diff --git a/litellm/litellm_core_utils/health_check_helpers.py b/litellm/litellm_core_utils/health_check_helpers.py
index a0f027cd58f..d1381e06bcf 100644
--- a/litellm/litellm_core_utils/health_check_helpers.py
+++ b/litellm/litellm_core_utils/health_check_helpers.py
@@ -251,7 +251,7 @@ class HealthCheckHelpers:
             ),
             "responses": lambda: litellm.aresponses(
                 **_filter_model_params(model_params=model_params),
-                input=prompt or "test",
+                input=[{"role": "user", "content": prompt or "test"}],
             ),
             "ocr": lambda: litellm.aocr(
                 **_filter_model_params(model_params=model_params),
diff --git a/litellm/main.py b/litellm/main.py
index 6c85adf3ae8..3923dc91701 100644
--- a/litellm/main.py
+++ b/litellm/main.py
@@ -8755,6 +8755,10 @@ async def ahealth_check(
             mode = litellm.model_cost[model].get("mode")
 
         model_params["cache"] = {"no-cache": True}  # don't used cached responses for making health check calls
+        # chatgpt provider uses the Responses API exclusively; default to "responses" so the
+        # health check doesn't send a Chat Completions payload that the backend rejects.
+        if mode is None and custom_llm_provider == "chatgpt":
+            mode = "responses"
         mode = mode or "chat"
         if "*" in model:
             return await HealthCheckHelpers.ahealth_check_wildcard_models(