@aggc/or-info 0.2.11 → 0.2.13

package/README.md

@@ -157,14 +157,23 @@ or-info compare openai/gpt-4o deepseek/deepseek-chat-v3-0324 --json
 or-info top --task coding             # Best coding models
 or-info top --task reasoning          # Best reasoning models
 or-info top --task general            # Best all-rounders
-or-info top --task vision             # Best vision models
+or-info top --task vision             # Best vision models (requires image input)
 or-info top --task cheap              # Best value for money
+or-info top --task premium            # Highest quality, ignoring price
 or-info top --task coding --budget 2  # Best coders under $2/M output
 or-info top --task general --limit 10
 ```
 
-Ranking combines LMArena ELO with price. `--task vision` and `--task coding` additionally
-filter for models that support the required capability (image input / tool use).
+Ranking combines LMArena ELO with price and context window size. Task behaviour:
+
+| Task | Price weight | Capability filter |
+|------|-------------|-------------------|
+| `general` | standard penalty | none |
+| `coding` | standard penalty | soft penalty (−15%) if no tool support |
+| `reasoning` | standard penalty | none |
+| `vision` | standard penalty | hard filter: image input required |
+| `cheap` | steep penalty | none |
+| `premium` | ignored | none |
 
 ### Cache management
 
@@ -184,7 +193,7 @@ or-info refresh         # Force-refresh OpenRouter catalog + LMArena ELO
 | `models.get` | Pricing, context, architecture, features and LMArena ELO for a model |
 | `models.list` | List models with optional filter, sort and limit |
 | `models.compare` | Side-by-side comparison of two models |
-| `models.top` | Ranked top models for coding/reasoning/general/vision/cheap |
+| `models.top` | Ranked top models for coding/reasoning/general/vision/cheap/premium |
 | `benchmarks.get` | LMArena ELO score, global rank, vote count and confidence interval for a model |
 | `cache.refresh` | Force-refresh OpenRouter catalog + LMArena ELO |
 

package/bin/or-info.mjs

@@ -22,7 +22,7 @@ function die(msg) {
   process.exit(1);
 }
 
-const TOP_TASKS = new Set(['coding', 'reasoning', 'general', 'vision', 'cheap']);
+const TOP_TASKS = new Set(['coding', 'reasoning', 'general', 'vision', 'cheap', 'premium']);
 
 function parsePositiveInteger(value) {
   const n = Number.parseInt(value, 10);

package/lib/openrouter.mjs

@@ -66,7 +66,11 @@ export async function fetchModels({ apiKey, force = false } = {}) {
 }
 
 export function findModel(models, id) {
-  return models.find((m) => m?.id === id) ?? null;
+  const exact = models.find((m) => m?.id === id);
+  if (exact) return exact;
+  const norm = (s) => s.replace(/\./g, '-');
+  const needle = norm(id);
+  return models.find((m) => m?.id && norm(m.id) === needle) ?? null;
 }
 
 export function pricePerMillion(model) {

package/lib/scorer.mjs

@@ -1,8 +1,8 @@
-import { pricePerMillion, supportsFeature } from './openrouter.mjs';
+import { pricePerMillion, supportsFeature, contextLength } from './openrouter.mjs';
 
-// ELO range observed on LMArena (2026): ~1050 (weak) to ~1500 (best)
-const ELO_MIN = 1050;
-const ELO_MAX = 1500;
+// ELO range observed on LMArena (2026): ~1000 (weak) to ~1540+ (best)
+const ELO_MIN = 1000;
+const ELO_MAX = 1600;
 
 function normaliseElo(elo) {
   return Math.max(0, Math.min(100, ((elo - ELO_MIN) / (ELO_MAX - ELO_MIN)) * 100));
@@ -18,32 +18,58 @@ function pricePenalty(outputPerM) {
   return Math.max(0.1, 1 - Math.log10(outputPerM + 1) * 0.15);
 }
 
-function requiresCapability(task) {
-  if (task === 'vision') return 'vision';
-  if (task === 'coding') return 'tools';
-  return null;
+// Steeper penalty for cheap task — more spread between free and expensive.
