@gabrielsmartin/orbit-sdk 0.1.0

package/README.md

@@ -0,0 +1,202 @@
+# orbit-ai
+
+> Stop blasting every query at GPT-4o. Route intelligently. Save 85%.
+
+`orbit-ai` is a drop-in routing layer that reads the fingerprint of every AI query and sends it to the optimal model — automatically, in under 1ms.
+
+```bash
+npm install orbit-ai
+```
+
+---
+
+## The problem
+
+You're probably doing this:
+
+```javascript
+const res = await openai.chat.completions.create({
+  model: "gpt-4o",  // $30/1M tokens — every single query
+  messages
+})
+```
+
+GPT-4o costs **30x more per token** than Gemini Flash. For "what's 2+2?" — you're paying Ferrari prices to drive to the mailbox.
+
+ORBIT fixes this. One line.
+
+---
+
+## How it works
+
+Every query gets fingerprinted across **8 axes** in under 1ms:
+
+| Axis | What it measures |
+|---|---|
+| **Complexity** | Depth of reasoning required |
+| **Creativity** | Open-ended vs deterministic |
+| **Emotional Weight** | Sensitivity — crisis queries always go to Claude |
+| **Recency** | Need for live/current data → Grok |
+| **Context Load** | Window size needed → Claude 200k |
+| **Speed** | Latency sensitivity |
+| **Domain** | Code · Creative · Medical · Legal · General |
+| **Cost Tolerance** | Budget tier (overridable) |
+
+Then it routes to the right model. Invisibly.
+
+---
+
+## Usage
+
+### Zero-config routing decision
+
+```javascript
+import orbit from 'orbit-ai'
+
+// Get the routing decision
+const decision = orbit.route("write a haiku about recursion")
+
+console.log(decision.model.name)    // "Claude Sonnet"
+console.log(decision.reason)        // "High creativity score (8/10)..."
+console.log(decision.savings)       // { savings: 0.007245, reductionPct: 97 }
+```
+
+### With your own API keys
+
+```javascript
+import { OrbitClient } from 'orbit-ai'
+
+const orbit = new OrbitClient({
+  cost_tolerance: 'low',  // 'low' | 'medium' | 'high'
+  log: true,              // logs routing decisions to console
+})
+
+// Route the query
+const { model, reason, savings } = orbit.route("explain blockchain simply")
+// [ORBIT] → Gemini 2.5 Flash | cost_gemini | saved $0.01455 (97% reduction)
+
+// Now call the model yourself with your keys
+// model.id = 'gemini-2.5-flash', model.provider = 'google'
+```
+
+### Full pipeline example (with Anthropic SDK)
+
+```javascript
+import { OrbitClient } from 'orbit-ai'
+import Anthropic from '@anthropic-ai/sdk'
+import OpenAI from 'openai'
+import { GoogleGenerativeAI } from '@google/generative-ai'
+
+const orbit = new OrbitClient({ log: true })
+
+async function smartQuery(text) {
+  const { model, reason } = orbit.route(text)
+
+  if (model.provider === 'anthropic') {
+    const client = new Anthropic()
+    return client.messages.create({
+      model: model.id,
+      max_tokens: 1024,
+      messages: [{ role: 'user', content: text }]
+    })
+  }
+
+  if (model.provider === 'openai') {
+    const client = new OpenAI()
+    return client.chat.completions.create({
+      model: model.id,
+      messages: [{ role: 'user', content: text }]
+    })
+  }
+
+  if (model.provider === 'google') {
+    const client = new GoogleGenerativeAI(process.env.GOOGLE_API_KEY)
+    const genModel = client.getGenerativeModel({ model: model.id })
+    return genModel.generateContent(text)
+  }
+}
+
+// Routes each query to the optimal model
+await smartQuery("write a poem about the ocean")           // → Claude Sonnet
+await smartQuery("what's the latest news on AI funding?")  // → Grok
+await smartQuery("what is 2+2")                            // → Gemini Flash
+await smartQuery("I've been feeling really overwhelmed")   // → Claude Sonnet (ethics-first)
+```
+
+### Session stats
+
+```javascript
+const stats = orbit.stats()
+console.log(stats.total_savings_formatted)  // "$0.2341"
+console.log(stats.model_usage)              // { "Claude Sonnet": 4, "Gemini 2.5 Flash": 12, ... }
+```
+
+### Fingerprint only (no routing)
+
+```javascript
+import { fingerprint } from 'orbit-ai'
+
+const scores = fingerprint("architect a distributed caching system for 10M users")
+// {
+//   complexity: 9,
+//   creativity: 0,
+//   domain: 'code',
+//   emotional_weight: 0,
+//   recency: 0,
+//   context_load: 3,
+//   ...
+// }
+```
+
+---
+
+## Model matrix
+
+| Model | Provider | Cost/1M | Best for |
+|---|---|---|---|
+| Claude Sonnet | Anthropic | $15 | Complex reasoning · Ethics · Long context |
+| Claude Haiku | Anthropic | $1 | Speed · Summaries · Medium tasks |
+| Gemini 2.5 Flash | Google | $0.50 | High volume · Simple queries · Cost |
+| GPT-4o | OpenAI | $30 | Structured output · Broad knowledge |
+| GPT-4o Mini | OpenAI | $0.30 | Classification · Filler tasks |
+| Grok | xAI | $10 | Trending · Real-time web |
+
+---
+
+## The math
+
+Validated by [RouteLLM (UC Berkeley · ICLR 2025)](https://arxiv.org/abs/2406.18665): intelligent routing achieves **85% cost reduction** while maintaining 95% of quality.
+
+For a team running 100k queries/month at GPT-4o:
+- Without ORBIT: **$1,500/month**
+- With ORBIT: **~$225/month**
+- Savings: **$1,275/month · $15,300/year**
+
+---
+
+## Live demo
+
+[orbit-model-flow.base44.app](https://orbit-model-flow.base44.app) — route a real query, see the fingerprint radar, watch the savings counter.
+
+---
+
+## Roadmap
+
+- [x] 8-axis fingerprinting engine
+- [x] 6-model routing matrix
+- [x] TypeScript types
+- [ ] Streaming support
+- [ ] Custom model matrix (bring your own models)
+- [ ] Automatic provider failover
+- [ ] Usage analytics dashboard integration
+- [ ] Browser extension
+
+---
+
+## License
+
+MIT · Built by [Gabriel Martin](https://github.com/gabrielsmartin)
+
+*"Every model has a gravitational pull. ORBIT decides which one you need."*
+
+777 · 555 · 333

