diff --git a/lib/hooks/image-analyzer-transformer.cjs b/lib/hooks/image-analyzer-transformer.cjs
index b041dd17..8fa3686c 100755
--- a/lib/hooks/image-analyzer-transformer.cjs
+++ b/lib/hooks/image-analyzer-transformer.cjs
@@ -177,16 +177,16 @@ function getModelsToTry() {
     seen.add(providerModels[currentProvider]);
   }
 
-  // 2. Default model
-  if (!seen.has(DEFAULT_MODEL)) {
+  // 2. Default model — only when no provider-specific model is configured,
+  // since DEFAULT_MODEL (gemini-2.5-flash) may not be routable on the
+  // current provider's CLIProxy endpoint (e.g. codex, claude)
+  if (models.length === 0 && !seen.has(DEFAULT_MODEL)) {
     models.push(DEFAULT_MODEL);
     seen.add(DEFAULT_MODEL);
   }
 
-  // 3. ANTHROPIC_MODEL fallback
-  if (anthropicModel && !seen.has(anthropicModel)) {
-    models.push(anthropicModel);
-  }
+  // 3. ANTHROPIC_MODEL fallback — skip, the main chat model is not
+  // necessarily vision-capable and adds unnecessary retry latency
 
   return models;
 }
@@ -382,7 +382,10 @@ function analyzeViaCliProxy(base64Data, mediaType, model, timeoutMs) {
 
           try {
             const response = JSON.parse(data);
-            const text = response.content?.[0]?.text;
+            // Find the first text block — skip thinking blocks which use
+            // .thinking instead of .text (e.g. codex models with thinking enabled)
+            const textBlock = response.content?.find(b => b.type === 'text' && b.text);
+            const text = textBlock?.text;
 
             if (!text) {
               reject(new Error('No text content in response'));