"blockquote",{},[27,165,166],{},"小规模总成本：1 台 2 核 4G VPS 跑 Postgres + LiteLLM 月 $20–30，团队 10 人完全够。",[22,168,170],{"id":169},"实测10-人-saas-微服务架构","实测（10 人 SaaS \u002F 微服务架构）",[27,172,173],{},[48,174,175],{},"亮点：",[42,177,178,181,184,187,190,193,196],{},[45,179,180],{},"Docker Compose 40 分钟拉起完整生产栈",[45,182,183],{},"虚拟 Key + 预算让微服务团队成本归因清晰",[45,185,186],{},"三类 fallback 配齐后可用率从 99.2% → 99.8%",[45,188,189],{},"admin UI 看每团队每天成本省了一堆自研 dashboard",[45,191,192],{},"Postgres + master key 模式，密钥 + 配置一致性高",[45,194,195],{},"与 Langfuse 集成做 trace + cost 双视角",[45,197,198],{},"多语言客户端（Python \u002F Node \u002F Go \u002F Rust）走同一 endpoint 体验一致",[27,200,201],{},[48,202,203],{},"踩坑：",[42,205,206,213,216,219,226],{},[45,207,208,209,212],{},"2026-03 供应链事件让团队对 ",[31,210,211],{},":latest"," tag 警惕，生产必须固定签名版本（如 v1.85.0）",[45,214,215],{},"Postgres salt key 一旦生成不能轮换，初始化前要谨慎备份",[45,217,218],{},"admin UI 早期版本英文为主，中文文档少",[45,220,221,222,225],{},"模型 ID 命名复杂（",[31,223,224],{},"provider\u002Fmodel-name","），各厂商命名规则不一",[45,227,228],{},"自托管 = 自付运维（Postgres backup \u002F 升级 \u002F 监控）",[22,230,232],{"id":231},"上手proxy-模式","上手（Proxy 模式）",[234,235,240],"pre",{"className":236,"code":237,"language":238,"meta":239,"style":239},"language-bash shiki shiki-themes github-light github-dark","mkdir llm-gateway && cd llm-gateway\nopenssl rand -hex 32 > .master_key\nopenssl rand -hex 32 > .salt_key\n\n# docker-compose.yml 拉 ghcr.io\u002Fberriai\u002Flitellm:v1.85.0\n# config.yaml 写 model_list \u002F fallbacks \u002F litellm_settings\n\ndocker compose up -d\n\n# 生成虚拟 Key\ncurl -X POST http:\u002F\u002Flocalhost:4000\u002Fkey\u002Fgenerate \\\n  -H \"Authorization: Bearer $LITELLM_MASTER_KEY\" \\\n  -d '{\"models\":[\"gpt-5.4\",\"claude-sonnet-4.6\"],\"max_budget\":100}'\n\n# 业务方调用\ncurl http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions \\\n  -H \"Authorization: Bearer sk-xxx\" \\\n  -d '{\"model\":\"gpt-5.4\",\"messages\":[...]}'\n","bash","",[31,241,242,266,288,304,311,318,324,329,344,349,355,373,390,399,404,410,420,430],{"__ignoreMap":239},[243,244,247,251,255,259,263],"span",{"class":245,"line":246},"line",1,[243,248,250],{"class":249},"sScJk","mkdir",[243,252,254],{"class":253},"sZZnC"," llm-gateway",[243,256,258],{"class":257},"sVt8B"," && ",[243,260,262],{"class":261},"sj4cs","cd",[243,264,265],{"class":253}," llm-gateway\n",[243,267,269,272,275,278,281,285],{"class":245,"line":268},2,[243,270,271],{"class":249},"openssl",[243,273,274],{"class":253}," rand",[243,276,277],{"class":261}," -hex",[243,279,280],{"class":261}," 32",[243,282,284],{"class":283},"szBVR"," >",[243,286,287],{"class":253}," .master_key\n",[243,289,291,293,295,297,299,301],{"class":245,"line":290},3,[243,292,271],{"class":249},[243,294,274],{"class":253},[243,296,277],{"class":261},[243,298,280],{"class":261},[243,300,284],{"class":283},[243,302,303],{"class":253}," .salt_key\n",[243,305,307],{"class":245,"line":306},4,[243,308,310],{"emptyLinePlaceholder":309},true,"\n",[243,312,314],{"class":245,"line":313},5,[243,315,317],{"class":316},"sJ8bj","# docker-compose.yml 