package/package.json

@@ -0,0 +1,52 @@
+{
+  "name": "@gabrielsmartin/orbit-sdk",
+  "version": "0.1.0",
+  "description": "Intelligent AI model routing. Drop-in replacement for OpenAI/Anthropic. Routes every query to the optimal model automatically.",
+  "type": "module",
+  "main": "src/index.js",
+  "types": "src/index.d.ts",
+  "exports": {
+    ".": {
+      "import": "./src/index.js",
+      "types": "./src/index.d.ts"
+    }
+  },
+  "files": [
+    "src",
+    "README.md"
+  ],
+  "scripts": {
+    "test": "node test.js"
+  },
+  "keywords": [
+    "ai",
+    "llm",
+    "routing",
+    "openai",
+    "anthropic",
+    "gemini",
+    "orbit",
+    "cost-optimization",
+    "model-routing",
+    "selective-model-matching",
+    "gpt4",
+    "claude",
+    "gemini-flash",
+    "grok",
+    "ai-infrastructure"
+  ],
+  "author": "Gabriel Martin <gabriel@gtll.app>",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/gabrielsmartin/orbit.git",
+    "directory": "sdk"
+  },
+  "homepage": "https://orbit-model-flow.base44.app",
+  "bugs": {
+    "url": "https://github.com/gabrielsmartin/orbit/issues"
+  },
+  "engines": {
+    "node": ">=16"
+  }
+}

