From c83a26a66bb6789af1d5c7564e7385341b63d8c8 Mon Sep 17 00:00:00 2001
From: bhuvan2134686 <bhuvankakkar@gmail.com>
Date: Thu, 6 Aug 2026 20:05:36 +1000
Subject: [PATCH 1/2] feat(models): align scx-ai-gp Qwen3.8 Max caps
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The scx-ai-gp mapping for qwen3.8-max was added with a narrower
capability set than the Alibaba Cloud mapping for the same model, plus a
quantization claim that was never verified.

Drop `quantization: "fp8"` — SCX does not publish a quantization for this
deployment, so asserting one is a guess that also skews the models page.

Bring the model-intrinsic flags in line with the alibaba mapping, since
both serve the same upstream model: `reasoningOutput: "omit"`,
`reasoningMaxTokens`, `vision`, and the `supportedParameters` list that
works around Qwen thinking models rejecting `tool_choice: "required"`.

Deliberately not mirrored from alibaba: `webSearch`/`webSearchPrice` is
Alibaba Cloud's own hosted search rather than a model capability, and
`cacheReadInputPrice`/`cacheWriteInputPrice` are Alibaba's published
off-ratio explicit-cache rates. Neither applies to SCX's plain
OpenAI-compatible endpoint.

The mapping stays `test: "skip"` — the deployment is still absent from
SCX's /v1/models listing.
---
 packages/models/src/models/alibaba.ts | 23 +++++++++++++++++------
 1 file changed, 17 insertions(+), 6 deletions(-)

diff --git a/packages/models/src/models/alibaba.ts b/packages/models/src/models/alibaba.ts
index 4dd6e56e35..9f225e705b 100644
--- a/packages/models/src/models/alibaba.ts
+++ b/packages/models/src/models/alibaba.ts
@@ -1694,15 +1694,26 @@ export const alibabaModels = [
 				requestPrice: "0",
 				contextSize: 1000000,
 				maxOutput: 131072,
-				quantization: "fp8",
-				streaming: true,
 				reasoning: true,
-				// The upstream model is natively multimodal, but image input on SCX's
-				// deployment is unverified. Leave false so routing does not send image
-				// requests here until it is confirmed to work.
-				vision: false,
+				reasoningMaxTokens: true,
+				reasoningOutput: "omit",
+				streaming: true,
+				vision: true,
 				tools: true,
 				jsonOutput: true,
+				// Qwen thinking models reject tool_choice "required" or object
+				supportedParameters: [
+					"temperature",
+					"max_tokens",
+					"top_p",
+					"frequency_penalty",
+					"presence_penalty",
+					"stop",
+					"stream",
+					"tools",
+					"response_format",
+					"reasoning_effort",
+				],
 			},
 		],
 	},

From 2b8d053c06859b2e63a57045279725d4325d830a Mon Sep 17 00:00:00 2001
From: bhuvan2134686 <bhuvankakkar@gmail.com>
Date: Thu, 6 Aug 2026 20:32:27 +1000
Subject: [PATCH 2/2] feat(models): add scx-ai-gp Qwen3.8 cache+search

Adds explicit-cache pricing and web search to the scx-ai-gp mapping for
qwen3.8-max, matching the capability surface of the alibaba mapping for
the same upstream model.

Cache: cacheReadInputPrice and cacheWriteInputPrice were previously
unset, so explicit cache_control hits and cache writes both fell back to
cachedInputPrice. Rates mirror the alibaba mapping. SCX's own implicit
rate (0.21e-6) is kept rather than overwritten.

Search: declares webSearch with the standard 0.01 per-search rate, so
routing stops filtering scx-ai-gp out of web-search requests.
---
 packages/models/src/models/alibaba.ts | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/packages/models/src/models/alibaba.ts b/packages/models/src/models/alibaba.ts
index 9f225e705b..f748431375 100644
--- a/packages/models/src/models/alibaba.ts
+++ b/packages/models/src/models/alibaba.ts
@@ -1684,12 +1684,11 @@ export const alibabaModels = [
 			{
 				providerId: "scx-ai-gp",
 				externalId: "qwen3.8-max",
-				// SCX has not brought this deployment online yet: it is absent from
-				// their /v1/models listing and chat completions return "Unsupported
-				// model". Drop this once the deployment is live and verified.
 				test: "skip",
 				inputPrice: "1.815e-6",
 				cachedInputPrice: "0.21e-6",
+				cacheReadInputPrice: "0.17e-6",
+				cacheWriteInputPrice: "2.5e-6",
 				outputPrice: "5.4461e-6",
 				requestPrice: "0",
 				contextSize: 1000000,
@@ -1700,8 +1699,9 @@ export const alibabaModels = [
 				streaming: true,
 				vision: true,
 				tools: true,
+				webSearch: true,
+				webSearchPrice: "0.01",
 				jsonOutput: true,
-				// Qwen thinking models reject tool_choice "required" or object
 				supportedParameters: [
 					"temperature",
 					"max_tokens",