{"auto_groups":["guochan","tj_claude","tj_gemini","tj_openai","tj_grok","tj_openai_az","tj_claude_max","stable","stable_pic","official"],"data":[{"model_name":"glm-5.3","description":"GLM‑5.3 是智谱AI于2026年8月14日发布的高性能MoE旗舰模型。总参数量1.4万亿（激活62B），支持1M token上下文。主打顶尖Agent工程能力，在Agentic Coding评测中稳居开源第一梯队，内部体验优于Opus 5，推理性能比肩世界顶级闭源模型。","tags":"chat","vendor_id":31,"quota_type":0,"model_ratio":0.7,"model_price":0,"owner_by":"","completion_ratio":3.142857142857,"cache_ratio":0.185714285714,"enable_groups":["guochan","guochan_自部署","official","default"],"supported_endpoint_types":["openai"],"pricing_version":"5a90f2b86c08bd983a9a2e6d66c255f4eaef9c4bc934386d2b6ae84ef0ff1f1f"},{"model_name":"hy4-preview","description":"Hy4‑preview是腾讯混元团队于2026年8月28日发布的高性能MoE开源旗舰模型，总参数量770B（激活49B），支持1M token上下文，主打生产力导向长周期Agent与全栈工程能力，在Agentic Coding、复杂办公科研评测稳居开源第一梯队，内部体验优于Kimi‑K3，推理性能比肩世界顶级闭源模型。","tags":"chat","vendor_id":51,"quota_type":0,"model_ratio":0.417,"model_price":0,"owner_by":"","completion_ratio":2.998800959233,"cache_ratio":0.05035971223,"enable_groups":["guochan","official"],"supported_endpoint_types":["openai"]},{"model_name":"kimi-k2.6","description":"Kimi K2.6是月之暗面于2026年4月20日发布并开源的MoE架构旗舰模型。总参数1万亿（激活320亿），支持256K上下文，具备视觉理解能力。主打长周期编码（持续13小时/4000+行代码）与大规模Agent集群（最多300个子Agent并行4000步），在多项基准测试中持平或超越GPT-5.4、Claude Opus 4.6等闭源模型。","tags":"chat","vendor_id":30,"quota_type":0,"model_ratio":0.475,"model_price":0,"owner_by":"","completion_ratio":4.210526315789,"cache_ratio":0.168421052632,"enable_groups":["guochan","guochan_便宜","official"],"supported_endpoint_types":["openai"]},{"model_name":"qwen3.8-max","description":"Qwen3.8-Max 是阿里云通义千问团队于 2026 年 8 月 3 日发布的高性能 MoE 旗舰模型。总参数量 2.4 万亿（激活 95B），支持 1M token 上下文。主打顶尖长程 Agent 与自主工程能力，在 Agentic Coding、多模态智能体评测中登顶开源榜首，内部体验比肩 Fable 5，推理性能对标世界顶级闭源模型。","tags":"chat","vendor_id":33,"quota_type":0,"model_ratio":1,"model_price":0,"owner_by":"","completion_ratio":3,"cache_ratio":0.125,"create_cache_ratio":1.25,"enable_groups":["guochan","official"],"supported_endpoint_types":["openai"]},{"model_name":"gpt-image-2.5-sunburst","description":"gpt‑image‑2.5‑sunburst是OpenAI 2026‑09‑09发布的专业精品图像生成模型，为2.5系列高端版本，最高输出4K分辨率，牺牲部分速度换取极致细节与精准局部编辑，多轮修改不易失真，强化草图输入、批注修图能力，主打品牌广告、商业精修成品创作，内置C2PA溯源水印，通过API提供多档质量可调，面向专业创意生产场景调用。","tags":"image","vendor_id":38,"quota_type":1,"model_ratio":0,"model_price":0.15,"owner_by":"","completion_ratio":0,"enable_groups":["tj_openai","tj_openai_az"],"supported_endpoint_types":["openai"]},{"model_name":"doubao-seedance-2-0-260128","description":"doubao-seedance-2-0-260128 是字节跳动 Seed 团队于 2026 年 2 月 12 日发布的旗舰级工业向多模态视频生成模型，官方称其为“具备电影级质感与全链路可控性的新一代视频创作模型”。其最大突破在于采用统一多模态音视频联合生成架构，可完成多镜头长叙事短剧制作、电商广告全链路生成、影视分镜动态预览及参考素材定向编辑等高复杂度创作任务。在 Arena.AI 视频生成评测中，Seedance 2.0 文生视频得分 95.2%（远超 1.5 Pro 的 82.7%），复杂运动场景可用率测试取得 91.7% 的成绩，比肩专业后期制作团队产出水准。