diff --git a/src/agents/run_internal/items.py b/src/agents/run_internal/items.py
index aadba1d361..042cfb4854 100644
--- a/src/agents/run_internal/items.py
+++ b/src/agents/run_internal/items.py
@@ -67,13 +67,26 @@ def run_item_to_input_item(
     run_item: RunItem,
     reasoning_item_id_policy: ReasoningItemIdPolicy | None = None,
 ) -> TResponseInputItem | None:
-    """Convert a run item to model input, optionally stripping reasoning IDs."""
+    """Convert a run item to model input, optionally stripping reasoning IDs.
+
+    Returns ``None`` for items that should be omitted from model input, including
+    tool-approval placeholders and reasoning items whose ``summary`` list is empty
+    (e.g. encrypted-reasoning turns where the model emits no visible summary text).
+    Persisting an empty reasoning item to the Conversations API causes a 400 error,
+    so we skip them rather than forwarding a ``{"type": "reasoning", "summary": []}``
+    payload that the API would reject.
+    """
     if run_item.type == "tool_approval_item":
         return None
     to_input = getattr(run_item, "to_input_item", None)
     input_item = to_input() if callable(to_input) else cast(TResponseInputItem, run_item.raw_item)
     if isinstance(input_item, dict) and input_item.get("status") is None:
         input_item = {k: v for k, v in input_item.items() if k != "status"}
+    # Skip reasoning items that carry no summary content.  The Conversations API
+    # rejects these with a 400, and they add no useful context for future turns.
+    if run_item.type == "reasoning_item" and isinstance(input_item, dict):
+        if not input_item.get("summary"):
+            return None
     if (
         _should_omit_reasoning_item_ids(reasoning_item_id_policy)
         and run_item.type == "reasoning_item"
diff --git a/tests/test_run_internal_items.py b/tests/test_run_internal_items.py
index 019a8ededa..9ec9350935 100644
--- a/tests/test_run_internal_items.py
+++ b/tests/test_run_internal_items.py
@@ -8,7 +8,7 @@
     ResponseToolSearchCall,
     ResponseToolSearchOutputItem,
 )
-from openai.types.responses.response_reasoning_item import ResponseReasoningItem
+from openai.types.responses.response_reasoning_item import ResponseReasoningItem, Summary
 
 from agents import Agent
 from agents.exceptions import AgentsException
@@ -440,7 +440,7 @@ def test_run_item_to_input_item_preserves_reasoning_item_ids_by_default() -> Non
         raw_item=ResponseReasoningItem(
             type="reasoning",
             id="rs_123",
-            summary=[],
+            summary=[Summary(type="summary_text", text="thinking")],
         ),
     )
 
@@ -458,7 +458,7 @@ def test_run_item_to_input_item_omits_reasoning_item_ids_when_configured() -> No
         raw_item=ResponseReasoningItem(
             type="reasoning",
             id="rs_456",
-            summary=[],
+            summary=[Summary(type="summary_text", text="thinking")],
         ),
     )
 
@@ -469,6 +469,32 @@ def test_run_item_to_input_item_omits_reasoning_item_ids_when_configured() -> No
     assert "id" not in result
 
 
+def test_run_item_to_input_item_skips_reasoning_item_with_empty_summary() -> None:
+    """Regression test for #3268.
+
+    When a reasoning item has an empty summary list (e.g. encrypted-reasoning turns
+    where the model emits no visible summary text), ``run_item_to_input_item`` must
+    return ``None`` so the empty item is not forwarded to session persistence.  The
+    Conversations API rejects ``{"type": "reasoning", "summary": []}`` with a 400,
+    and the payload carries no useful context for subsequent turns anyway.
+    """
+    agent = Agent(name="A")
+    reasoning = ReasoningItem(
+        agent=agent,
+        raw_item=ResponseReasoningItem(
+            type="reasoning",
+            id="rs_empty",
+            summary=[],
+        ),
+    )
+
+    result = run_items.run_item_to_input_item(reasoning)
+
+    assert result is None, (
+        f"Expected None for a reasoning item with an empty summary list, got {result!r} instead"
+    )
+
+
 def test_run_item_to_input_item_preserves_tool_search_items() -> None:
     agent = Agent(name="A")
     tool_search_call = ToolSearchCallItem(