diff --git a/services/discord-gateway/src/modules/ai-moderation/llmClient.ts b/services/discord-gateway/src/modules/ai-moderation/llmClient.ts
index 4309423..efbc58e 100644
--- a/services/discord-gateway/src/modules/ai-moderation/llmClient.ts
+++ b/services/discord-gateway/src/modules/ai-moderation/llmClient.ts
@@ -158,6 +158,12 @@ export interface LlmCallOpts {
   stream?: boolean;
   /** Optional AbortSignal to cancel the API request */
   signal?: AbortSignal;
+  /**
+   * Per-request timeout in ms. Falls back to the client-level default
+   * (60s) when omitted. Vision/image analysis passes a longer budget here
+   * so a single large-image call isn't killed early by the shared default.
+   */
+  timeout?: number;
 }
 
 /**
@@ -242,6 +248,7 @@ export async function llmChat(
         ) => {
           const response = await client.chat.completions.create(currentParams, {
             signal,
+            ...(opts.timeout ? { timeout: opts.timeout } : {}),
           });
           if (currentParams.stream) {
             let content = "";
@@ -352,6 +359,7 @@ export async function llmVision(
     top_p: 0.9,
     retries: 0,
     stream: true, // router always streams SSE; non-stream waits for full body and times out
+    timeout: config.AI_LLM_VISION_ANALYSIS_TIMEOUT_MS ?? 60_000,
   });
 
   if (!completion) return null;
diff --git a/services/discord-gateway/src/shared/config/index.ts b/services/discord-gateway/src/shared/config/index.ts
index 1659809..0a9f7f8 100644
--- a/services/discord-gateway/src/shared/config/index.ts
+++ b/services/discord-gateway/src/shared/config/index.ts
@@ -189,6 +189,15 @@ export const configSchema = z
       .int()
       .positive()
       .default(60000),
+    // Standalone image/sticker/emoji vision analysis (analyzeSingleMediaImage
+    // → llmVision → llmChat). Decoupled from the media *batch* timeout above so
+    // a single vision call can be tuned independently. 1 minute by default —
+    // vision models (especially behind a router) need headroom for large images.
+    AI_LLM_VISION_ANALYSIS_TIMEOUT_MS: z.coerce
+      .number()
+      .int()
+      .positive()
+      .default(60000),
     // Text-only moderation batches are cheaper than media (no downloads /
     // vision pre-pass), so they get their own (shorter) timeout instead of
     // being tied to the media budget.