Intelligently select the optimal model tier for AI tasks of varying complexity, saving 70-90% on token costs without sacrificing quality. 为不同复杂度的 AI 任务智能选择最合适的模型层级,在保证质量的前提下节省 70-90% 的 Token 成本。
- 4-Dimension Complexity Assessment: Reasoning depth, output length, precision requirement, context dependency (1-5 scale), auto-mapped to L0-L3 model tiers
- Safety-Forced Upgrades: Auto-upgrade to L2+ for finance/legal/medical/security/production scenarios
- Progressive User Profiling: Learn preferences from conversation, no questionnaires
- 3 Routing Modes: Direct (daily), Cascading (batch), Hybrid (Agent workflows)
- Multi-Platform Support: Trae IDE, Claude Code, OpenClaw, Hermes Agent, and more
| Tier | Role | Representative Models | Typical Scenarios |
|---|---|---|---|
| L0 | Router | DeepSeek-V4 Flash, GPT-4.1 nano, Gemini 2.5 Flash-Lite | Classification, extraction, formatting, routing |
| L1 | Executor | MiniMax M3, Claude Haiku 4.5, Gemini 3.5 Flash | Summarization, translation, simple QA, structured output |
| L2 | Reasoner | Claude Sonnet 4.6, GPT-5.5, Qwen 3.7 Max | Code generation, analysis reports, multi-step reasoning |
| L3 | Creator | Claude Opus 4.8, GPT-5.5 Pro, o3 | Architecture design, creative writing, complex planning |
This skill activates when users discuss:
- "API costs too high", "save money on AI", "token cost optimization", "reduce API bill"
- "which model should I use", "Sonnet vs Opus", "help me pick a model"
- "configure Hermes/OpenClaw multi-model routing", "model tiers", "intelligent scheduling"
- Solo founder / indie developer AI tool costs
- Building Agent workflows with different models for different tasks
Simple Task → Lightweight Model
User: Classify these 100 emails, find which ones are complaints
Recommend: DeepSeek-V4 Flash (L0) — email classification is pattern matching, lightweight model sufficient
Estimate: $0.02 (vs flagship $0.50, save 96%)
Complex Task → Flagship Model
User: Design a microservice architecture supporting million-level concurrency
Recommend: Claude Opus 4.8 (L3) — architecture design requires deep reasoning
Agent Workflow → Hybrid Routing
Email classification → L0 (DeepSeek V4 Flash)
Daily report generation → L1 (MiniMax M3)
Code review → L2 (Claude Sonnet 4.6)
Exception escalation → L3 (Claude Opus 4.8)
Estimated monthly cost: $3-8
token-router/
├── SKILL.md # Main skill file (decision flow, output formats, platform guides)
├── README.md # This file
├── references/
│ ├── model-tiers.md # 20+ model tiers with pricing (with freshness disclaimer)
│ ├── routing-strategies.md # Deep routing guide (cascading, caching strategies)
│ └── config-templates.md # Config templates (Trae/OpenClaw/Hermes-specific)
└── evals/
└── evals.json # Evaluation test cases (20 scenarios, 95% pass rate)
| Dimension | Score |
|---|---|
| Feature Completeness | 9/10 |
| Instruction Clarity | 9.5/10 |
| Test Pass Rate | 19/20 (95%) |
| Structural Compliance | 9/10 |
| Platform Coverage | 9/10 |
| Overall | 9.0/10 |
- 四维复杂度评估:推理深度、输出长度、精度要求、上下文依赖(1-5分制),自动映射到 L0-L3 四个模型层级
- 安全强制升级:金钱/法律/医疗/安全/生产环境场景自动升级到 L2+ 模型
- 渐进式用户画像:从对话中学习偏好,不问问卷
- 三种路由方式:单次路由(日常)、级联路由(批量)、混合路由(Agent工作流)
- 多平台支持:Trae IDE、Claude Code、OpenClaw、Hermes Agent 等
| 层级 | 定位 | 代表模型 | 典型场景 |
|---|---|---|---|
| L0 | 路由级 | DeepSeek-V4 Flash、GPT-4.1 nano、Gemini 2.5 Flash-Lite | 分类、提取、格式化、路由 |
| L1 | 执行级 | MiniMax M3、Claude Haiku 4.5、Gemini 3.5 Flash | 摘要、翻译、简单QA、结构化输出 |
| L2 | 推理级 | Claude Sonnet 4.6、GPT-5.5、Qwen 3.7 Max | 代码生成、分析报告、多步推理 |
| L3 | 创造级 | Claude Opus 4.8、GPT-5.5 Pro、o3 | 架构设计、创意写作、复杂规划 |
当用户涉及以下场景时自动触发:
- "API费用太贵"、"帮我省钱"、"Token太贵"、"怎么降本"
- "用哪个模型"、"Sonnet和Opus选哪个"、"帮我选模型"
- "配置Hermes/OpenClaw多模型路由"、"模型分级"、"智能调度"
- 一人公司/独立开发者的AI工具成本问题
- 搭建需要不同模型处理不同任务的Agent工作流
简单任务 → 轻量模型
用户:帮我给100封邮件分个类,看哪些是投诉
推荐:DeepSeek-V4 Flash(L0)— 邮件分类是模式匹配,轻量模型足够
预估:$0.02(vs 旗舰 $0.50,省96%)
复杂任务 → 旗舰模型
用户:帮我设计一个微服务架构,要支持百万级并发
推荐:Claude Opus 4.8(L3)— 架构设计需要深度推理
Agent工作流 → 混合路由
邮件分类 → L0(DeepSeek V4 Flash)
日报生成 → L1(MiniMax M3)
代码review → L2(Claude Sonnet 4.6)
异常升级 → L3(Claude Opus 4.8)
预估月成本:$3-8
token-router/
├── SKILL.md # 主技能文件(决策流程、输出格式、平台指引)
├── README.md # 本文件
├── references/
│ ├── model-tiers.md # 20+ 模型详细分级与定价(含时效性声明)
│ ├── routing-strategies.md # 路由策略深度指南(级联实现、缓存策略)
│ └── config-templates.md # 配置模板(Trae/OpenClaw/Hermes 专属配置)
└── evals/
└── evals.json # 评估测试用例(20个场景,95%通过率)
| 维度 | 评分 |
|---|---|
| 功能完整性 | 9/10 |
| 指令清晰度 | 9.5/10 |
| 测试通过率 | 19/20(95%) |
| 结构合规性 | 9/10 |
| 平台覆盖 | 9/10 |
| 综合 | 9.0/10 |