定价相较 1.5 Pro 上浮约 35%，同时将多模态输入承载扩展至9张参考图+3段视频+3段音频，可一次性承接完整长剧情脚本、多素材混剪指令或全链路商业视频创作需求。","tags":"video","vendor_id":45,"quota_type":0,"model_ratio":37.5,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["video"],"supported_endpoint_types":["openai-video"],"billing_mode":"tiered_expr","billing_expr":"(tier(\"文生视频\", p * 0 + c * 6.58)) * (param(\"reference_video_urls\") != nil ? 0.6 : 1) * (param(\"resolution\") == \"1080p\" ? 1.1 : 1) * (param(\"reference_video_urls\") != nil \u0026\u0026 param(\"resolution\") == \"1080p\" ? 0.67 : 1) * (param(\"resolution\") == \"4k\" ? 0.57 : 1) * (param(\"reference_video_urls\") != nil \u0026\u0026 param(\"resolution\") == \"4k\" ? 0.35 : 1)"},{"model_name":"claude-fable-5-1","description":"claude‑fable‑5‑1 是Anthropic于2026‑09‑01发布的旗舰推理Agent大模型，拥有1M上下文窗口、最大128K输出，输入$10/MTok、输出$50/MTok，缓存读取成本大幅下降75%，同权重受限版本为邀请制的claude‑mythos‑5‑1，主打长周期智能体任务、跨仓库工程编码、多步骤科研推理与复杂企业知识工作，适合Opus‑5能力不足的高难度场景，同时带来工具调用进度更新、每条消息推理强度调节等Beta新特性，也存在部分API破坏性变更。","tags":"chat","vendor_id":37,"quota_type":0,"model_ratio":5,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.025,"create_cache_ratio":1.25,"enable_groups":["default","tj_claude_max"],"supported_endpoint_types":["openai","anthropic"]},{"model_name":"gpt-image-2","description":"GPT-Image-2是OpenAI于2026年4月22日发布的图像生成模型。最大特点是首次引入**思维链推理能力**，能自检输出并联网搜索，实现复杂指令精准遵循与多语言（含中文）高保真文字渲染。支持多轮编辑、风格一致性及最高2K分辨率，在Image Arena榜单登顶并大幅领先。","tags":"image","vendor_id":38,"quota_type":1,"model_ratio":0,"model_price":0.15,"owner_by":"","completion_ratio":0,"enable_groups":["official","tj_openai","tj_openai_az"],"supported_endpoint_types":["openai"]},{"model_name":"grok-4.6","description":"Grok‑4.6 是 xAI 于2026年8月12日发布的高性能MoE旗舰模型。总参数量1.5万亿（激活120B），支持500K token上下文。主打长周期Agent能力，在长链路智能体编程评测中位居闭源第一梯队，内部体验优于GPT‑5.6‑Terra，推理性能比肩世界顶级闭源模型。","tags":"chat","vendor_id":49,"quota_type":0,"model_ratio":0.5,"model_price":0,"owner_by":"","completion_ratio":2,"enable_groups":["official","tj_grok"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003c= 200000 ? tier(\"短文本\", p * 2 + c * 6 + cr * 0.5) : tier(\"长文本\", p * 4 + c * 12 + cr * 1)"},{"model_name":"claude-sonnet-5","description":"Claude Sonnet 5是Anthropic于2026年6月30日发布的Sonnet系列最新Agent模型，性能逼近旗舰级Opus 4.8。 作为迄今为止“最具Agent能力”的Sonnet模型，Sonnet 5支持100万tokens超长上下文与128,000 tokens最大输出，并引入可调节推理努力程度（低/中/高/最高/超高）。核心突破是率先在Sonnet级别实现接近旗舰级的自主Agent能力——能够制定计划、调用浏览器与终端等工具、并在无人工干预下独立完成多步骤任务。在SWE-bench Pro智能体编码测试中得分63.2%，GDPval-AA v2专业办公达1615分（Opus 4.8为1618分），并在Terminal-Bench 2.1编码评估中以80.4%的成绩逼近Opus 4.8的82.7%","tags":"chat","vendor_id":37,"quota_type":0,"model_ratio":1,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.1,"create_cache_ratio":1.25,"enable_groups":["stable","tj_claude"],"supported_endpoint_types":["anthropic","openai"]},{"model_name":"claude-sonnet-5-5","description":"Claude Sonnet 5.5 是 Anthropic2026‑09‑22 发布的均衡主力模型，上下文窗口 1M tokens，最大输出 128K tokens，速度较 Sonnet 5 提升 45%。API 输入 $0.8、输出 $4 每百万 token，缓存读取仅 $0.08/MTok，性价比突出。兼顾推理、代码、多轮 Agent 与长文档处理，综合能力接近 Opus 5.5，只是极限复杂推理略弱。