+// Free → 1.0, $1/M → 0.88, $5/M → 0.72, $20/M → 0.59
+function cheapPenalty(outputPerM) {
+  if (outputPerM === null || outputPerM === 0) return 1.0;
+  return Math.max(0.1, 1 - Math.log10(outputPerM + 1) * 0.25);
+}
+
+// Context window bonus [0.9, 1.0]. Larger context is more useful,
+// especially for coding where long files are common.
+function contextBonus(ctx, task) {
+  if (!ctx) return 0.9;
+  if (ctx >= 128_000) return 1.0;
+  if (ctx >= 64_000) return 0.97;
+  if (ctx >= 32_000) return 0.93;
+  return 0.9;
 }
 
 // Score a model for a task.
 // Returns { score, qualityScore } or null if not eligible.
 export function scoreForTask(model, eloEntry, task = 'general') {
-  const cap = requiresCapability(task);
-  if (cap && !supportsFeature(model, cap)) return null;
+  // Hard filter: vision requires vision capability
+  if (task === 'vision' && !supportsFeature(model, 'vision')) return null;
   if (!eloEntry?.elo) return null;
 
   const quality = normaliseElo(eloEntry.elo);
   const price = pricePerMillion(model);
-  const penalty = task === 'cheap'
-    ? pricePenalty(price.output) * 1.4   // aggressively favour cheap
-    : pricePenalty(price.output);
+  const ctx = contextLength(model);
+
+  // Price penalty: premium ignores price, cheap uses steep curve, others standard
+  let penalty;
+  if (task === 'premium') {
+    penalty = 1.0;
+  } else if (task === 'cheap') {
+    penalty = cheapPenalty(price.output);
+  } else {
+    penalty = pricePenalty(price.output);
+  }
+
+  const ctxB = contextBonus(ctx, task);
+
+  // Soft penalty for coding without tools (still eligible, just less ideal)
+  const capPenalty = (task === 'coding' && !supportsFeature(model, 'tools')) ? 0.85 : 1.0;
+
+  const rawScore = quality * penalty * ctxB * capPenalty;
 
   return {
-    score: Math.round(quality * penalty * 10) / 10,
+    score: Math.round(Math.min(100, rawScore) * 10) / 10,
     qualityScore: Math.round(quality * 10) / 10,
   };
 }
 
-const CATEGORY_FOR_TASK = { coding: 'coding', reasoning: 'math', vision: null };
+const CATEGORY_FOR_TASK = { coding: 'coding', reasoning: 'math', vision: null, premium: null };
 
 export function rankModels(models, allElo, { task = 'general', maxPricePerMOutput, limit = 5 } = {}) {
   // allElo may be a by-category map { overall: [...], coding: [...], ... }
@@ -69,7 +95,15 @@ export function rankModels(models, allElo, { task = 'general', maxPricePerMOutpu
     scored.push({ model, score: result.score, qualityScore: result.qualityScore, eloEntry });
   }
 
-  return scored.sort((a, b) => b.score - a.score).slice(0, limit);
+  // Dedup :free variants — keep highest-scoring variant per base model
+  const byBase = new Map();
+  for (const entry of scored) {
+    const baseId = entry.model.id.replace(/:free$/, '');
+    const prev = byBase.get(baseId);
+    if (!prev || entry.score > prev.score) byBase.set(baseId, entry);
+  }
+
+  return [...byBase.values()].sort((a, b) => b.score - a.score).slice(0, limit);
 }
 
 // Inline minimal name matching (mirrors lmarena.mjs logic without importing it)

package/mcp/server.mjs

@@ -147,7 +147,7 @@ const CANONICAL_TOOLS = [
       properties: {
         task: {
           type: 'string',
-          enum: ['coding', 'reasoning', 'general', 'vision', 'cheap'],
+          enum: ['coding', 'reasoning', 'general', 'vision', 'cheap', 'premium'],
           description: 'Task type to optimise for',
         },
         max_price_per_m_output: {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aggc/or-info",
-  "version": "0.2.11",
+  "version": "0.2.13",
   "description": "CLI + MCP server for OpenRouter models: prices, benchmarks, context and comparisons",
   "type": "module",
   "engines": {