diff --git a/docs/index.md b/docs/index.md
index d1627c9d..a05a2bce 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -324,9 +324,6 @@ print(Settings().model_dump())
 `env_nested_delimiter` can be configured via the `model_config` as shown above, or via the
 `_env_nested_delimiter` keyword argument on instantiation.
 
-JSON is only parsed in top-level fields, if you need to parse JSON in sub-models, you will need to implement
-validators on those models.
-
 Nested environment variables take precedence over the top-level environment variable JSON
 (e.g. in the example above, `SUB_MODEL__V2` trumps `SUB_MODEL`).
 
diff --git a/pydantic_settings/sources.py b/pydantic_settings/sources.py
index 335ba9dd..c9a32430 100644
--- a/pydantic_settings/sources.py
+++ b/pydantic_settings/sources.py
@@ -11,8 +11,8 @@
 
 from dotenv import dotenv_values
 from pydantic import AliasChoices, AliasPath, BaseModel, Json, TypeAdapter
-from pydantic._internal._typing_extra import origin_is_union
-from pydantic._internal._utils import deep_update, lenient_issubclass
+from pydantic._internal._typing_extra import WithArgsTypes, origin_is_union
+from pydantic._internal._utils import deep_update, is_model_class, lenient_issubclass
 from pydantic.fields import FieldInfo
 from typing_extensions import get_args, get_origin
 
@@ -188,6 +188,8 @@ def _extract_field_info(self, field: FieldInfo, field_name: str) -> list[tuple[s
                         )
             else:  # string validation alias
                 field_info.append((v_alias, self._apply_case_sensitive(v_alias), False))
+        elif origin_is_union(get_origin(field.annotation)) and _union_is_complex(field.annotation, field.metadata):
+            field_info.append((field_name, self._apply_case_sensitive(self.env_prefix + field_name), True))
         else:
             field_info.append((field_name, self._apply_case_sensitive(self.env_prefix + field_name), False))
 
@@ -478,16 +480,13 @@ def prepare_field_value(self, field_name: str, field: FieldInfo, value: Any, val
             # simplest case, field is not complex, we only need to add the value if it was found
             return value
 
-    def _union_is_complex(self, annotation: type[Any] | None, metadata: list[Any]) -> bool:
-        return any(_annotation_is_complex(arg, metadata) for arg in get_args(annotation))
-
     def _field_is_complex(self, field: FieldInfo) -> tuple[bool, bool]:
         """
         Find out if a field is complex, and if so whether JSON errors should be ignored
         """
         if self.field_is_complex(field):
             allow_parse_failure = False
-        elif origin_is_union(get_origin(field.annotation)) and self._union_is_complex(field.annotation, field.metadata):
+        elif origin_is_union(get_origin(field.annotation)) and _union_is_complex(field.annotation, field.metadata):
             allow_parse_failure = True
         else:
             return False, False
@@ -495,7 +494,7 @@ def _field_is_complex(self, field: FieldInfo) -> tuple[bool, bool]:
         return True, allow_parse_failure
 
     @staticmethod
-    def next_field(field: FieldInfo | None, key: str) -> FieldInfo | None:
+    def next_field(field: FieldInfo | Any | None, key: str) -> FieldInfo | None:
         """
         Find the field in a sub model by key(env name)
 
@@ -524,11 +523,17 @@ class Cfg(BaseSettings):
         Returns:
             Field if it finds the next field otherwise `None`.
         """
-        if not field or origin_is_union(get_origin(field.annotation)):
-            # no support for Unions of complex BaseSettings fields
+        if not field:
             return None
-        elif field.annotation and hasattr(field.annotation, 'model_fields') and field.annotation.model_fields.get(key):
-            return field.annotation.model_fields[key]
+
+        annotation = field.annotation if isinstance(field, FieldInfo) else field
+        if origin_is_union(get_origin(annotation)) or isinstance(annotation, WithArgsTypes):
+            for type_ in get_args(annotation):
+                type_has_key = EnvSettingsSource.next_field(type_, key)
+                if type_has_key:
+                    return type_has_key
+        elif is_model_class(annotation) and annotation.model_fields.get(key):
+            return annotation.model_fields[key]
 
         return None
 
@@ -716,3 +721,7 @@ def _annotation_is_complex_inner(annotation: type[Any] | None) -> bool:
     return lenient_issubclass(annotation, (BaseModel, Mapping, Sequence, tuple, set, frozenset, deque)) or is_dataclass(
         annotation
     )
+
+
+def _union_is_complex(annotation: type[Any] | None, metadata: list[Any]) -> bool:
+    return any(_annotation_is_complex(arg, metadata) for arg in get_args(annotation))
diff --git a/tests/test_settings.py b/tests/test_settings.py
index b2371953..a7d6912b 100644
--- a/tests/test_settings.py
+++ b/tests/test_settings.py
@@ -198,6 +198,22 @@ class Cfg(BaseSettings):
     }
 
 
+def test_nested_env_optional_json(env):
+    class Child(BaseModel):
+        num_list: Optional[List[int]] = None
+
+    class Cfg(BaseSettings, env_nested_delimiter='__'):
+        child: Optional[Child] = None
+
+    env.set('CHILD__NUM_LIST', '[1,2,3]')
+    cfg = Cfg()
+    assert cfg.model_dump() == {
+        'child': {
+            'num_list': [1, 2, 3],
+        },
+    }
+
+
 def test_nested_env_delimiter_with_prefix(env):
     class Subsettings(BaseSettings):
         banana: str
@@ -1212,6 +1228,21 @@ class Settings(BaseSettings):
     assert Settings().model_dump() == {'foo': {'a': 'b'}}
 
 
+def test_secrets_nested_optional_json(tmp_path):
+    p = tmp_path / 'foo'
+    p.write_text('{"a": 10}')
+
+    class Foo(BaseModel):
+        a: int
+
+    class Settings(BaseSettings):
+        foo: Optional[Foo] = None
+
+        model_config = SettingsConfigDict(secrets_dir=tmp_path)
+
+    assert Settings().model_dump() == {'foo': {'a': 10}}
+
+
 def test_secrets_path_invalid_json(tmp_path):
     p = tmp_path / 'foo'
     p.write_text('{"a": "b"')