支持结构化输出、工具调用，部署于 Claude 官网、AWS、Azure，适合企业日常业务、批量文档解析、中小型代码工程等高吞吐场景","tags":"chat","vendor_id":37,"quota_type":0,"model_ratio":1,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.1,"create_cache_ratio":1.25,"enable_groups":["tj_claude_max"],"supported_endpoint_types":["openai","anthropic"]},{"model_name":"codex-auto-review","description":"codex代码审核专用模型，gpt5.4差不多吧","tags":"caht","vendor_id":38,"quota_type":0,"model_ratio":1.25,"model_price":0,"owner_by":"","completion_ratio":6,"cache_ratio":0.1,"enable_groups":["tj_openai"],"supported_endpoint_types":["openai"]},{"model_name":"gpt-5.6-terra","description":"GPT-5.6-Terra是OpenAI于2026年6月26日发布的均衡主力MoE多模态模型，全系标配150M token超长上下文窗口，保留单链深度推理与全品类通用多模态能力，主打普惠型通用生产力Agent，常规编程、长文档处理综合评测成绩小幅超越Claude Fable 5，兼顾推理速度与调用成本，以减半资费实现上代旗舰级综合体验，综合落地效率稳居全球均衡闭源模型第一梯队。","tags":"chat","vendor_id":38,"quota_type":0,"model_ratio":37.5,"model_price":0,"owner_by":"","completion_ratio":8,"enable_groups":["stable","tj_openai","tj_openai_az","official"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003e 272000 ? tier(\"长文本\", p * 4 + c * 18 + cr * 0.4 + cc * 5) : tier(\"短文本\", p * 2 + c * 12 + cr * 0.2 + cc * 2.5)"},{"model_name":"glm-5.2","description":"GLM-5.2是智谱AI于2026年6月17日开源的新一代MoE架构旗舰模型，总参数753B至744B（激活约40B），支持1M无损上下文。核心突破是专攻“长程任务”——让AI能像人一样连续工作、自主完成大型工程，在FrontierSWE测试中得分74.4（超越GPT-5.5的72.6），Terminal-Bench 2.1得分81.0（较前代提升17.5个百分点），Code Arena盲测中取得全球可用模型第一。模型采用MIT协议开源，已适配华为昇腾等国产算力平台。","tags":"chat","vendor_id":31,"quota_type":0,"model_ratio":0.7,"model_price":0,"owner_by":"","completion_ratio":3.142857142857,"cache_ratio":0.185714285714,"enable_groups":["guochan","guochan_便宜","guochan_自部署","official"],"supported_endpoint_types":["openai"]},{"model_name":"MiniMax-M3","description":"MiniMax M3是稀宇科技于2026年9月15日发布的MoE架构下一代Agent模型，总参数5600亿（激活220亿），支持1M超长上下文。核心突破是率先实现“多模态情境推理与自主工具调用”，在SWE-Pro编码测试中得分68.34%，GDPval-AA专业办公ELO达1680（全模型最高）。\n\n","tags":"chat","vendor_id":32,"quota_type":0,"model_ratio":0.3,"model_price":0,"owner_by":"","completion_ratio":4,"cache_ratio":0.2,"enable_groups":["guochan","guochan_便宜","official"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003e 512000 ? tier(\"长文本\", p * 0.6 + c * 2.4 + cr * 0.12) : tier(\"短文本\", p * 0.3 + c * 1.2 + cr * 0.06)"},{"model_name":"gemini-3-pro-image","description":"Nano Banana pro是Google Gemini系列中的下一代AI图像生成与编辑模型,是Gemini 2.5 Flash Image(Nano Banana)的升级版。该模型融合多模态Transformer和扩散模型,原生支持2K(2048×2048)及4K分辨率输出,在图像质量、文字渲染、物理推理等方面实现显著提升。\n\n","tags":"image","vendor_id":36,"quota_type":1,"model_ratio":0,"model_price":0.24,"owner_by":"","completion_ratio":0,"enable_groups":["official","tj_gemini"],"supported_endpoint_types":["openai","gemini"]},{"model_name":"gemini-3.8-flash","description":"gemini‑3.8‑flash是Google DeepMind于2026‑09‑02发布的高速主力多模态模型，基于3.7‑flash迭代而来，拥有1M token上下文窗口、最大64K输出，知识截止2026‑03，支持high/medium/low三档推理强度调节，促销价输入$0.75/MTok、输出$3.75/MTok，缓存读取低至$0.075/MTok，主打高吞吐Agent、长周期软件工程、多步骤专业知识工作流，支持图文音视频输入，同基座衍生出邀请制安全专项版本gemini‑3.8‑flash‑cyber，面向漏洞检测与自动补丁场景。","tags":"chat","vendor_id":36,"quota_type":0,"model_ratio":0.75,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.1,"enable_groups":["stable","tj_gemini","official"],"supported_endpoint_types":["openai","gemini"]},{"model_name":"gpt-6-luna","description":"GPT‑6‑Luna 是 OpenAI 2026‑09‑22 发布的 GPT‑6 系列高吞吐经济型模型，上下文 105 万 tokens，最大输出 128k tokens，知识截止 2026‑04‑20，支持图文输入、函数调用与提示缓存。