package/src/fingerprint.js

@@ -0,0 +1,102 @@
+/**
+ * ORBIT · 8-Axis Query Fingerprinting Engine
+ * Scores every query across 8 dimensions in <1ms
+ * 
+ * interval(n) = base / 2^n — Recursive Beat Engine
+ * 777 · 555 · 333
+ */
+
+const COMPLEXITY_SIGNALS = [
+  'architect','distributed','implement','design','optimize','analyze',
+  'complex','system','algorithm','concurrent','scale','infrastructure',
+  'microservice','kubernetes','terraform','recursive','refactor'
+];
+
+const CREATIVITY_SIGNALS = [
+  'write','create','generate','poem','story','haiku','imagine',
+  'creative','brainstorm','ideas','invent','compose','design','narrative'
+];
+
+const RECENCY_SIGNALS = [
+  'latest','today','current','news','now','recent','trending',
+  '2024','2025','2026','live','breaking','update','this week'
+];
+
+const EMOTIONAL_SIGNALS = [
+  'feel','feeling','anxious','anxiety','depressed','depression','overwhelmed',
+  'scared','lonely','hurt','grief','trauma','crisis','suicid','help me',
+  'relationship','breakup','divorce','mental health','therapy','struggling'
+];
+
+const CODE_SIGNALS = [
+  'code','function','class','api','debug','error','syntax','compile',
+  'deploy','database','query','algorithm','javascript','python','typescript',
+  'react','node','sql','git','docker','bug','fix','implement'
+];
+
+/**
+ * Fingerprint a query across 8 axes
+ * @param {string} text - The query text
+ * @returns {Object} scores - 0-10 score for each axis
+ */
+export function fingerprint(text) {
+  const t = text.toLowerCase();
+  const words = t.split(/\s+/);
+  const wordCount = words.length;
+
+  // Complexity (0-10): reasoning depth required
+  const complexitySignals = COMPLEXITY_SIGNALS.filter(s => t.includes(s)).length;
+  const questionDepth = (t.match(/\b(why|how|explain|compare|tradeoff|pros|cons|difference|analyze)\b/g) || []).length;
+  const complexity = Math.min(10, Math.round(
+    complexitySignals * 2 +
+    questionDepth * 1.5 +
+    (wordCount > 30 ? 3 : wordCount > 15 ? 1.5 : 0)
+  ));
+
+  // Creativity (0-10): open-ended vs deterministic
+  const creativitySignals = CREATIVITY_SIGNALS.filter(s => t.includes(s)).length;
+  const creativity = Math.min(10, Math.round(creativitySignals * 2.5));
+
+  // Speed (0-10): latency sensitivity
+  const speedSignals = (t.match(/\b(quick|fast|brief|tldr|short|simple|summarize|quickly)\b/g) || []).length;
+  const speed = Math.min(10, Math.round(speedSignals * 3 + (wordCount < 8 ? 3 : 0)));
+
+  // Cost tolerance (0-10): inferred from query type (overridable)
+  // Default: medium. Will be set by user config.
+  const costTolerance = 5;
+
+  // Emotional weight (0-10): sensitivity of content
+  const emotionalSignals = EMOTIONAL_SIGNALS.filter(s => t.includes(s)).length;
+  const emotional_weight = Math.min(10, Math.round(emotionalSignals * 3));
+
+  // Recency (0-10): need for live/current data
+  const recencySignals = RECENCY_SIGNALS.filter(s => t.includes(s)).length;
+  const recency = Math.min(10, Math.round(recencySignals * 3.5));
+
+  // Context load (0-10): context window size needed
+  const context_load = Math.min(10, Math.round(
+    (wordCount > 100 ? 8 : wordCount > 50 ? 5 : wordCount > 20 ? 3 : 1)
+  ));
+
+  // Domain classification
+  const codeSignals = CODE_SIGNALS.filter(s => t.includes(s)).length;
+  let domain = 'general';
+  if (codeSignals >= 2) domain = 'code';
+  else if (creativity >= 6) domain = 'creative';
+  else if (recency >= 6) domain = 'current_events';
+  else if (emotional_weight >= 5) domain = 'emotional';
+  else if (t.match(/\b(legal|law|contract|lawsuit|rights|regulation)\b/)) domain = 'legal';
+  else if (t.match(/\b(medical|doctor|symptom|diagnosis|health|drug|medicine)\b/)) domain = 'medical';
+  else if (t.match(/\b(research|study|paper|academic|journal|cite|evidence)\b/)) domain = 'research';
+
+  return {
+    complexity,
+    creativity,
+    speed,
+    cost_tolerance: costTolerance,
+    emotional_weight,
+    recency,
+    context_load,
+    domain,
+  };
+}

