diff --git a/pi-extension/index.ts b/pi-extension/index.ts
index 4dcc6be..32635b0 100644
--- a/pi-extension/index.ts
+++ b/pi-extension/index.ts
@@ -3,6 +3,7 @@ import {
   type ExtensionAPI,
   type ExtensionContext,
 } from "@mariozechner/pi-coding-agent";
+import { getModel } from "@earendil-works/pi-ai";
 import { existsSync, readFileSync } from "node:fs";
 import { dirname, join } from "node:path";
 import { fileURLToPath } from "node:url";
@@ -59,6 +60,15 @@ function isIntegrationBaseUrl(baseUrl: string | undefined, info: IntegrationProv
 
 type CurrentModel = NonNullable<ExtensionContext["model"]>;
 
+function builtInModelCapabilities(provider: string, modelID: string) {
+  const model = getModel(provider, modelID);
+  if (!model) return undefined;
+  return {
+    reasoning: model.reasoning,
+    thinkingLevelMap: model.thinkingLevelMap,
+  };
+}
+
 let procEnvCache: Map<string, string> | null = null;
 
 // envValue mirrors pi-ai's Bun compiled-binary workaround for sandboxed Linux
@@ -118,7 +128,12 @@ export default async function (pi: ExtensionAPI) {
   const integrationNames = discovered.integrations.map((integration) => integration.name);
   const routeLabel = integrationProviderDisplayName(integrationNames);
   const availableIntegrationsLabel = integrationPromptAvailabilityLabel(integrationNames);
-  const integrationInfos = providerInfosFromIntegrationCatalogs(discovered.integrations, pricingCatalog);
+  const integrationInfos = providerInfosFromIntegrationCatalogs(
+    discovered.integrations,
+    pricingCatalog,
+    console.warn,
+    builtInModelCapabilities,
+  );
   if (discovered.found && integrationInfos.size === 0 && !disabled) {
     console.warn(`[pi-exe-dev] LLM integration discovered, but no supported models were available`);
   }
diff --git a/pi-extension/integration_catalog.test.ts b/pi-extension/integration_catalog.test.ts
index 508aae3..0ebdcf0 100644
--- a/pi-extension/integration_catalog.test.ts
+++ b/pi-extension/integration_catalog.test.ts
@@ -229,6 +229,110 @@ test("namespaces reflected xAI models and never creates routes from pricing meta
   ]);
 });
 
+test("inherits built-in thinking capabilities before aliasing integration models", () => {
+  const thinkingLevelMap = {
+    off: "none",
+    minimal: null,
+    low: "low",
+    medium: "medium",
+    high: "high",
+    xhigh: "xhigh",
+    max: "max",
+  };
+  const lookups: string[] = [];
+  const infos = providerInfosFromIntegrationCatalogs(
+    [
+      {
+        name: "llm",
+        baseURL: "https://llm.int.exe.xyz",
+        catalog: { schema_version: 1, models: [openAIGPTModel("chatgpt")] },
+      },
+    ],
+    undefined,
+    () => {},
+    (provider, modelID) => {
+      lookups.push(`${provider}/${modelID}`);
+      return { reasoning: true, thinkingLevelMap };
+    },
+  );
+
+  assert.deepEqual(lookups, ["openai/gpt-5.5"]);
+  const model = infos.get("exe-dev-openai")?.config.models?.[0];
+  assert.equal(model?.id, "gpt-5.5@llm");
+  assert.equal(model?.reasoning, true);
+  assert.deepEqual(model?.thinkingLevelMap, thinkingLevelMap);
+});
+
+test("preserves explicit false from built-in capabilities over pricing metadata", () => {
+  const pricingCatalog: Catalog = {
+    schemaVersion: 1,
+    providers: [
+      {
+        id: "openai",
+        path: "openai/v1",
+        models: [
+          {
+            id: "gpt-5.5",
+            reasoning: true,
+            cost: { input: 1, output: 2, cacheRead: 0.1, cacheWrite: 0.2 },
+          },
+        ],
+      },
+    ],
+  };
+  const infos = providerInfosFromIntegrationCatalogs(
+    [
+      {
+        name: "llm",
+        baseURL: "https://llm.int.exe.xyz",
+        catalog: { schema_version: 1, models: [openAIGPTModel("managed")] },
+      },
+    ],
+    pricingCatalog,
+    () => {},
+    () => ({ reasoning: false, thinkingLevelMap: undefined }),
+  );
+
+  const model = infos.get("exe-dev-openai")?.config.models?.[0];
+  assert.equal(model?.reasoning, false);
+  assert.equal("thinkingLevelMap" in (model ?? {}), false);
+});
+
+test("retains pricing-catalog reasoning for unknown built-in models", () => {
+  const pricingCatalog: Catalog = {
+    schemaVersion: 1,
+    providers: [
+      {
+        id: "xai",
+        path: "xai/v1",
+        models: [
+          {
+            id: "grok-4.5",
+            reasoning: true,
+            cost: { input: 2, output: 6, cacheRead: 0.5, cacheWrite: 0 },
+          },
+        ],
+      },
+    ],
+  };
+  const infos = providerInfosFromIntegrationCatalogs(
+    [
+      {
+        name: "llm",
+        baseURL: "https://llm.int.exe.xyz",
+        catalog: { schema_version: 1, models: [xaiGrokModel()] },
+      },
+    ],
+    pricingCatalog,
+    () => {},
+    () => undefined,
+  );
+
+  const model = infos.get("exe-dev-xai")?.config.models?.[0];
+  assert.equal(model?.reasoning, true);
+  assert.equal("thinkingLevelMap" in (model ?? {}), false);
+});
+
 test("routes arbitrary providers by client protocol priority", () => {
   const infos = providerInfosFromIntegrationCatalogs(
     [
diff --git a/pi-extension/integration_catalog.ts b/pi-extension/integration_catalog.ts
index c57d8d4..297cc28 100644
--- a/pi-extension/integration_catalog.ts
+++ b/pi-extension/integration_catalog.ts
@@ -108,6 +108,11 @@ type CompatBag = {
 