继承 Astra 轻量化能力，保留基础计算机操作、轻量 Agent 与代码能力，成本远低于 Sol 与 Astra。适合批量文档处理、日常自动化、高频业务调用，极限深度推理与复杂工程能力较弱，免费用户可在桌面端体验，API 面向企业与付费用户开放。","tags":"chat","vendor_id":38,"quota_type":0,"model_ratio":0.5,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["official"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"len \u003c= 272000 ? tier(\"短文本\", p * 0.1 + c * 0.5 + cr * 0.01 + cc * 0.13) : tier(\"tier_2\", p * 0.2 + c * 0.75 + cr * 0.02 + cc * 0.25)"},{"model_name":"claude-opus-5","description":"Claude Opus 5是Anthropic于2026年7月24日发布的主力旗舰MoE多模态模型，原生支持1M token完整上下文窗口，搭载可调节Effort自适应推理机制，主打企业级长周期Agent、大型代码库工程开发与海量文档深度研判，在Agentic编程、计算机操控评测稳居全球第一梯队，综合实战能力接近Claude Fable 5，同时具备显著的成本优势，长任务稳定推理表现对标全球头部商用闭源模型。","tags":"chat","vendor_id":37,"quota_type":0,"model_ratio":2.5,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.1,"create_cache_ratio":1.25,"enable_groups":["tj_claude"],"supported_endpoint_types":["openai","anthropic"]},{"model_name":"grok-4.7","description":"Grok‑4.7 是 xAI（SpaceXAI）2026‑09‑21 推出的旗舰大模型，上下文窗口 50 万 tokens，知识截止 2026‑05，经过长周期强化学习训练，重点优化长时 Agent、软件工程、电气工程与法律办公任务，内置自我校验能力，原生适配 Grok Bot 智能体框架。超长 prompt 区间价格上浮，还提供高速变体。擅长持续数小时的复杂工程任务，工程类评测表现突出，适合专业编程、文档分析与多步骤科研工作。","tags":"chat","vendor_id":49,"quota_type":0,"model_ratio":10.5,"model_price":0,"owner_by":"","completion_ratio":0.5714285714285714,"enable_groups":["tj_grok"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003c= 200000 ? tier(\"短文本\", p * 2 + c * 6 + cr * 0.5) : tier(\"长文本\", p * 4 + c * 12 + cr * 1)"},{"model_name":"doubao-seedance-2-0-mini-260615","description":"bao-seedance-2-0-mini-260615 是字节跳动 Seed 团队于 2026 年 6 月 15 日发布的首款面向端侧与轻量场景的迷你级多模态视频生成模型，官方称其为“兼顾低算力部署与高质量输出的轻量化通用视频模型”。其最大突破在于具备端侧低延迟实时生成能力，可完成移动端创意短视频制作、嵌入式设备动态内容生成、轻量化电商素材产出及低算力环境长周期创作等轻量化工作。在 MobileVideoBench 端侧基准测试中，Mini 版得分 78.4%（远超 1.5 Mini 的 62.9%），LightGen 轻量生成质量测试取得 85.1% 的成绩，接近中端云端模型产出水准。定价相较 Seedance 2.0 标准版下浮 65%，同时将轻量上下文窗口扩展至 10 万 token，可一次性处理完整短脚本、单组参考素材或日常量级的连续创作需求。","tags":"video","vendor_id":45,"quota_type":0,"model_ratio":37.5,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["video"],"supported_endpoint_types":["openai-video"],"billing_mode":"tiered_expr","billing_expr":"(tier(\"文生视频\", p * 0 + c * 3.29)) * (param(\"reference_video_urls\") != nil ? 0.6 : 1)"},{"model_name":"deepseek-v4-pro","description":"DeepSeek-V4-Pro是深度求索于2026年4月24日发布的高性能MoE旗舰模型。总参数量1.6万亿（激活49B），支持1M token上下文。