拉 ghcr.io\u002Fberriai\u002Flitellm:v1.85.0\n",[243,319,321],{"class":245,"line":320},6,[243,322,323],{"class":316},"# config.yaml 写 model_list \u002F fallbacks \u002F litellm_settings\n",[243,325,327],{"class":245,"line":326},7,[243,328,310],{"emptyLinePlaceholder":309},[243,330,332,335,338,341],{"class":245,"line":331},8,[243,333,334],{"class":249},"docker",[243,336,337],{"class":253}," compose",[243,339,340],{"class":253}," up",[243,342,343],{"class":261}," -d\n",[243,345,347],{"class":245,"line":346},9,[243,348,310],{"emptyLinePlaceholder":309},[243,350,352],{"class":245,"line":351},10,[243,353,354],{"class":316},"# 生成虚拟 Key\n",[243,356,358,361,364,367,370],{"class":245,"line":357},11,[243,359,360],{"class":249},"curl",[243,362,363],{"class":261}," -X",[243,365,366],{"class":253}," POST",[243,368,369],{"class":253}," http:\u002F\u002Flocalhost:4000\u002Fkey\u002Fgenerate",[243,371,372],{"class":261}," \\\n",[243,374,376,379,382,385,388],{"class":245,"line":375},12,[243,377,378],{"class":261},"  -H",[243,380,381],{"class":253}," \"Authorization: Bearer ",[243,383,384],{"class":257},"$LITELLM_MASTER_KEY",[243,386,387],{"class":253},"\"",[243,389,372],{"class":261},[243,391,393,396],{"class":245,"line":392},13,[243,394,395],{"class":261},"  -d",[243,397,398],{"class":253}," '{\"models\":[\"gpt-5.4\",\"claude-sonnet-4.6\"],\"max_budget\":100}'\n",[243,400,402],{"class":245,"line":401},14,[243,403,310],{"emptyLinePlaceholder":309},[243,405,407],{"class":245,"line":406},15,[243,408,409],{"class":316},"# 业务方调用\n",[243,411,413,415,418],{"class":245,"line":412},16,[243,414,360],{"class":249},[243,416,417],{"class":253}," http:\u002F\u002Flocalhost:4000\u002Fv1\u002Fchat\u002Fcompletions",[243,419,372],{"class":261},[243,421,423,425,428],{"class":245,"line":422},17,[243,424,378],{"class":261},[243,426,427],{"class":253}," \"Authorization: Bearer sk-xxx\"",[243,429,372],{"class":261},[243,431,433,435],{"class":245,"line":432},18,[243,434,395],{"class":261},[243,436,437],{"class":253}," '{\"model\":\"gpt-5.4\",\"messages\":[...]}'\n",[22,439,440],{"id":440},"对比",[442,443,444,465],"table",{},[445,446,447],"thead",{},[448,449,450,454,456,459,462],"tr",{},[451,452,453],"th",{},"维度",[451,455,12],{},[451,457,458],{},"OpenRouter",[451,460,461],{},"One-API",[451,463,464],{},"Portkey",[466,467,468,484,499,514,531,544,558,574,588],"tbody",{},[448,469,470,474,477,480,482],{},[471,472,473],"td",{},"形态",[471,475,476],{},"OSS + Enterprise",[471,478,479],{},"SaaS",[471,481,150],{},[471,483,479],{},[448,485,486,489,492,495,497],{},[471,487,488],{},"协议",[471,490,491],{},"MIT",[471,493,494],{},"–",[471,496,491],{},[471,498,494],{},[448,500,501,504,507,510,512],{},[471,502,503],{},"自托管",[471,505,506],{},"✅",[471,508,509],{},"❌",[471,511,506],{},[471,513,156],{},[448,515,516,519,522,525,528],{},[471,517,518],{},"模型数",[471,520,521],{},"100+",[471,523,524],{},"300+",[471,526,527],{},"30+",[471,529,530],{},"250+",[448,532,533,536,538,540,542],{},[471,534,535],{},"虚拟 Key + 