package/src/index.d.ts

@@ -0,0 +1,80 @@
+/**
+ * orbit-ai · TypeScript definitions
+ * 777 · 555 · 333
+ */
+
+export interface QueryScores {
+  complexity: number;        // 0-10
+  creativity: number;        // 0-10
+  speed: number;             // 0-10
+  cost_tolerance: number;    // 0-10
+  emotional_weight: number;  // 0-10
+  recency: number;           // 0-10
+  context_load: number;      // 0-10
+  domain: 'general' | 'code' | 'creative' | 'current_events' | 'emotional' | 'legal' | 'medical' | 'research';
+}
+
+export interface ModelInfo {
+  id: string;
+  name: string;
+  provider: 'anthropic' | 'openai' | 'google' | 'xai' | 'meta';
+  costPer1M: number;
+  strengths: string[];
+  maxContext: number;
+  tier: 'very_low' | 'low' | 'medium' | 'high';
+}
+
+export interface SavingsInfo {
+  premiumCost: number;
+  actualCost: number;
+  savings: number;
+  reductionPct: number;
+}
+
+export interface RoutingDecision {
+  model: ModelInfo;
+  reason: string;
+  rule: string;
+  scores: QueryScores;
+  savings: SavingsInfo;
+  timestamp: string;
+}
+
+export interface OrbitConfig {
+  apiKey?: string;
+  cost_tolerance?: 'low' | 'medium' | 'high';
+  blocked_models?: string[];
+  log?: boolean;
+  on_route?: (decision: RoutingDecision) => void;
+  anthropic_key?: string;
+  openai_key?: string;
+  google_key?: string;
+}
+
+export interface RouteOptions {
+  cost_tolerance?: 'low' | 'medium' | 'high';
+  estimated_tokens?: number;
+  blocked_models?: string[];
+}
+
+export interface SessionStats {
+  total_queries: number;
+  total_savings: number;
+  total_savings_formatted: string;
+  model_usage: Record<string, number>;
+}
+
+export declare class OrbitClient {
+  constructor(config?: OrbitConfig);
+  route(text: string, options?: RouteOptions): RoutingDecision;
+  fingerprint(text: string): QueryScores;
+  stats(): SessionStats;
+}
+
+export declare function fingerprint(text: string): QueryScores;
+export declare function route(scores: QueryScores, config?: OrbitConfig): { model: ModelInfo; reason: string; rule: string };
+export declare function calculateSavings(model: ModelInfo, estimatedTokens?: number): SavingsInfo;
+export declare const MODEL_MATRIX: Record<string, ModelInfo>;
+
+declare const orbit: OrbitClient;
+export default orbit;

package/src/index.js

@@ -0,0 +1,122 @@
+/**
+ * orbit-ai · v0.1.0
+ * Intelligent AI model routing. Drop in. Save 85%.
+ * 
+ * https://orbit-model-flow.base44.app
+ * github.com/gabrielsmartin/orbit
+ * 
+ * 777 · 555 · 333
+ */
+
+import { fingerprint } from './fingerprint.js';
+import { route, calculateSavings, MODEL_MATRIX } from './router.js';
+
+export { fingerprint, route, calculateSavings, MODEL_MATRIX };
+
+/**
+ * OrbitClient — the main class
+ * 
+ * @example
+ * import { OrbitClient } from 'orbit-ai'
+ * const orbit = new OrbitClient({ apiKey: 'your-orbit-key' })
+ * const result = await orbit.query("explain quantum entanglement simply")
+ */
+export class OrbitClient {
+  constructor(config = {}) {
+    this.config = {
+      cost_tolerance: config.cost_tolerance || 'medium', // 'low' | 'medium' | 'high'
+      blocked_models: config.blocked_models || [],
+      api_key: config.apiKey || config.api_key || null,
+      log: config.log !== false, // log routing decisions by default
+      on_route: config.on_route || null, // callback: (decision) => void
+      // Provider API keys (optional — falls back to env vars)
+      anthropic_key: config.anthropic_key || null,
+      openai_key: config.openai_key || null,
+      google_key: config.google_key || null,
+    };
+
+    this._stats = {
+      total_queries: 0,
+      total_savings: 0,
+      model_usage: {},
+    };
+  }
+
+  /**
+   * Route a query to the optimal model
+   * Returns the routing decision — you call the model yourself with your keys
+   * 
+   * @param {string} text - The query text
+   * @param {Object} options - Override options for this query
+   * @returns {Object} decision - { model, reason, rule, scores, savings }
+   */
+  route(text, options = {}) {
+    const scores = fingerprint(text);
+
+    // Allow per-query cost_tolerance override
+    if (options.cost_tolerance) {
+      scores.cost_tolerance = options.cost_tolerance === 'low' ? 2
+        : options.cost_tolerance === 'high' ? 9 : 5;
+    }
+
+    const config = { ...this.config, ...options };
+    const decision = route(scores, config);
+    const savings = calculateSavings(decision.model, options.estimated_tokens || 500);
+
+    const result = {
+      model: decision.model,
+      reason: decision.reason,
+      rule: decision.rule,
+      scores,
+      savings,
+      timestamp: new Date().toISOString(),
+    };
+
+    // Update stats
+    this._stats.total_queries++;
+    this._stats.total_savings += savings.savings;
+    const modelName = decision.model.name;
+    this._stats.model_usage[modelName] = (this._stats.model_usage[modelName] || 0) + 1;
+
+    // Log routing decision
+    if (this.config.log) {
+      console.log(`[ORBIT] → ${decision.model.name} | ${decision.rule} | saved $${savings.savings.toFixed(5)} (${savings.reductionPct}% reduction)`);
+    }
+
+    // Fire callback
+    if (this.config.on_route) {
+      this.config.on_route(result);
+    }
+
+    return result;
+  }
+
+  /**
+   * Get cumulative stats for this session
+   */
+  stats() {
+    return {
+      ...this._stats,
+      total_savings_formatted: `$${this._stats.total_savings.toFixed(4)}`,
+    };
+  }
+
+  /**
+   * Fingerprint a query without routing
+   * Useful for debugging or building custom routing logic on top
+   */
+  fingerprint(text) {
+    return fingerprint(text);
+  }
+}
+
+/**
+ * Default singleton client — zero config, ready to use
+ * 
+ * @example
+ * import orbit from 'orbit-ai'
+ * const decision = orbit.route("write a haiku about recursion")
+ * // → { model: { name: 'Claude Sonnet', ... }, reason: '...', savings: { ... } }
+ */
+const orbit = new OrbitClient();
+export default orbit;