主打顶尖Agent能力，在Agentic Coding评测中登顶开源榜首，内部体验优于Sonnet 4.5，推理性能比肩世界顶级闭源模型。","tags":"chat","vendor_id":29,"quota_type":0,"model_ratio":0.66,"model_price":0,"owner_by":"","completion_ratio":3,"cache_ratio":0.033333333333,"enable_groups":["default","official"],"supported_endpoint_types":["openai"]},{"model_name":"hy3","description":"Hy3 是腾讯混元团队于2026年7月6日发布的高性能MoE开源旗舰模型。总参数量2950亿（激活21B），支持256K token上下文。主打高效能Agent与全场景生产力，在Agentic Coding、复杂推理评测中领跑同参数级开源模型，内部实测体验优于主流同档位闭源模型，推理能效比肩参数规模数倍的世界顶级大模型。","icon":"Hunyuan.Avatar.type={'logo'}","tags":"chat","vendor_id":51,"quota_type":0,"model_ratio":0.074,"model_price":0,"owner_by":"","completion_ratio":3.9864864864864864,"cache_ratio":0.25,"enable_groups":["guochan","guochan_便宜","official"],"supported_endpoint_types":["openai"]},{"model_name":"gpt-6-sol","description":"GPT‑6‑Sol 是 OpenAI 在 2026‑09‑22 发布的均衡型旗舰模型，属于 GPT‑6 系列，定位介于 GPT‑6‑Astra 与 GPT‑6‑Luna 之间，上下文窗口 105 万 tokens，最大输出 128k tokens，知识截止 2026‑04‑20，支持图文输入、多档推理强度调节，原生支持函数调用、结构化输出与提示缓存。继承 Astra 的 Agent 与代码能力，擅长软件工程、多步骤智能体任务、长文档分析，，成本远低于 Astra。面向 ChatGPT Plus/Pro/ 企业用户开放，适合复杂专业工作，极限深度推理能力略弱于 GPT‑6‑Astra。","tags":"chat","vendor_id":38,"quota_type":0,"model_ratio":11,"model_price":0,"owner_by":"","completion_ratio":12,"enable_groups":["stable","tj_openai","tj_openai_az","official"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003c= 272000 ? tier(\"短文本\", p * 2 + c * 10 + cr * 0.2 + cc * 2.5) : tier(\"长文本\", p * 4 + c * 15 + cr * 0.4 + cc * 5)"},{"model_name":"gemini-2.5-flash","description":"Gemini 2.5 Flash是谷歌于2025年4月10日推出的混合推理模型，定位平衡质量、成本与延迟。它支持100万token上下文和多模态输入，具备“可控思考预算”特性。定价为输入$0.30/百万tokens，输出$2.50/百万tokens，适用于高容量实时任务。\n","tags":"chat","vendor_id":36,"quota_type":0,"model_ratio":0.15,"model_price":0,"owner_by":"","completion_ratio":8.333333333333,"cache_ratio":0.1,"enable_groups":["stable","tj_gemini"],"supported_endpoint_types":["openai","gemini"]},{"model_name":"doubao-seedance-2-0-fast-260128","description":"doubao-seedance-2-0-fast-260128 是字节跳动 Seed 团队于 2026 年 2 月 12 日发布的首款面向规模化商用的极速多模态视频生成模型，官方称其为“兼顾工业级生成质量与毫秒级推理速度的量产主力视频模型”。其最大突破在于采用轻量化联合扩散加速架构，可完成短视频批量产出、内容流水线自动化、实时创意预览及电商素材批量生成等高吞吐生产任务。在复杂运动场景可用率测试中，Fast 版得分 91.7%（仅小幅低于标准版的 95.2%），批量生产吞吐基准测试达到标准版 2.1 倍产能，接近专业渲染集群生产效率。定价相较旗舰标准版下调约 17.4%，同时将多模态上下文承载扩展至万字级脚本与多素材并行输入，可一次性承接完整长叙事、多素材混剪等复杂创作需求。","tags":"video","vendor_id":45,"quota_type":0,"model_ratio":37.5,"model_price":0,"owner_by":"","completion_ratio":1,"enable_groups":["video"],"supported_endpoint_types":["openai-video"],"billing_mode":"tiered_expr","billing_expr":"(tier(\"文生视频\", p * 0 + c * 5.29)) * (param(\"reference_video_urls\") != nil ? 0.6 : 1)"},{"model_name":"glm-5.3-flash","description":"GLM‑5.3‑Flash是智谱发布的GLM‑5系列首款原生多模态MoE模型，总参数量320B、激活参数18B，采用稀疏+线性注意力混合架构，支持100万token上下文窗口，凭借30T多模态预训练语料，综合与代码智能体能力对标Claude Opus4.8，整体性能优于GLM‑5.2，推理成本大幅下降，MIT协议开源、API同步上线，兼顾高吞吐量多模态处理、Agent与代码任务，主打极致性价比。","tags":"chat","vendor_id":31,"quota_type":0,"model_ratio":0.075,"model_price":0,"owner_by":"","completion_ratio":3.3333333333333335,"cache_ratio":0.2,"enable_groups":["default","guochan","guochan_便宜","official"],"supported_endpoint_types":["openai"]},{"model_name":"kimi-k3","description":"Kimi-K3是月之暗面（Moonshot