diff --git a/.github/workflows/test.yaml b/.github/workflows/test.yaml
index 8e1e7a4..907cd3d 100644
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@@ -9,8 +9,19 @@ jobs:
 
     strategy:
       matrix:
-        os: [ubuntu-latest, macos-latest, windows-latest]
+        os: [ubuntu-latest, windows-latest, macos-latest]
         python-version: ["3.8", "3.9", "3.10", "3.11", "3.12"]
+        numpy-version: ["1.24.2", "1.26.4", "2.0.0"]
+        exclude:
+          - python-version: "3.8"
+            numpy-version: "2.0.0"
+          - python-version: "3.8"
+            numpy-version: "1.26.4"
+          - python-version: "3.12"
+            numpy-version: "1.24.2"
+          - os: windows-latest
+            numpy-version: "2.0.0"
+
 
     steps:
       - uses: actions/checkout@v3
@@ -21,6 +32,7 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
+          pip install numpy==${{ matrix.numpy-version }}
           pip install -r requirements.txt
       - name: Test with pytest
         run: |
diff --git a/README.md b/README.md
index 9e9e8b1..b8dcd43 100644
--- a/README.md
+++ b/README.md
@@ -32,6 +32,8 @@ def f(large_argument, not_json_argument, transient_flag):
 
 The dictionary has keys that correspond to each of the arguments, and the values are applied to them before placing them in the key. Here, `stable_hash` can be used to hash the json stringification of the value, saving disk space but making recovering the value impossible if you want to do that later. Additionally, `str` can be used to stringify objects that you are convinced have stable `str` representations but cannot be represented in json. Finally, the flag argument is ignored in the JSON representation, this is useful for verbosity flags, etc., that don't affect the output.
 
+**Extremely important note**: If you use `stable_hash`, it is only guaranteed to be stable for the same major version of `numpy` (numpy `1.*.*` vs `2.*.*`) and the same operating system. On Ubuntu it does appear to be stable across versions of `numpy`, but this is not true on Mac OS.
+
 ## Aliasing
 
 Permacache uses the underlying function signature to construct the key. For example, for the function
diff --git a/tests/stringify_test.py b/tests/stringify_test.py
index 273a5d0..3d90371 100644
--- a/tests/stringify_test.py
+++ b/tests/stringify_test.py
@@ -8,18 +8,26 @@
 
 from permacache import stable_hash, stringify
 
+NUMPY_VERSION = np.version.version.split(".", maxsplit=1)[0]
+
 
 class StringifyTest(unittest.TestCase):
     data = np.random.RandomState(0).randn(1000)
     slow_hash = {
         "linux": "16469cff96525e3e190758d793e61f9d798cb87617787bc1312cf7a8b59aa4b2",
         "win32": "faaa45dfa377870db4d547ca91187566b6b74983adaf16241a80ef96af592285",
-        "darwin": "97d8947421010821dd8f3f7046ef0eccd5c27602c7505a5749c1090a0fe7435b",
+        "darwin": {
+            "1": "97d8947421010821dd8f3f7046ef0eccd5c27602c7505a5749c1090a0fe7435b",
+            "2": "f80b352d29bedc2147f6d5a3623226c7f350bbab1f743ffe0a257132c5a60597",
+        }[NUMPY_VERSION],
     }[sys.platform]
     fast_hash = {
         "linux": "c97fda7d817a32aad65ce77f5043a51410c5893e6bab8e746a23f68c8e483774",
         "win32": "387b1ef12b70d4df0984846b753fbcceb1b7d079256e75479810b6a15d668b70",
-        "darwin": "76ce19673caf9532d68730c7f83bfe8e6f521cc001c2d43373f9c4f33c925037",
+        "darwin": {
+            "1": "76ce19673caf9532d68730c7f83bfe8e6f521cc001c2d43373f9c4f33c925037",
+            "2": "a0128e05b5d5ceab4d1b5716c5ce3339b25f3ef6a1208383e6c99afdc2b519c1",
+        }[NUMPY_VERSION],
     }[sys.platform]
 
     def test_stringify_json(self):
@@ -39,6 +47,7 @@ class X:
         )
 
     def test_stringify_numpy(self):
+        print(self.data.tobytes())
         self.assertEqual(
             self.slow_hash,
             stable_hash(self.data, fast_bytes=False),