package/src/router.js

@@ -0,0 +1,173 @@
+/**
+ * ORBIT · Selective Model Matching (SMM) Router
+ * Routes queries to optimal models based on 8-axis fingerprints
+ * 
+ * Proprietary routing logic — open SDK, closed engine weights
+ * 777 · 555 · 333
+ */
+
+export const MODEL_MATRIX = {
+  claude_sonnet: {
+    id: 'claude-sonnet-4-5',
+    name: 'Claude Sonnet',
+    provider: 'anthropic',
+    costPer1M: 15,
+    strengths: ['nuance', 'long_context', 'ethics', 'reasoning'],
+    maxContext: 200000,
+    tier: 'medium',
+  },
+  claude_haiku: {
+    id: 'claude-haiku-3-5',
+    name: 'Claude Haiku',
+    provider: 'anthropic',
+    costPer1M: 1,
+    strengths: ['speed', 'efficiency', 'summaries'],
+    maxContext: 200000,
+    tier: 'low',
+  },
+  gemini_flash: {
+    id: 'gemini-2.5-flash',
+    name: 'Gemini 2.5 Flash',
+    provider: 'google',
+    costPer1M: 0.5,
+    strengths: ['speed', 'multimodal', 'cost', 'volume'],
+    maxContext: 1000000,
+    tier: 'very_low',
+  },
+  gpt4o: {
+    id: 'gpt-4o',
+    name: 'GPT-4o',
+    provider: 'openai',
+    costPer1M: 30,
+    strengths: ['broad_knowledge', 'instruction_following', 'structured_output'],
+    maxContext: 128000,
+    tier: 'high',
+  },
+  gpt4o_mini: {
+    id: 'gpt-4o-mini',
+    name: 'GPT-4o Mini',
+    provider: 'openai',
+    costPer1M: 0.3,
+    strengths: ['classification', 'simple_tasks', 'cost'],
+    maxContext: 128000,
+    tier: 'very_low',
+  },
+  grok: {
+    id: 'grok-2',
+    name: 'Grok',
+    provider: 'xai',
+    costPer1M: 10,
+    strengths: ['realtime_web', 'trending', 'social_intelligence'],
+    maxContext: 131072,
+    tier: 'medium',
+  },
+};
+
+/**
+ * Core SMM routing logic
+ * Returns the selected model + reasoning
+ * 
+ * @param {Object} scores - 8-axis fingerprint scores
+ * @param {Object} config - User config (cost_tolerance override, blocked_models, etc.)
+ * @returns {Object} { model, reason, fallback }
+ */
+export function route(scores, config = {}) {
+  const {
+    complexity, creativity, speed, emotional_weight,
+    recency, context_load, domain, cost_tolerance
+  } = scores;
+
+  const blocked = config.blocked_models || [];
+  const preferLow = config.cost_tolerance === 'low' || cost_tolerance <= 3;
+  const preferHigh = config.cost_tolerance === 'high' || cost_tolerance >= 8;
+
+  // Rule 1: ETHICS FIRST — emotional/crisis queries always go to Claude
+  if (emotional_weight >= 6) {
+    return {
+      model: MODEL_MATRIX.claude_sonnet,
+      reason: 'Emotional weight detected — routing to Claude for ethics-first handling. Never use a cheap model for sensitive content.',
+      rule: 'ethics_first',
+    };
+  }
+
+  // Rule 2: Realtime / current events → Grok
+  if (recency >= 7 && !blocked.includes('grok')) {
+    return {
+      model: MODEL_MATRIX.grok,