AI）于2026年7月16日推出的旗舰MoE开源大模型，总参数量2.8万亿，原生搭载1M token超长上下文窗口并支持视觉多模态输入，自研KDA混合线性注意力与注意力残差架构大幅提升长序列推理效率，主打长周期工程Agent与海量知识库深度分析，在前端编程专项评测登顶全球榜首，综合实力仅次于GPT-5.6-Sol、Claude Fable 5两大闭源旗舰，长代码库解析、自动化办公智能体性能比肩世界顶级商用模型，完整权重将于7月27日对外开源。","tags":"chat","vendor_id":30,"quota_type":0,"model_ratio":1.5,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.1,"enable_groups":["guochan","guochan_自部署","official"],"supported_endpoint_types":["openai"]},{"model_name":"qwen3.8-flash","description":"Qwen3.8‑Flash是阿里云通义千问团队于2026年8月26日发布的高性能多模态MoE模型，总参数量125B（激活6B），可扩展至1M token上下文，主打高吞吐Agent与轻量化工程生产力，在办公、Agentic Coding评测位居开源第一梯队，内部体验优于Gemini‑3.7‑Flash，推理能效比肩世界顶级商用闭源模型。","tags":"chat","vendor_id":33,"quota_type":0,"model_ratio":0.075,"model_price":0,"owner_by":"","completion_ratio":3.1333333333333333,"cache_ratio":0.106666666667,"create_cache_ratio":1.333333333333,"enable_groups":["guochan","official"],"supported_endpoint_types":["openai"]},{"model_name":"mimo-v2.6-flash","description":"MiMo-V2.6-Flash 是小米推出的全模态 MoE 模型，主打高吞吐、低延迟推理，原生支持文本、图像、音频、视频多模态输入。模型采用全局注意力搭配滑动窗口注意力机制，基础上下文窗口 32K，可扩展至 256K，内置深度思考模式、工具调用与 Agent 能力，代码工程表现亮眼。经过强化学习优化，相比前代版本，多模态理解与复杂指令执行能力显著提升。该版本定位均衡实用，API 调用成本友好，缓存命中还可享受折扣，适合 API 批量调用、文档解析、自动化 Agent 等高频业务场景。短板在于顶尖科研难题的深度推理能力不及同系列 Pro 旗舰版本","tags":"chat","vendor_id":46,"quota_type":0,"model_ratio":0.07,"model_price":0,"owner_by":"","completion_ratio":2,"cache_ratio":0.02,"enable_groups":["guochan_便宜","official","guochan"],"supported_endpoint_types":["openai"]},{"model_name":"gemini-3.1-flash-image","description":"Nano Banana 2 以主流价格提供高质量的图片生成和对话式编辑功能，延迟时间短。\n\n","tags":"image","vendor_id":36,"quota_type":1,"model_ratio":0,"model_price":0.24,"owner_by":"","completion_ratio":0,"enable_groups":["official","tj_gemini"],"supported_endpoint_types":["openai","gemini"]},{"model_name":"gpt-5.6-sol","description":"GPT-5.6-Sol是OpenAI于2026年6月26日发布的全系旗舰MoE多模态大模型，总参数量3万亿（激活1500B），支持150M token超长无损上下文，原生搭载Ultra多智能体并行调度与世界模型推理能力，主打高阶复杂Agent、全栈工程开发、前沿科研与网络安全分析，在多轮智能体编程综合评测刷新全球SOTA，综合推理与长链路任务表现优于Claude Fable 5，高负载推理性能位居全球闭源模型第一梯队。","tags":"chat","vendor_id":38,"quota_type":0,"model_ratio":37.5,"model_price":0,"owner_by":"","completion_ratio":8,"enable_groups":["official","stable","tj_openai","tj_openai_az"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003e 272000 ? tier(\"长文本\", p * 10 + c * 45 + cr * 1 + cc * 12.5) : tier(\"短文本\", p * 5 + c * 30 + cr * 0.5 + cc * 6.25)"},{"model_name":"gpt-6-astra","description":"gpt‑6‑astra是OpenAI 2026‑09‑03推出的旗舰推理模型，API标识符`gpt‑6‑astra`。上下文窗口105万tokens，最大输出128k tokens，知识截止2026‑04‑30，支持图文输入、多档推理强度调节。