@hiper2d/ai-agents 0.38.0 → 0.39.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/dist/index.mjs CHANGED
@@ -1279,16 +1279,25 @@ var SupportedAiModels = {
1279
1279
  // Speed tags from the 2026-08-05 live measurements (two samples each): plus 17.4s/14.5s,
1280
1280
  // flash 14.3s/16.4s (both slow); max 30.6s/100.5s — its latency tracks how long it decides
1281
1281
  // to think (4.2K reasoning tokens on the slow run), hence the budget cap below.
1282
+ //
1283
+ // `qwen-max` repointed 2026-11-16 from qwen3.8-max ($2/$6, hits $0.25) to qwen4-max
1284
+ // ($1.50/$4.50, hits $0.075; not yet in GET /models, but the chat endpoint serves it and
1285
+ // echoes the name). Probe on a real day-4 prompt (32K-char system prompt, ~10K tokens in,
1286
+ // cold cache, budget 1024, one call at a time): discussion 7-9s vs 5-7s, vote 9s vs 24s
1287
+ // (3.8 Max spent ~1,000 output tokens on a vote, 4 Max ~220), night action 12s vs 17s,
1288
+ // end-of-day notes 33s vs 34s; every answer valid through the shared schema. Three
1289
+ // identical calls sent AT ONCE timed out on both models (night action, notes) — the account
1290
+ // under a burst, not the model. The 'slow' tag is 3.8 Max's production grade (p90 26.2s);
1291
+ // re-grade from production turns once 4 Max has 100+.
1282
1292
  [LLM_CONSTANTS.QWEN_MAX]: {
1283
- displayName: "Qwen3.8 Max",
1284
- modelApiName: "qwen3.8-max",
1293
+ displayName: "Qwen4 Max",
1294
+ modelApiName: "qwen4-max",
1285
1295
  apiKeyName: API_KEY_CONSTANTS.QWEN,
1286
1296
  hasThinking: true,
1287
1297
  temperature: 0.7,
1288
- // Caps `thinking_budget` to bound the 30–100s latency variance. The same knob works on
1289
- // the 3.7 models (verified live) — add it to their entries if they ever need taming.
1298
+ // Caps `thinking_budget` to bound the latency variance (3.8 Max swung 30–100s without
1299
+ // it). Uncapped, 4 Max's vote went 9s → 14s.
1290
1300
  thinkingBudgetTokens: 1024,
1291
- // Capped it measures 25-26s → the >25s tier.
1292
1301
  tags: ["slow"]
1293
1302
  },
1294
1303
  // qwen3.8-flash replaced qwen3.7-flash on 2026-08-30 (same 1M context, 128k max output);
@@ -1642,10 +1651,11 @@ var MODEL_PRICING = {
1642
1651
  // $0.016. Neither has input-length tiers (the tier column is "-" for both). These
1643
1652
  // published cached rates supersede the 20%-of-input rule charged before 2026-08-30; we
1644
1653
  // still don't send explicit cache_control.
1654
+ // qwen4-max (2026-11-16, same page): $1.50/$4.50, implicit-cache hits $0.075, no tiers.
1645
1655
  [SupportedAiModels[LLM_CONSTANTS.QWEN_MAX].modelApiName]: {
1646
- inputPrice: 2,
1647
- outputPrice: 6,
1648
- cacheHitPrice: 0.25
1656
+ inputPrice: 1.5,
1657
+ outputPrice: 4.5,
1658
+ cacheHitPrice: 0.075
1649
1659
  },
1650
1660
  [SupportedAiModels[LLM_CONSTANTS.QWEN_FLASH].modelApiName]: {
1651
1661
  inputPrice: 0.15,