预算",[471,537,506],{},[471,539,494],{},[471,541,506],{},[471,543,506],{},[448,545,546,549,552,554,556],{},[471,547,548],{},"Fallback",[471,550,551],{},"✅ 三类",[471,553,506],{},[471,555,506],{},[471,557,506],{},[448,559,560,563,566,569,572],{},[471,561,562],{},"admin UI",[471,564,565],{},"✅ 内置",[471,567,568],{},"dashboard",[471,570,571],{},"✅ 中文 UI",[471,573,506],{},[448,575,576,579,581,583,586],{},[471,577,578],{},"中文支付",[471,580,509],{},[471,582,509],{},[471,584,585],{},"✅ EPay 内置",[471,587,509],{},[448,589,590,593,596,598,601],{},[471,591,592],{},"集成观测",[471,594,595],{},"Langfuse\u002FProm",[471,597,568],{},[471,599,600],{},"仪表盘",[471,602,603],{},"原生",[22,605,606],{"id":606},"避坑",[42,608,609,617,623,629,635,645,651,665],{},[45,610,611,66,614,616],{},[48,612,613],{},"必固定版本号",[31,615,211],{}," \u002F 滚动 tag 在 2026-03 供应链事件后已是禁忌；用带签名验证的具体版本（如 v1.85.0）",[45,618,619,622],{},[48,620,621],{},"salt_key 不能轮换","：初始化前生成 + 加密备份",[45,624,625,628],{},[48,626,627],{},"Postgres 备份","：所有虚拟 Key + 预算都在 DB，必须定期备份",[45,630,631,634],{},[48,632,633],{},"fallback 链别堆超 3 个","：失败叠加延迟 + 计费混乱",[45,636,637,640,641,644],{},[48,638,639],{},"drop_params 谨慎开","：开 ",[31,642,643],{},"drop_params: true"," 会静默丢不兼容字段，调试时容易 confused",[45,646,647,650],{},[48,648,649],{},"PII guardrail 不是 0 延迟","：Presidio 调用增 50–100ms，敏感场景再用",[45,652,653,656,657,660,661,664],{},[48,654,655],{},"monorepo + workspace 模型映射","：业务方调 ",[31,658,659],{},"gpt-4"," → config 映射到 ",[31,662,663],{},"openai\u002Fgpt-5.4-mini","，要规划稳定映射表",[45,666,667,670],{},[48,668,669],{},"国内自托管","：BYOK Key 走海外 API 仍然受网络影响，需要部署在能直连厂商的节点",[22,672,674],{"id":673},"适合-不适合","适合 \u002F 不适合",[42,676,677,680,683,686,689,692,695,698],{},[45,678,679],{},"✅ 中大型团队 \u002F 微服务架构",[45,681,682],{},"✅ 合规 \u002F 数据主权要求",[45,684,685],{},"✅ 多团队预算治理",[45,687,688],{},"✅ 想叠加 Langfuse \u002F Helicone \u002F Prometheus 观测",[45,690,691],{},"✅ BYOK 模式跨多厂商",[45,693,694],{},"❌ 不想运维（走 OpenRouter \u002F Portkey）",[45,696,697],{},"❌ 国内业务支付 + 中文运营（走 one-api \u002F new-api）",[45,699,700],{},"❌ 纯个人项目（SDK 足够，不需要 Proxy）",[22,702,703],{"id":703},"相关阅读",[42,705,706,713,719,725],{},[45,707,708],{},[709,710,712],"a",{"href":711},"\u002Ftools\u002Fcoding\u002Fapi\u002Fopenrouter","OpenRouter 评测",[45,714,715],{},[709,716,718],{"href":717},"\u002Ftools\u002Fcoding\u002Fapi\u002Fone-api","One-API 评测",[45,720,721],{},[709,722,724],{"href":723},"\u002Ftools\u002Fcoding\u002Fapi\u002Fportkey","Portkey 评测",[45,726,727],{},[709,728,730],{"href":729},"\u002Ftools\u002Fcoding\u002Fapi\u002Fhelicone","Helicone 评测",[22,732,733],{"id":733},"来源",[735,736,737,745,752,759],"ol",{},[45,738,739,740],{},"NerdLevelTech — LiteLLM Proxy Production Tutorial 2026（含 v1.85.0 + 供应链事件细节）",[709,741,742],{"href":742,"rel":743},"https:\u002F\u002Fnerdleveltech.com\u002Flitellm-proxy-production-llm-gateway-tutorial",[744],"nofollow",[45,746,747,748],{},"LiteLLM 官方文档 — Fallbacks \u002F Retries \u002F Cooldowns ",[709,749,750],{"href":750,"rel":751},"https:\u002F\u002Fdocs.litellm.ai\u002Fdocs\u002Fproxy\u002Freliability",[744],[45,753,754,755],{},"Youngju.dev — LiteLLM Complete Guide 2026（SDK vs Proxy \u002F cost-based routing）",[709,756,757],{"href":757,"rel":758},"https:\u002F\u002Fwww.youngju.dev\u002Fblog\u002Fculture\u002F2026-03-25-litellm-unified-llm-api-proxy-guide-2025.en",[744],[45,760,761,762],{},"GitHub — BerriAI\u002Flitellm README ",[709,763,764],{"href":764,"rel":765},"https:\u002F\u002Fgithub.com\u002FBerriAI\u002Flitellm",[744],[767,768,769],"style",{},"html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .sj4cs, html code.shiki .sj4cs{--shiki-default:#005CC5;--shiki-dark:#79B8FF}html pre.shiki code .szBVR, html code.shiki .szBVR{--shiki-default:#D73A49;--shiki-dark:#F97583}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":239,"searchDepth":290,"depth":290,"links":771},[772,773,774,775,776,777,778,779,780,781],{"id":24,"depth":268,"text":25},{"id":40,"depth":268,"text":40},{"id":143,"depth":268,"text":143},{"id":169,"depth":268,"text":170},{"id":231,"depth":268,"text":232},{"id":440,"depth":268,"text":440},{"id":606,"depth":268,"text":606},{"id":673,"depth":268,"text":674},{"id":703,"depth":268,"text":703},{"id":733,"depth":268,"text":733},"api","\u002Fimg\u002Ftools\u002Flitellm.webp","LiteLLM 真实评测：MIT 开源 LLM 网关，BerriAI 维护。SDK 模式（pip install litellm）+ Proxy 模式（Docker Compose 自托管）双形态，100+ 厂商统一 OpenAI 接口，虚拟 Key + 团队预算 + 三类 fallback + 成本追踪 + 内置 UI。2026-03 经历供应链事件后 v1.83+ 强化签名 + 镜像验证，生产请固定版本号。",false,"md",[788,791,794,797],{"q":789,"a":790},"SDK 和 Proxy 模式区别？","SDK 模式（`from litellm import completion`）直接在 Python 里调，适合个人项目 \u002F 单服务。Proxy 模式启动一个 HTTP 服务（默认 4000 端口）+ Postgres，所有应用通过 OpenAI 兼容 endpoint 调用，支持虚拟 Key \u002F 团队 \u002F 预算 \u002F 配额 \u002F admin UI——团队 \u002F 多语言客户端 \u002F 生产建议走 Proxy。",{"q":792,"a":793},"Fallback 有几种？","三类：(1) general fallback——5xx\u002F429 错误切到备用模型；(2) content policy fallback——内容审查拒绝时切；(3) context window fallback——输入超 context 切大窗口模型（比如超 GPT-4 32k 自动切 Claude 200k）。可叠加 num_retries \u002F cooldown \u002F timeout 做完整 SRE 策略。",{"q":795,"a":796},"2026-03 供应链事件是什么？","2026-03-24 PyPI 上的 v1.82.7 \u002F v1.82.8 在被替换约 40 分钟内可能植入 Trivy CI token 外泄代码。BerriAI 在 v1.83+ 强化签名 + 镜像验证，生产部署务必固定到带签名验证的版本（如 v1.85.0），不要用 `:latest` tag。",{"q":798,"a":799},"和 OpenRouter 怎么选？","OpenRouter = SaaS（不用自己运维 + 直接付钱）；LiteLLM = OSS 自托管（自付服务器 + 完全控制数据 + BYOK 所有 Key）。中大型团队 \u002F 合规 \u002F 长期成本敏感 \u002F 自建 → LiteLLM Proxy；小团队 \u002F 不想运维 \u002F 快速迭代 → OpenRouter。",[801],"en",{},"\u002Ftools\u002Fcoding\u002Fapi\u002Flitellm","coding",[806,807,334,808],"sdk","proxy","kubernetes",[810,815],{"plan":811,"price":812,"features":813,"notes":814},"Open Source","$0（MIT）","SDK + Proxy + Postgres 虚拟 Key + 预算 + fallback + admin UI + 100+ providers","完全自托管，模型 \u002F 平台费走自付",{"plan":156,"price":816,"features":817,"notes":818},"Custom","SSO \u002F SAML + Audit + JWT auth + 高级路由 + SLA + on-prem 部署支持","联系 BerriAI 销售","MIT 开源免费 \u002F Enterprise SaaS 定制","2026-06-19",[822],"onboarding\u002Frag-pipeline-build",[824],"llm-gateway-comparison",{"power":313,"ux":306,"price":313,"cn_support":290,"stability":306},{"title":12,"description":784},"coding\u002Fapi\u002Flitellm",[829,832,834,836],{"name":830,"url":742,"accessed":831},"NerdLevelTech — LiteLLM Proxy 生产教程 2026","2026-06-24",{"name":833,"url":750,"accessed":831},"LiteLLM 官方文档 — Fallbacks",{"name":835,"url":757,"accessed":831},"Youngju.dev — LiteLLM Complete Guide 2026",{"name":837,"url":764,"accessed":831},"GitHub — BerriAI\u002Flitellm","tools\u002Fcoding\u002Fapi\u002Flitellm","MIT 开源 LLM 网关——SDK + Proxy 双形态，100+ 厂商统一 OpenAI 接口 + 虚拟 Key + 团队预算",[841,807,806,842,843,844,845,846],"llm-gateway","opensource","mit","fallback","virtual-keys","byok","想要『SaaS 体验 + 完全自托管』的开源 LLM 网关首选。组合『LiteLLM 网关 + Helicone\u002FLangfuse 观测』在中大型 dev org 是黄金组合。中文中转 + 业务支付走 one-api。","https:\u002F\u002Fwww.litellm.ai","JD7TeoQ-by7hjinNw1E4qfhTneex-5pJiPI0H8uth5Y",{"id":851,"title":458,"alternatives":852,"api_compatible":9,"body":853,"category":782,"chinese_friendly":268,"cover":1314,"description":1315,"domestic":785,"extension":786,"faq":1316,"free":785,"github":9,"languages":1329,"lastVerified":9,"meta":1331,"models":9,"navigation":309,"notSuitable":9,"opensource":785,"path":711,"pillar":804,"platforms":1332,"priceTable":1333,"pricing":1349,"published":820,"relatedPlaybooks":1350,"relatedReviews":1352,"score":1353,"self_host":785,"seo":1354,"seoTitle":9,"slug":14,"sources":1355,"stem":1364,"suitable":9,"tagline":1365,"tags":1366,"updated":831,"verdict":1371,"website":1372,"__hash__":1373},"tools\u002Ftools\u002Fcoding\u002Fapi\u002Fopenrouter.md",[827,15,16,17],{"type":19,"value":854,"toc":1302},[855,857,860,863,865,939,941,971,976,980,984,1001,1006,1030,1033,1065,1067,1178,1180,1218,1220,1249,1251,1270,1272],[22,856,25],{"id":24},[27,858,859],{},"OpenRouter 是 AI API 聚合先驱——一个 OpenAI 兼容 endpoint + 一个 Key，调 60+ 厂商 300+ 模型，5M+ 用户。模型 passthrough 定价 0 markup，但充值有 5.5% 信用卡平台费（$0.80 最低）+ BYOK 月 100 万请求后 5%。自动 fallback \u002F playground \u002F 海量模型选型是杀手锏。国内直连延迟 1500–3000ms + 不支持支付宝是硬伤。",[27,861,862],{},"适合：海外开发者；多模型选型 \u002F 横评；生产 fallback 高可用；零迁移成本（OpenAI SDK 改 baseURL）。不适合：国内对延迟敏感的 AI 编程工具（用 OfoxAI \u002F 自建中转）；要支付宝 \u002F 微信 \u002F 银联；要 self-host（用 one-api \u002F LiteLLM）。",[22,864,40],{"id":40},[42,866,867,873,882,893,903,909,915,921,927,933],{},[45,868,869,872],{},[48,870,871],{},"60+ 厂商 300+ 模型","：OpenAI \u002F Anthropic \u002F Google \u002F Meta \u002F DeepSeek \u002F MiniMax \u002F Mistral \u002F Cohere \u002F xAI \u002F NVIDIA \u002F Qwen \u002F 字节豆包 等",[45,874,875,66,878,881],{},[48,876,877],{},"OpenAI 兼容 endpoint",[31,879,880],{},"https:\u002F\u002Fopenrouter.ai\u002Fapi\u002Fv1","，SDK 改 baseURL 即用",[45,883,884,66,887,889,890,80],{},[48,885,886],{},"模型 ID 格式",[31,888,224],{},"（如 ",[31,891,892],{},"anthropic\u002Fclaude-sonnet-4.6",[45,894,895,898,899,902],{},[48,896,897],{},"自动 fallback","：请求里列 ",[31,900,901],{},"models: [...]","，前者失败自动切后者",[45,904,905,908],{},[48,906,907],{},"Playground","：5 模型并排测同 prompt + 成本 + 延迟 + 输出",[45,910,911,914],{},[48,912,913],{},"BYOK","：自带 OpenAI \u002F Anthropic Key，OpenRouter 只做网关；前 100 万 req\u002F月免费，超 5% fee",[45,916,917,920],{},[48,918,919],{},"Passthrough 