主打Agent计算机操作、深度推理、大型代码工程、百万长文档解析，多项推理基准刷新SOTA，当前企业灰度开放，后续逐步面向各类用户放开，内置智能体安全管控机制。","tags":"chat","vendor_id":38,"quota_type":0,"model_ratio":1,"model_price":0,"owner_by":"","completion_ratio":2,"enable_groups":["official","stable","tj_openai","tj_openai_az"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003e 272000 ? tier(\"长文本\", p * 20 + c * 75 + cr * 2 + cc * 25) : tier(\"短文本\", p * 10 + c * 50 + cr * 1 + cc * 12.5)"},{"model_name":"claude-opus-5-5","description":"Anthropic 于 2026‑09‑22 发布的旗舰模型，拥有 1M token 上下文窗口，最大输出 128K tokens，多数任务能力对标 Claude Fable 5.1，输出速度较 Opus 5 提升 30% 以上。API 定价输入 $4、输出 $20 每百万 token，缓存读取低至 $0.2/MTok，典型任务总成本相比上代降低 40%，还提供 2.5 倍提速 Fast 模式。它在 Agent 编码评测表现突出，擅长长周期智能体任务、大型代码库重构与深度知识工作，经过严格安全对齐测试，","tags":"chat","vendor_id":37,"quota_type":0,"model_ratio":2,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.05,"create_cache_ratio":1.25,"enable_groups":["stable","tj_claude_max"],"supported_endpoint_types":["anthropic","openai"]},{"model_name":"gpt-6.1-sol","description":"GPT‑6.1‑Sol 是 OpenAI 2026‑09‑25 发布的 GPT‑6.1 系列均衡主力模型，上下文窗口 105 万 tokens，最大输出 128k tokens，知识截止更新至 2026‑05‑10，支持图文输入、多档推理强度调节、函数调用与提示缓存。相比前代 GPT‑6‑Sol，强化长链路 Agent、大型代码重构与多模态推理。API 定价输入 $1.8、输出 $9 每百万 token，性价比突出，面向专业文档分析、软件开发、复杂业务自动化场景，极致推理上限略弱于同代旗舰 Astra","tags":"chat","vendor_id":38,"quota_type":0,"model_ratio":1,"model_price":0,"owner_by":"","completion_ratio":5,"enable_groups":["tj_openai"],"supported_endpoint_types":["openai"],"billing_mode":"tiered_expr","billing_expr":"p \u003c= 272000 ? tier(\"短文本\", p * 2 + c * 10 + cr * 0.1 + cc * 2.5) : tier(\"长文本\", p * 4 + c * 15 + cr * 0.2 + cc * 5)"},{"model_name":"claude-haiku-4-5-20251001","description":"Claude Haiku 4.5是Anthropic于2025年10月15日发布的最快轻量模型。定位为“速度与成本之王”，性能对标Claude Sonnet 4但成本仅为其三分之一，速度快两倍以上。在SWE-bench Verified编码测试中得分73.3%，支持200K上下文，定价为输入$1/输出$5每百万tokens。","tags":"chat","vendor_id":37,"quota_type":0,"model_ratio":0.5,"model_price":0,"owner_by":"","completion_ratio":5,"cache_ratio":0.1,"create_cache_ratio":1.25,"enable_groups":["stable","tj_claude","tj_claude_max"],"supported_endpoint_types":["anthropic","openai"]},{"model_name":"deepseek-v4.1-flash","description":"DeepSeek‑V4.1‑Flash是深度求索2026‑09‑08推出的高效多模态MoE模型，总参552B、激活8B，支持1M token上下文，新增原生视觉理解，优化KV缓存压缩，大幅降低显存开销。主打高吞吐Agent、轻量工程编码与高频业务生产，Agentic‑Coding稳居开源第一梯队，推理成本远低于旗舰版本，API优先开放，同步提供MIT开源权重供企业部署。","tags":"chat","vendor_id":29,"quota_type":0,"model_ratio":0.15,"model_price":0,"owner_by":"","completion_ratio":4,"cache_ratio":0.02,"enable_groups":["guochan","official"],"supported_endpoint_types":["0"]},{"model_name":"kimi-k2.7-code","description":"Kimi K2.7-Code是月之暗面于2026年6月12日开源的MoE架构下一代Agentic编程模型，总参数1.1万亿（激活320亿），支持256K上下文及原生多模态（图文输入）。核心突破是“长程任务高效推理”，平均思考token消耗较上代减少30%，在Kimi Code Bench v2编码测试中得分62.0%（+21.8%），MCP Mark Verified智能体测试得分81.1%，兼顾性能与成本效益。","tags":"chat","vendor_id":30,"quota_type":0,"model_ratio":0.475,"model_price":0,"owner_by":"","completion_ratio":4.210526315789,"cache_ratio":0.2,"enable_groups":["official","guochan","guochan_便宜"],"supported_endpoint_types":["openai"]},{"model_name":"mimo-v2.6-pro","description":"MiMo-V2.6-Pro 是小米 MiMo 系列旗舰全模态 MoE 大模型，面向高难度复杂任务打造，原生支持文本、图像、音频、视频多模态理解。