 export type JSONFetcher = (url: string) => Promise<unknown | undefined>;
 
+export type ModelCapabilityLookup = (
+  provider: string,
+  modelID: string,
+) => Pick<ProviderModelConfig, "reasoning" | "thinkingLevelMap"> | undefined;
+
 // Both the integration catalog and the bundled pricing sidecar currently use
 // schema version 1. Unknown shapes are ignored rather than becoming routes.
 export const SCHEMA_VERSION = 1;
@@ -322,6 +327,7 @@ function configFromIntegrationModel(
   integration: DiscoveredIntegration,
   model: IntegrationModel,
   fallback: CatalogModel | undefined,
+  lookupCapabilities: ModelCapabilityLookup,
 ): ProviderModelConfig | undefined {
   if (!validIntegrationProviderID(model.provider)) return undefined;
   const adapter = integrationAPIAdapter(model);
@@ -330,13 +336,17 @@ function configFromIntegrationModel(
 
   const modelID = integrationModelID(model);
   if (!modelID) return undefined;
+  const capabilities = lookupCapabilities(model.provider, modelID);
   const compat = sanitizeCompat(fallback?.compat, model.provider, modelID);
   return {
     id: modelID,
     name: model.name || fallback?.name || modelID,
     api: adapter.piAPI,
     baseUrl,
-    reasoning: fallback?.reasoning ?? false,
+    reasoning: capabilities?.reasoning ?? fallback?.reasoning ?? false,
+    ...(capabilities?.thinkingLevelMap !== undefined
+      ? { thinkingLevelMap: capabilities.thinkingLevelMap }
+      : {}),
     input: inputModalities(model, fallback),
     contextWindow: model.limits?.context_window ?? fallback?.contextWindow ?? 128000,
     maxTokens: model.limits?.max_output_tokens ?? fallback?.maxTokens ?? 4096,
@@ -381,6 +391,7 @@ export function providerInfosFromIntegrationCatalogs(
   integrations: DiscoveredIntegration[],
   pricingCatalog: Catalog | undefined,
   warn: WarnFn = console.warn,
+  lookupCapabilities: ModelCapabilityLookup = () => undefined,
 ): Map<string, IntegrationProviderInfo> {
   const costs = costCatalogIndex(pricingCatalog);
   const grouped = new Map<string, { names: string[]; candidates: IntegrationModelCandidate[] }>();
@@ -393,7 +404,7 @@ export function providerInfosFromIntegrationCatalogs(
       const modelID = integrationModelID(model);
       if (!modelID) continue;
       const fallback = fallbackCatalogModel(costs, provider, modelID, model.id);
-      const config = configFromIntegrationModel(integration, model, fallback);
+      const config = configFromIntegrationModel(integration, model, fallback, lookupCapabilities);
       if (!config) continue;
       if (!fallback) warnedMissingPricing.add(`${provider}/${modelID}`);