+      reason: `High recency score (${recency}/10) — Grok has live web access for current events and trending topics.`,
+      rule: 'recency_grok',
+    };
+  }
+
+  // Rule 3: Long context load → Claude Sonnet (200k window)
+  if (context_load >= 8 && !blocked.includes('claude_sonnet')) {
+    return {
+      model: MODEL_MATRIX.claude_sonnet,
+      reason: `High context load (${context_load}/10) — Claude's 200k window is the only safe choice.`,
+      rule: 'context_claude',
+    };
+  }
+
+  // Rule 4: High complexity code/reasoning
+  if (complexity >= 7 && domain === 'code' && !blocked.includes('claude_sonnet')) {
+    return {
+      model: MODEL_MATRIX.claude_sonnet,
+      reason: `Complex code task (complexity ${complexity}/10) — Claude Sonnet for deep reasoning and long context.`,
+      rule: 'complex_code',
+    };
+  }
+
+  // Rule 5: High complexity general → GPT-4o (if cost tolerance allows)
+  if (complexity >= 7 && !preferLow && !blocked.includes('gpt4o')) {
+    return {
+      model: MODEL_MATRIX.gpt4o,
+      reason: `High complexity (${complexity}/10) — GPT-4o for broad knowledge and structured output.`,
+      rule: 'complex_gpt4o',
+    };
+  }
+
+  // Rule 6: Creative writing → Claude Sonnet
+  if (creativity >= 5 && !blocked.includes('claude_sonnet')) {
+    return {
+      model: MODEL_MATRIX.claude_sonnet,
+      reason: `High creativity score (${creativity}/10) — Claude Sonnet for nuanced creative writing.`,
+      rule: 'creative_claude',
+    };
+  }
+
+  // Rule 7: Cost sensitive OR simple queries → Gemini Flash
+  if ((preferLow || complexity <= 3) && !blocked.includes('gemini_flash')) {
+    return {
+      model: MODEL_MATRIX.gemini_flash,
+      reason: `Low complexity (${complexity}/10) — Gemini 2.5 Flash delivers 95% quality at 2% of GPT-4o cost.`,
+      rule: 'cost_gemini',
+    };
+  }
+
+  // Rule 8: Medium complexity → Claude Haiku (fast + cheap + capable)
+  if (complexity <= 5 && !blocked.includes('claude_haiku')) {
+    return {
+      model: MODEL_MATRIX.claude_haiku,
+      reason: `Medium complexity (${complexity}/10) — Claude Haiku balances speed, cost, and quality.`,
+      rule: 'medium_haiku',
+    };
+  }
+
+  // Default: Claude Sonnet (safest general choice)
+  return {
+    model: MODEL_MATRIX.claude_sonnet,
+    reason: 'Default routing — Claude Sonnet for reliable, high-quality responses.',
+    rule: 'default',
+  };
+}
+
+/**
+ * Calculate savings vs always using GPT-4o (premium baseline)
+ */
+export function calculateSavings(selectedModel, estimatedTokens = 500) {
+  const premiumCost = (MODEL_MATRIX.gpt4o.costPer1M / 1_000_000) * estimatedTokens;
+  const actualCost = (selectedModel.costPer1M / 1_000_000) * estimatedTokens;
+  const savings = premiumCost - actualCost;
+  const reductionPct = Math.round((savings / premiumCost) * 100);
+  return { premiumCost, actualCost, savings, reductionPct };
+}