pi-hypercharm-provider 1.3.20 → 1.3.22

package/README.md

@@ -32,13 +32,13 @@ _Hyperoptimized coding models — DeepSeek, GLM, Kimi, Qwen, MiniMax, Gemma, GPT
 | DeepSeek V4 Pro | Text | 1.0M | 384K | $2.40 | $4.80 |
 | DeepSeek V4 Pro 0813 | Text | 1.0M | 262K | $1.44 | $4.31 |
 | DeepSeek V4.1 Flash | Text + Image | 1.0M | 26K | $0.30 | $1.20 |
-| Gemma 4 26B A4B | Text | 256K | 26K | $0.12 | $0.42 |
+| Gemma 4 26B A4B | Text | 256K | 26K | $0.10 | $0.36 |
 | GLM 5.3 | Text + Image | 1.0M | 262K | $1.52 | $4.79 |
 | GLM 5.3 Flash | Text + Image | 1.0M | 131K | $0.16 | $0.54 |
-| GLM-5 | Text | 203K | 20K | $0.86 | $2.78 |
-| GLM-5.1 | Text | 203K | 3K | $1.33 | $4.22 |
+| GLM-5 | Text | 203K | 20K | $0.86 | $2.75 |
+| GLM-5.1 | Text | 203K | 3K | $1.32 | $4.31 |
 | GLM-5.2 | Text | 1.0M | 33K | $1.52 | $4.79 |
-| gpt-oss-120b | Text | 128K | 13K | $0.18 | $0.68 |
+| gpt-oss-120b | Text | 128K | 13K | $0.17 | $0.66 |
 | Inkling | Text + Image | 1.0M | 33K | $1.09 | $4.41 |
 | Kimi K2 Thinking | Text | 262K | 26K | $0.60 | $2.50 |
 | Kimi K2.5 | Text | 262K | 26K | $0.56 | $2.94 |
@@ -46,8 +46,8 @@ _Hyperoptimized coding models — DeepSeek, GLM, Kimi, Qwen, MiniMax, Gemma, GPT
 | Kimi K2.7 Code | Text + Image | 262K | 16K | $1.03 | $4.36 |
 | Kimi K3 | Text + Image | 1.0M | 16K | $3.27 | $16.33 |
 | Llama 3.3 70B Instruct | Text | 128K | 13K | $0.61 | $1.04 |
-| Llama 4 Maverick 17B 128E Instruct FP8 | Text | 430K | 43K | $0.27 | $0.90 |
-| MiniMax M2.7 | Text | 262K | 7K | $0.40 | $1.50 |
+| Llama 4 Maverick 17B 128E Instruct FP8 | Text | 430K | 43K | $0.26 | $0.84 |
+| MiniMax M2.7 | Text | 262K | 7K | $0.40 | $1.46 |
 | MiniMax M3 | Text + Image | 512K | 512K | $0.33 | $1.31 |
 | Qwen3 Coder 480B A35B Instruct INT4 Mixed AR | Text | 106K | 11K | $0.45 | $2.15 |
 | Qwen3 Next 80B A3B Instruct | Text | 262K | 26K | $0.12 | $1.14 |
@@ -225,11 +225,18 @@ Model metadata matches Charm's official [`@charmland/pi-hyper-provider`](https:/
 - `cacheRead` from `cost_per_1m_out_cached` (discounted cached-output price) and `cacheWrite` from `cost_per_1m_in_cached`, exactly as the official provider maps them
 - `/v1/credits` balance accepted in either hypercredits (`balance`) or USD (`balance_usd`, converted at the observed 20 hc = $1 rate)
 
-`patch.json` is reserved only for a verified provider regression and is currently empty.
+`patch.json` is reserved for verified provider regressions. It currently re-exposes Pi's `max` thinking level on the DeepSeek V4 models whose catalog `reasoning_levels` dropped it (`deepseek-v4-flash`, `deepseek-v4-pro`, `deepseek-v4.1-flash`), matching the levels their pinned snapshots still publish.
 