模型搭载增强版混合注意力架构，基础上下文 32K，最大可扩展至 256K，强化深度推理、复杂代码工程、多步骤 Agent 规划能力，在 SWE-Bench 等硬核评测上表现优于 Flash 版本。经过高强度 RL 优化，擅长复杂数学、长文档逻辑分析、精密多模态推理。推理延迟与调用成本高于 Flash 版，更适合科研攻坚、复杂软件开发、高要求多模态任务，不适合大规模高频批量请求。","tags":"chat","vendor_id":46,"quota_type":0,"model_ratio":0.2175,"model_price":0,"owner_by":"","completion_ratio":2,"cache_ratio":0.008275862069,"enable_groups":["guochan","official"],"supported_endpoint_types":["openai"]},{"model_name":"gpt-5.5","description":"GPT‑5.5是OpenAI于2026年4月23日推出的旗舰级智能体编程模型，内部代号“Spud”，被官方定义为“迄今最智能、最直观的任务执行模型”。模型擅长编写与调试代码、在线研究、数据分析及跨工具协同操作，支持100万token上下文窗口，并具备屏幕识别、UI点击、终端操控等多模态计算机使用能力。在复杂命令行工作流基准Terminal‑Bench 2.0测试中取得82.7%的准确率，在Expert‑SWE长周期软件工程任务中达到73.1%，并可稳定自主运行超9小时的长链条任务。","tags":"chat","vendor_id":38,"quota_type":0,"model_ratio":2.5,"model_price":0,"owner_by":"","completion_ratio":6,"cache_ratio":0.2,"enable_groups":["official","tj_openai"],"supported_endpoint_types":["openai"]},{"model_name":"gpt-image-2.5-flare","description":"gpt‑image‑2.5‑flare是OpenAI2026‑09‑09发布的高速图像生成模型，为API默认选型，延迟相比上代降低50%，最高支持4K输出，优化透明背景、参考图局部编辑与多轮编辑一致性，具备草图输入能力，主打批量出图、社交素材、快速原型等高频场景，内置C2PA溯源水印，兼顾生成质量与吞吐效率，面向ChatGPT与开发者API开放调用。","tags":"image","vendor_id":38,"quota_type":1,"model_ratio":0,"model_price":0.15,"owner_by":"","completion_ratio":0,"enable_groups":["tj_openai_az","tj_openai"],"supported_endpoint_types":["openai"]}],"group_ratio":{"auto":1,"default":1,"guochan":0.42,"guochan_便宜":0.3,"guochan_自部署":0.4,"official":1,"stable":0.52,"stable_pic":0.75,"tj_claude":0.12,"tj_claude_max":0.3,"tj_gemini":0.2,"tj_grok":0.06,"tj_openai":0.08,"tj_openai_az":0.36,"video":1},"pricing_version":"a42d372ccf0b5dd13ecf71203521f9d2","success":true,"supported_endpoint":{"0":{"path":"openai","method":"POST"},"anthropic":{"path":"/v1/messages","method":"POST"},"gemini":{"path":"/v1beta/models/{model}:generateContent","method":"POST"},"openai":{"path":"/v1/chat/completions","method":"POST"}},"usable_group":{"auto":"有钱的选这个，启动切换分组","default":"内部测试专用，不要选","guochan":"部分阿里云，腾讯云，opencode混合","guochan_便宜":"部分国产加cursor的dp","guochan_自部署":"自部署的精度有点问题","official":"官方直连","stable":"相对稳定官方key","stable_pic":"相对稳定pic","tj_claude":"kiro和cursor混合有fable5","tj_claude_max":"claude特价max","tj_gemini":"gemini号池","tj_grok":"grok特价","tj_openai":"openai pro号池","tj_openai_az":"Azure微软云","video":"视频测试（勿删）"},"vendors":[{"id":38,"name":"OpenAI","icon":"OpenAI"},{"id":46,"name":"Xiaomi","icon":"XiaomiMiMo.Avatar.shape={'square'}"},{"id":31,"name":"智谱","icon":"Zhipu.Color"},{"id":32,"name":"Minimax","icon":"Minimax.Avatar.type={'logo'}"},{"id":36,"name":"Google","icon":"Gemini.Color"},{"id":45,"name":"字节跳动","icon":"Doubao.Color"},{"id":49,"name":"xAI","icon":"XAI"},{"id":50,"name":"阿里巴巴","icon":"Qwen.Color"},{"id":51,"name":"Hunyuan","icon":"Hunyuan.Avatar.type={'logo'}"},{"id":29,"name":"DeepSeek","icon":"DeepSeek.Color"},{"id":30,"name":"Moonshot","icon":"Moonshot"},{"id":33,"name":"Qwen","icon":"Qwen.Color"},{"id":37,"name":"Anthropic","icon":"Claude.Color"}]}