diff --git a/docs/about/releases.md b/docs/about/releases.md
index a5dc07368..3d4e902d4 100644
--- a/docs/about/releases.md
+++ b/docs/about/releases.md
@@ -16,6 +16,8 @@ Adds a `nrefs` accessor for counting virtual chunk references, a `mode` paramete
 ### Bug fixes
 - Fix parsing kerchunk references that use a **structured (record) dtype**, as produced for a FITS `BinTableHDU` (e.g. an SDSS spectrum). Such references round-trip the dtype through JSON as a list of `[name, format]` lists and encode the `fill_value` as base64 raw bytes; `from_kerchunk_refs` now coerces the field specs back to tuples before `np.dtype` and decodes the base64 `fill_value` into a structured scalar, instead of raising `TypeError`.
   By [David Stuebe](https://github.com/emfdavid).
+- Read the bytes codec's `endian` robustly across zarr versions when converting v3 metadata to v2. zarr's `BytesCodec.endian` was an `Endian` enum through 3.2.x and is a plain `str` on zarr's `main` (following zarr's enum-to-string-literal deprecation, zarr-developers/zarr-python#3968); `convert_v3_to_v2_metadata` read `endian.value`, which raises `AttributeError` on the plain string, so writing kerchunk references for big-endian data would break against newer zarr. The endianness is now normalized to a string either way.
+  By [David Stuebe](https://github.com/emfdavid).
 
 - Writing a virtual `DataTree` with `region` or `append_dim` (forwarded to each node) previously always failed with `ContainsGroupError`, because every group was unconditionally created; the existing groups are now opened instead.
   By [Aaron Spring](https://github.com/aaronspring).
diff --git a/virtualizarr/tests/test_writers/test_kerchunk.py b/virtualizarr/tests/test_writers/test_kerchunk.py
index 705853f28..7a6434821 100644
--- a/virtualizarr/tests/test_writers/test_kerchunk.py
+++ b/virtualizarr/tests/test_writers/test_kerchunk.py
@@ -351,6 +351,26 @@ def test_convert_v3_to_v2_metadata(
     assert v2_metadata.attributes == {}
 
 
+def test_codec_endian_normalizes_enum_and_str():
+    # zarr's BytesCodec.endian was an Endian enum through 3.2.x and became a plain str on main
+    # (zarr-developers/zarr-python#3968). _codec_endian must return the string either way, so
+    # the str shape (which released zarr can't construct) is fed via a stand-in codec.
+    from types import SimpleNamespace
+    from typing import cast
+
+    from zarr.abc.codec import ArrayBytesCodec
+    from zarr.codecs import BytesCodec
+
+    from virtualizarr.utils import _codec_endian
+
+    assert _codec_endian(BytesCodec(endian="big")) == "big"
+    assert _codec_endian(BytesCodec(endian="little")) == "little"
+    assert _codec_endian(BytesCodec(endian=None)) is None
+    # zarr-main shape: endian is already a plain string
+    assert _codec_endian(cast(ArrayBytesCodec, SimpleNamespace(endian="big"))) == "big"
+    assert _codec_endian(cast(ArrayBytesCodec, SimpleNamespace(endian=None))) is None
+
+
 def test_warn_if_no_virtual_vars():
     non_virtual_ds = xr.Dataset({"foo": ("x", [10, 20, 30]), "x": ("x", [1, 2, 3])})
     with pytest.warns(UserWarning, match="non-virtual"):
diff --git a/virtualizarr/utils.py b/virtualizarr/utils.py
index 87af9094b..a83a840c2 100644
--- a/virtualizarr/utils.py
+++ b/virtualizarr/utils.py
@@ -96,6 +96,20 @@ def determine_chunk_grid_shape(
     return tuple(ceildiv(length, chunksize) for length, chunksize in zip(shape, chunks))
 
 
+def _codec_endian(codec: ArrayBytesCodec) -> str | None:
+    """The bytes codec's endianness as a plain string, or None if it has none.
+
+    ``BytesCodec.endian`` was a ``zarr.codecs.bytes.Endian`` enum through zarr 3.2.x and became a
+    plain ``str`` on zarr's ``main`` branch, following zarr's enum-to-string-literal deprecation
+    (zarr-developers/zarr-python#3968, #3457). Normalize both so the endianness check works across
+    zarr versions without touching the deprecated ``Endian`` enum.
+    """
+    endian = getattr(codec, "endian", None)
+    if endian is None:
+        return None
+    return endian if isinstance(endian, str) else endian.value
+
+
 def convert_v3_to_v2_metadata(
     v3_metadata: ArrayV3Metadata, fill_value: Any = None
 ) -> ArrayV2Metadata:
@@ -125,10 +139,9 @@ def convert_v3_to_v2_metadata(
     # This logic is based on the (default) Bytes codec's endian property,
     # but other codec pipelines could store endianness elsewhere.
     big_endian = any(
-        isinstance(codec, ArrayBytesCodec)
-        and getattr(codec, "endian", None) is not None
-        and codec.endian.value == "big"  # type: ignore[attr-defined]
+        _codec_endian(codec) == "big"
         for codec in v3_metadata.codecs
+        if isinstance(codec, ArrayBytesCodec)
     )
     if big_endian:
         na_dtype = v3_metadata.data_type.to_native_dtype().newbyteorder(">")