 ### Patch Overrides
 
-`patch.json` is applied on top of `models.json` only for verified endpoint corrections. It is currently empty because every live field comes from Charm's canonical provider catalog.
+`patch.json` is applied on top of `models.json` only for verified endpoint corrections; every other live field comes from Charm's canonical provider catalog.
+
+| Model | Override | Why |
+|-------|----------|-----|
+| `deepseek-v4-flash`, `deepseek-v4-pro` | `thinkingLevelMap.max = "max"` | The catalog now publishes only `high,xhigh`, while the pinned `-0731`/`-0813` snapshots publish `max`. |
+| `deepseek-v4.1-flash` | `thinkingLevelMap.max = "max"` | The catalog publishes `low,high,xhigh`; `max` mirrors the sibling DeepSeek V4 entries. |
+
+Each entry carries the model's full level map, because `patch.json` replaces `thinkingLevelMap` wholesale rather than merging it.
 
 ## Updating Models
 

package/models.json

@@ -172,9 +172,9 @@
       "text"
     ],
     "cost": {
-      "input": 0.122,
-      "output": 0.42,
-      "cacheRead": 0.061,
+      "input": 0.102,
+      "output": 0.356,
+      "cacheRead": 0.051,
       "cacheWrite": 0
     },
     "contextWindow": 256000,
@@ -268,7 +268,7 @@
     ],
     "cost": {
       "input": 0.86,
-      "output": 2.784,
+      "output": 2.752,
       "cacheRead": 0.43,
       "cacheWrite": 0
     },
@@ -298,9 +298,9 @@
       "text"
     ],
     "cost": {
-      "input": 1.326,
-      "output": 4.22,
-      "cacheRead": 0.663,
+      "input": 1.318,
+      "output": 4.308,
+      "cacheRead": 0.659,
       "cacheWrite": 0
     },
     "contextWindow": 202750,
@@ -360,9 +360,9 @@
       "text"
     ],
     "cost": {
-      "input": 0.178,
-      "output": 0.68,
-      "cacheRead": 0.089,
+      "input": 0.168,
+      "output": 0.66,
+      "cacheRead": 0.084,
       "cacheWrite": 0
     },
     "contextWindow": 128072,
@@ -612,9 +612,9 @@
       "text"
     ],
     "cost": {
-      "input": 0.274,
-      "output": 0.8992,
-      "cacheRead": 0.137,
+      "input": 0.255,
+      "output": 0.8365,
+      "cacheRead": 0.1275,
       "cacheWrite": 0
     },
     "contextWindow": 430000,
@@ -634,9 +634,9 @@
       "text"
     ],
     "cost": {
-      "input": 0.404,
-      "output": 1.496,
-      "cacheRead": 0.202,
+      "input": 0.396,
+      "output": 1.464,
+      "cacheRead": 0.198,
       "cacheWrite": 0
     },
     "contextWindow": 262100,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-hypercharm-provider",
-  "version": "1.3.20",
+  "version": "1.3.22",
   "description": "HyperCharm provider extension for pi - Access DeepSeek, GLM, Kimi, Qwen, MiniMax, Gemma, and GPT-OSS models through the Charm Hyper API",
   "type": "module",
   "main": "index.ts",

package/patch.json

@@ -1 +1,35 @@
-{}
+{
+  "deepseek-v4-flash": {
+    "thinkingLevelMap": {
+      "off": null,
+      "minimal": null,
+      "low": null,
+      "medium": null,
+      "high": "high",
+      "xhigh": "xhigh",
+      "max": "max"
+    }
+  },
+  "deepseek-v4-pro": {
+    "thinkingLevelMap": {
+      "off": null,
+      "minimal": null,
+      "low": null,
+      "medium": null,
+      "high": "high",
+      "xhigh": "xhigh",
+      "max": "max"
+    }
+  },
+  "deepseek-v4.1-flash": {
+    "thinkingLevelMap": {
+      "off": null,
+      "minimal": null,
+      "low": "low",
+      "medium": null,
+      "high": "high",
+      "xhigh": "xhigh",
+      "max": "max"
+    }
+  }
+}

package/scripts/update-models.js

@@ -9,8 +9,10 @@
  *
  * The endpoint provides canonical names, $/M pricing, context/output limits,
  * can_reason, optional reasoning levels, and attachment support. models.json is
- * pure API data. patch.json is reserved for verified endpoint regressions and
- * currently contains no overrides.
+ * pure API data. patch.json is reserved for verified endpoint regressions; it
+ * currently restores Pi's max thinking level on the DeepSeek V4 models whose
+ * catalog reasoning_levels no longer publish it. Patch entries replace
+ * thinkingLevelMap wholesale, so each one carries the model's full level map.
  *
  * Merge order for README: models.json → apply patch.json → merge custom-models.json
  *