定价","：模型按厂商官方价",[45,922,923,926],{},[48,924,925],{},"免费模型","：Llama 4 Scout \u002F DeepSeek V4 Flash \u002F Qwen3 Coder \u002F Gemma 3 等，每日 5\u002F200 次",[45,928,929,932],{},[48,930,931],{},"Provider ranking \u002F SLA","：dashboard 看每家 provider 延迟 + 错误率",[45,934,935,938],{},[48,936,937],{},"Rankings 页","：模型按 token 消耗排名（市场热度参考）",[22,940,143],{"id":143},[42,942,943,949,955,961,966],{},[45,944,945,948],{},[48,946,947],{},"Free","：$1 credit + 免费模型 5\u002F日 + 标准限流",[45,950,951,954],{},[48,952,953],{},"Pay-as-you-go","：模型 passthrough + 5.5% 卡费（最低 $0.80）+ 充值 $5 后免费模型涨到 200\u002F日",[45,956,957,960],{},[48,958,959],{},"Team","：$99\u002F月 + Priority routing + 99.9% SLA",[45,962,963,965],{},[48,964,156],{},"：Custom 专属基础设施",[45,967,968,970],{},[48,969,913],{},"：前 100 万 req\u002F月免费，超出 5%",[162,972,973],{},[27,974,975],{},"实际总成本 ≈ 模型价 × (1 + 5.5%) + BYOK 5% × 月 >1M 部分。预算按 5–7% overhead 算。",[22,977,979],{"id":978},"实测小型-saas-后端-海外节点","实测（小型 SaaS 后端 \u002F 海外节点）",[27,981,982],{},[48,983,175],{},[42,985,986,989,992,995,998],{},[45,987,988],{},"5 分钟接入，OpenAI SDK 改两行配置",[45,990,991],{},"Playground 横评新模型省一大堆 boilerplate",[45,993,994],{},"自动 fallback 救命：Anthropic 4 小时宕机时应用零中断",[45,996,997],{},"模型 ID 提示器 + Rankings 看市场热度方便选型",[45,999,1000],{},"海外节点延迟 300–600ms，CDN 体感好",[27,1002,1003],{},[48,1004,1005],{},"踩坑（国内）：",[42,1007,1008,1011,1014,1017,1024,1027],{},[45,1009,1010],{},"国内直连 1500–3000ms + 高峰超时率明显，Cursor \u002F Claude Code 高频调用体感糟",[45,1012,1013],{},"不支持支付宝 \u002F 微信 \u002F 银联，只 Visa\u002FMastercard \u002F USDC",[45,1015,1016],{},"信用卡费 5.5% 最低 $0.80，\u003C$15 充值实际打到 10%+",[45,1018,1019,1020,1023],{},"模型 ID 点号分隔（",[31,1021,1022],{},"claude-sonnet-4.6","）易跟连字符混",[45,1025,1026],{},"Claude Code 默认走 Anthropic 原生协议，OpenRouter 适配偶有不一致",[45,1028,1029],{},"「免费模型」rate limit 后可能静默切付费",[22,1031,1032],{"id":1032},"上手",[735,1034,1035,1038,1041,1044,1050,1056,1062],{},[45,1036,1037],{},"openrouter.ai → Google \u002F GitHub OAuth 注册",[45,1039,1040],{},"Keys → 创建 API Key（sk-or-xxx）",[45,1042,1043],{},"充值 $10+（Visa \u002F Mastercard \u002F USDC）→ 解锁更高免费配额",[45,1045,1046,1047],{},"代码：",[31,1048,1049],{},"OpenAI(baseURL='https:\u002F\u002Fopenrouter.ai\u002Fapi\u002Fv1', api_key=KEY)",[45,1051,1052,1053,1055],{},"模型用 ",[31,1054,224],{}," 格式",[45,1057,1058,1059],{},"加 fallback：",[31,1060,1061],{},"extra_body={\"models\": [\"anthropic\u002Fclaude-sonnet-4.6\", \"openai\u002Fgpt-5.4\"]}",[45,1063,1064],{},"Playground 横评再批量切",[22,1066,440],{"id":440},[442,1068,1069,1083],{},[445,1070,1071],{},[448,1072,1073,1075,1077,1079,1081],{},[451,1074,453],{},[451,1076,458],{},[451,1078,12],{},[451,1080,461],{},[451,1082,464],{},[466,1084,1085,1098,1110,1122,1134,1146,1162],{},[448,1086,1087,1089,1091,1094,1096],{},[471,1088,473],{},[471,1090,479],{},[471,1092,1093],{},"OSS + SaaS",[471,1095,150],{},[471,1097,479],{},[448,1099,1100,1102,1104,1106,1108],{},[471,1101,518],{},[471,1103,524],{},[471,1105,521],{},[471,1107,527],{},[471,1109,530],{},[448,1111,1112,1114,1116,1118,1120],{},[471,1113,503],{},[471,1115,509],{},[471,1117,506],{},[471,1119,506],{},[471,1121,156],{},[448,1123,1124,1126,1128,1130,1132],{},[471,1125,578],{},[471,1127,509],{},[471,1129,494],{},[471,1131,506],{},[471,1133,494],{},[448,1135,1136,1138,1140,1142,1144],{},[471,1137,548],{},[471,1139,506],{},[471,1141,506],{},[471,1143,506],{},[471,1145,506],{},[448,1147,1148,1151,1154,1157,1159],{},[471,1149,1150],{},"国内延迟",[471,1152,1153],{},"1500–3000ms",[471,1155,1156],{},"自托管自定",[471,1158,1156],{},[471,1160,1161],{},"海外为主",[448,1163,1164,1167,1170,1173,1175],{},[471,1165,1166],{},"成本",[471,1168,1169],{},"模型 + 5.5%\u002FBYOK 5%",[471,1171,1172],{},"自付服务器",[471,1174,1172],{},[471,1176,1177],{},"按 logs 计费",[22,1179,606],{"id":606},[42,1181,1182,1188,1194,1200,1206,1212],{},[45,1183,1184,1187],{},[48,1185,1186],{},"国内别裸用","：1500ms+ 起、超时率高，主链路上 OfoxAI \u002F 自建 one-api 中转，OpenRouter 当冷门模型补",[45,1189,1190,1193],{},[48,1191,1192],{},"预算多算 5–7%","：5.5% 信用卡 + BYOK 5% + 偶发汇率",[45,1195,1196,1199],{},[48,1197,1198],{},"免费模型生产慎用","：rate limit 后可能静默 fallback 付费，账单失控",[45,1201,1202,1205],{},[48,1203,1204],{},"模型 ID 格式坑","：点号 vs 连字符要测；新模型 ID 改名要监控",[45,1207,1208,1211],{},[48,1209,1210],{},"fallback 别堆太多","：5+ fallback 一次失败延迟叠加可能 10s+，前 2 个即可",[45,1213,1214,1217],{},[48,1215,1216],{},"Claude Code 原生模式建议直连","：OpenRouter 走 OpenAI 兼容协议有时与原生 Anthropic 协议有出入",[22,1219,674],{"id":673},[42,1221,1222,1225,1228,1231,1234,1237,1240,1243,1246],{},[45,1223,1224],{},"✅ 海外开发者 \u002F 跨境业务",[45,1226,1227],{},"✅ 多模型选型 \u002F 横评",[45,1229,1230],{},"✅ 生产 fallback 高可用",[45,1232,1233],{},"✅ 零迁移成本（OpenAI SDK 改 baseURL）",[45,1235,1236],{},"✅ BYOK 模式给团队统一 Key 管理",[45,1238,1239],{},"❌ 国内 AI 编程工具（Cursor \u002F Claude Code）高频调用",[45,1241,1242],{},"❌ 要支付宝 \u002F 微信 \u002F 银联",[45,1244,1245],{},"❌ Self-host \u002F 完全控制数据",[45,1247,1248],{},"❌ 极致低延迟实时应用",[22,1250,703],{"id":703},[42,1252,1253,1258,1262,1266],{},[45,1254,1255],{},[709,1256,1257],{"href":803},"LiteLLM 评测",[45,1259,1260],{},[709,1261,718],{"href":717},[45,1263,1264],{},[709,1265,724],{"href":723},[45,1267,1268],{},[709,1269,730],{"href":729},[22,1271,733],{"id":733},[735,1273,1274,1281,1288,1295],{},[45,1275,1276,1277],{},"OfoxAI — OpenRouter 完全指南 2026（国内延迟实测 + 替代方案）",[709,1278,1279],{"href":1279,"rel":1280},"https:\u002F\u002Fofox.ai\u002Fzh\u002Fblog\u002Fopenrouter-complete-guide-china-developers-2026",[744],[45,1282,1283,1284],{},"OfoxAI — OpenRouter Pricing 2026 Hidden Markup Breakdown ",[709,1285,1286],{"href":1286,"rel":1287},"https:\u002F\u002Fofox.ai\u002Fblog\u002Fopenrouter-pricing-hidden-markup-breakdown-2026\u002F",[744],[45,1289,1290,1291],{},"DigitalApplied — OpenRouter June 2026 New Models & Pricing Roundup ",[709,1292,1293],{"href":1293,"rel":1294},"https:\u002F\u002Fwww.digitalapplied.com\u002Fblog\u002Fopenrouter-new-models-june-2026-roundup-pricing-rankings",[744],[45,1296,1297,1298],{},"AIToolBox — OpenRouter Review 2026 ",[709,1299,1300],{"href":1300,"rel":1301},"https:\u002F\u002Fwww.aitoolbox.hk\u002Ftools\u002Fopen-router\u002F",[744],{"title":239,"searchDepth":290,"depth":290,"links":1303},[1304,1305,1306,1307,1308,1309,1310,1311,1312,1313],{"id":24,"depth":268,"text":25},{"id":40,"depth":268,"text":40},{"id":143,"depth":268,"text":143},{"id":978,"depth":268,"text":979},{"id":1032,"depth":268,"text":1032},{"id":440,"depth":268,"text":440},{"id":606,"depth":268,"text":606},{"id":673,"depth":268,"text":674},{"id":703,"depth":268,"text":703},{"id":733,"depth":268,"text":733},"\u002Fimg\u002Ftools\u002Fopenrouter.webp","OpenRouter 真实评测：AI API 聚合先驱，60+ 厂商 300+ 模型一个 OpenAI 兼容 endpoint。passthrough 定价 0 markup + 5.5% 信用卡平台费（$0.80 最低）+ BYOK 5% 月千万请求外。Pay-as-you-go \u002F Team $99\u002F月 SLA \u002F Enterprise。国内直连 1500–3000ms 高峰超时多。",[1317,1320,1323,1326],{"q":1318,"a":1319},"国内能用吗？","技术上能（OpenAI SDK 改 baseURL），但服务器在海外、无国内节点：实测首 Token 延迟代理 600–1200ms \u002F 直连 1500–3000ms 且不稳，高峰超时率高。AI 编程类高频调用（Cursor \u002F Claude Code \u002F Cline）国内体验糟糕，主流方案是用 OfoxAI \u002F 自建 one-api\u002Fnew-api 中转。",{"q":1321,"a":1322},"真的零 markup 吗？","模型本身按厂商官方价 passthrough，技术上 0 markup。但充值有两层平台费：信用卡 5.5%（最低 $0.80，小额充值实际打到 10–20%）、BYOK 月 100 万请求后 5%。预算时按上层多算 5–7%。",{"q":1324,"a":1325},"自动 fallback 怎么用？","请求里加 `models: ['anthropic\u002Fclaude-sonnet-4.6', 'openai\u002Fgpt-5.4']`，前者失败 \u002F 超时 \u002F 限流时自动切到后者。生产 24h+ 不间断 AI 服务几乎刚需——Anthropic 2026 年 5 月 4 小时宕机，用 fallback 的应用零中断。",{"q":1327,"a":1328},"免费模型怎么用？","模型列表筛选 $0 价（如 Llama 4 Scout \u002F DeepSeek V4 Flash \u002F Qwen3 Coder \u002F Gemma 3）。未充值每日 5 次，充 $5+ 涨到 200 次\u002F日，附带 20 req\u002Fmin。生产别依赖免费模型，限流 \u002F 排队 \u002F 暗切付费会让计费失控。",[801,1330],"zh",{},[782,806],[1334,1338,1342,1346],{"plan":947,"price":1335,"features":1336,"notes":1337},"$0","$1 credit + 免费模型 5\u002F日 + 标准限流","不需要信用卡",{"plan":953,"price":1339,"features":1340,"notes":1341},"Passthrough + 5.5% 卡费","300+ 模型 + 标准限流 + 充 $5 后免费模型 200\u002F日 + BYOK","信用卡平台费最低 $0.80",{"plan":959,"price":1343,"features":1344,"notes":1345},"$99\u002F月","Priority routing + 99.9% SLA + 更高限流","面向生产应用",{"plan":156,"price":816,"features":1347,"notes":1348},"专属基础设施 + 自定义合同","大规模生产","免费 $1 credit + 5\u002F日 → 200\u002F日（充 $5+） \u002F Pay-as-you-go passthrough + 5.5% 卡费 \u002F Team $99\u002F月 SLA \u002F Enterprise",[822,1351],"onboarding\u002Fclaude-code-getting-started",[824],{"power":313,"ux":306,"price":306,"cn_support":268,"stability":306},{"title":458,"description":1315},[1356,1358,1360,1362],{"name":1357,"url":1279,"accessed":831},"OfoxAI — OpenRouter 完全指南 2026",{"name":1359,"url":1286,"accessed":831},"OfoxAI — OpenRouter Pricing Hidden Markup",{"name":1361,"url":1293,"accessed":831},"DigitalApplied — OpenRouter June 2026 Roundup",{"name":1363,"url":1300,"accessed":831},"AIToolBox — OpenRouter Review 2026","tools\u002Fcoding\u002Fapi\u002Fopenrouter","AI API 聚合先驱——一个 Key + 一个 endpoint 调 300+ 模型，自动 fallback，OpenAI 兼容",[1367,1368,1369,844,1370,846],"api-gateway","aggregator","openai-compatible","multi-model","AI API 聚合的全球事实标准——300+ 模型 \u002F 一个 Key \u002F 自动 fallback。海外开发者首选；国内开发者延迟 1500–3000ms + 不支持支付宝是硬伤，主走 OfoxAI \u002F one-api 自建中转。","https:\u002F\u002Fopenrouter.ai","baZ8onQahFHHQRNSweqSrYSNISBgcVfv7vmfrkme054",1784565441108]