diff --git a/lib/api/__tests__/chat-stream-helpers-fallback.test.ts b/lib/api/__tests__/chat-stream-helpers-fallback.test.ts
index 6cb507a45..e4ab62e81 100644
--- a/lib/api/__tests__/chat-stream-helpers-fallback.test.ts
+++ b/lib/api/__tests__/chat-stream-helpers-fallback.test.ts
@@ -41,7 +41,6 @@ const DEEPSEEK_V4_PRO_0813_SLUG = "deepseek/deepseek-v4-pro-0813";
 const HIGH_REASONING_ROUTES = [
   "ask-model",
   "agent-model",
-  "agent-model-free",
   "model-grok-4.5-pro",
   "model-grok-4.6",
   "model-grok-4.6-pro",
@@ -51,7 +50,6 @@ const HIGH_REASONING_ROUTES = [
   "model-opus-4.6",
   "model-glm-5.2",
   "model-glm-5.3",
-  "model-glm-5.3-flash-agent",
   "model-kimi-k3",
   "fallback-agent-model",
   "fallback-ask-model",
@@ -175,12 +173,12 @@ describe("buildProviderOptions fallback chain", () => {
   });
 
   it.each([
-    ["model-glm-5.3-flash", "low"],
+    ["model-glm-5.3-flash", "high"],
     ["model-glm-5.3-flash-pro", "high"],
   ] as const)(
-    "routes %s directly with %s reasoning and DeepSeek Vision fallback",
+    "routes Ask %s directly with %s reasoning and DeepSeek Vision fallback",
     (modelName, effort) => {
-      const opts = buildProviderOptions(true, "user-1", modelName, "agent");
+      const opts = buildProviderOptions(false, "user-1", modelName, "ask");
       expect(opts.openrouter).toMatchObject({
         reasoning: { enabled: true, effort },
         models: [DEEPSEEK_VISION_SLUG],
@@ -281,7 +279,7 @@ describe("buildProviderOptions fallback chain", () => {
     expect(opts.openrouter.models).toHaveLength(3);
   });
 
-  it("runs free Agent on GLM Flash high and falls back through DeepSeek Flash, Pro 0813, and GLM 5.3", () => {
+  it("runs free Agent on the GLM Flash default and falls back through DeepSeek Flash, Pro 0813, and GLM 5.3", () => {
     const opts = buildProviderOptions(
       true,
       "user-1",
@@ -289,14 +287,14 @@ describe("buildProviderOptions fallback chain", () => {
       "agent",
     );
     expect(opts.openrouter).toMatchObject({
-      reasoning: { enabled: true, effort: "high" },
       provider: { sort: "latency", data_collection: "deny" },
       models: [DEEPSEEK_FLASH_SLUG, DEEPSEEK_V4_PRO_0813_SLUG, GLM_SLUG],
       user: "user-1",
     });
+    expect(opts.openrouter).not.toHaveProperty("reasoning");
   });
 
-  it("uses the same high-reasoning fallback route for the paid GLM Agent treatment", () => {
+  it("uses provider-default reasoning for the paid GLM Agent route", () => {
     const opts = buildProviderOptions(
       true,
       "user-1",
@@ -304,13 +302,29 @@ describe("buildProviderOptions fallback chain", () => {
       "agent",
     );
     expect(opts.openrouter).toMatchObject({
-      reasoning: { enabled: true, effort: "high" },
       provider: { sort: "latency", data_collection: "deny" },
       models: [DEEPSEEK_FLASH_SLUG, DEEPSEEK_V4_PRO_0813_SLUG, GLM_SLUG],
       user: "user-1",
     });
+    expect(opts.openrouter).not.toHaveProperty("reasoning");
   });
 
+  it.each([
+    "agent-model-free",
+    "model-glm-5.3-flash-agent",
+    "model-glm-5.3-flash",
+    "model-glm-5.3-flash-pro",
+  ] as const)(
+    "keeps Agent %s on provider-default reasoning even when an override is supplied",
+    (modelName) => {
+      const opts = buildProviderOptions(true, "user-1", modelName, "agent", {
+        reasoningOverride: { enabled: true, effort: "high" },
+      });
+
+      expect(opts.openrouter).not.toHaveProperty("reasoning");
+    },
+  );
+
   it("falls back from explicit DeepSeek Pro ask model through GLM 5.3 then Kimi K3", () => {
     const opts = buildProviderOptions(
       false,
diff --git a/lib/api/chat-stream-helpers.ts b/lib/api/chat-stream-helpers.ts
index 4da378cff..b8ff66fa0 100644
--- a/lib/api/chat-stream-helpers.ts
+++ b/lib/api/chat-stream-helpers.ts
@@ -685,7 +685,6 @@ export function isExplicitDeepSeekProSelectionForRetry({
 }
 
 const HIGH_REASONING_MODELS = [
-  "agent-model-free",
   "model-grok-4.5-pro",
   "model-grok-4.6",
   "model-grok-4.6-pro",
@@ -694,7 +693,6 @@ const HIGH_REASONING_MODELS = [
   "model-glm-5.2",
   "model-glm-5.3",
   "model-glm-5.3-flash-pro",
-  "model-glm-5.3-flash-agent",
   "model-opus-4.6",
 ] as const satisfies readonly ModelName[];
 
@@ -971,10 +969,13 @@ export function buildProviderOptions(
       })
     : fallbackSlugs;
   // OpenRouter applies one reasoning configuration to both the primary model
-  // and every provider fallback. Standard GLM vision uses low, legacy Standard
-  // Grok vision uses medium, and Pro/full reasoning routes remain high.
+  // and every provider fallback. Ask GLM vision uses high, legacy Standard Grok
+  // vision uses medium, and Pro/full reasoning routes remain high. Agent GLM
+  // Flash routes omit this option so each provider model uses its default.
   const isMediumGrok45Vision = modelName === "model-grok-4.5" && isGrok45;
   const isStandardGlmFlashVision = modelName === "model-glm-5.3-flash";
+  const usesDefaultGlmFlashAgentReasoning =
+    mode === "agent" && modelId === GLM_5_3_FLASH_SLUG;
   const routesThroughHighReasoningModel =
     isGrok45 ||
     isGrok46 ||
@@ -987,7 +988,7 @@ export function buildProviderOptions(
   const reasoning = isStandardGlmFlashVision
     ? {
         enabled: true,
-        effort: "low",
+        effort: "high",
       }
     : isMediumGrok45Vision
       ? {
@@ -1016,7 +1017,7 @@ export function buildProviderOptions(
 
   return {
     openrouter: {
-      reasoning,
+      ...(!usesDefaultGlmFlashAgentReasoning && { reasoning }),
       ...(options.hasPdfAttachments && isDeepSeekV4
         ? {
             plugins: [