diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PrepareModelConfig.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PrepareModelConfig.aidl
index 85c924f..1f44ba7 100644
--- a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PrepareModelConfig.aidl
+++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PrepareModelConfig.aidl
@@ -39,7 +39,8 @@
   long deadlineNs;
   ParcelFileDescriptor[] modelCache;
   ParcelFileDescriptor[] dataCache;
-  byte[] cacheToken;
+  byte[32] cacheToken;
   android.hardware.neuralnetworks.TokenValuePair[] compilationHints;
   android.hardware.neuralnetworks.ExtensionNameAndPrefix[] extensionNameToPrefix;
+  const int BYTE_SIZE_OF_CACHE_TOKEN = 32;
 }
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl
index 821b9fe..7808fc2 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl
@@ -39,7 +39,7 @@
     /**
      * The byte size of the cache token.
      */
-    const int BYTE_SIZE_OF_CACHE_TOKEN = 32;
+    const int BYTE_SIZE_OF_CACHE_TOKEN = PrepareModelConfig.BYTE_SIZE_OF_CACHE_TOKEN;
     /**
      * The maximum number of files for each type of cache in compilation caching.
      */
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl
index 949804e..f752750 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl
@@ -204,6 +204,12 @@
      * appropriate ErrorStatus value. If the inputs to the function are valid and there is no error,
      * createReusableExecution must construct a reusable execution.
      *
+     * This method will be called when a client requests a reusable execution with consistent
+     * request and execution config. For single-time execution,
+     * {@link IPreparedModel::executeSynchronouslyWithConfig} or
+     * {@link IPreparedModel::executeFencedWithConfig} is preferred, because the overhead of
+     * setting up a reusable execution can be avoided.
+     *
      * @param request The input and output information on which the prepared model is to be
      *                executed.
      * @param config Specifies the execution configuration parameters.
@@ -223,6 +229,10 @@
      * ExecutionConfig} instead of a list of configuration parameters, and ExecutionConfig contains
      * more configuration parameters than are passed to executeSynchronously.
      *
+     * This method is preferred when a client requests a single-time synchronous execution.
+     * For reusable execution with consistent request and execution config,
+     * {@link IPreparedModel::createReusableExecution} must be called.
+     *
      * @param request The input and output information on which the prepared model is to be
      *                executed.
      * @param config Specifies the execution configuration parameters.
@@ -246,6 +256,10 @@
      * ExecutionConfig} instead of a list of configuration parameters, and ExecutionConfig contains
      * more configuration parameters than are passed to executeFenced.
      *
+     * This method is preferred when a client requests a single-time fenced execution.
+     * For reusable execution with consistent request and execution config,
+     * {@link IPreparedModel::createReusableExecution} must be called.
+     *
      * @param request The input and output information on which the prepared model is to be
      *                executed. The outputs in the request must have fully specified dimensions.
      * @param waitFor A vector of sync fence file descriptors. Execution must not start until all
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/PrepareModelConfig.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/PrepareModelConfig.aidl
index 96df968..55bd291 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/PrepareModelConfig.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/PrepareModelConfig.aidl
@@ -28,6 +28,11 @@
 @VintfStability
 parcelable PrepareModelConfig {
     /**
+     * The byte size of the cache token.
+     */
+    const int BYTE_SIZE_OF_CACHE_TOKEN = 32;
+
+    /**
      * Indicates the intended execution behavior of a prepared model.
      */
     ExecutionPreference preference;
@@ -66,7 +71,7 @@
      */
     ParcelFileDescriptor[] dataCache;
     /**
-     * A caching token of length IDevice::BYTE_SIZE_OF_CACHE_TOKEN identifying
+     * A caching token of length BYTE_SIZE_OF_CACHE_TOKEN identifying
      * the prepared model. The same token will be provided when
      * retrieving the prepared model from the cache files with
      * IDevice::prepareModelFromCache.  Tokens should be chosen to have a low
@@ -77,7 +82,7 @@
      * indicating that caching information is not provided, this
      * token must be ignored.
      */
-    byte[] cacheToken;
+    byte[BYTE_SIZE_OF_CACHE_TOKEN] cacheToken;
     /**
      * A vector of token / value pairs represent vendor specific
      * compilation hints or metadata. The provided TokenValuePairs must not
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
index af58715..71a28ef 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
@@ -27,6 +27,7 @@
 #include <aidl/android/hardware/neuralnetworks/Extension.h>
 #include <aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.h>
 #include <aidl/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.h>
+#include <aidl/android/hardware/neuralnetworks/IDevice.h>
 #include <aidl/android/hardware/neuralnetworks/Memory.h>
 #include <aidl/android/hardware/neuralnetworks/Model.h>
 #include <aidl/android/hardware/neuralnetworks/Operand.h>
@@ -219,6 +220,7 @@
 #endif  // NN_AIDL_V4_OR_ABOVE
 
 nn::GeneralResult<std::vector<int32_t>> toSigned(const std::vector<uint32_t>& vec);
+std::vector<uint8_t> toVec(const std::array<uint8_t, IDevice::BYTE_SIZE_OF_CACHE_TOKEN>& token);
 
 }  // namespace aidl::android::hardware::neuralnetworks::utils
 
diff --git a/neuralnetworks/aidl/utils/src/Conversions.cpp b/neuralnetworks/aidl/utils/src/Conversions.cpp
index 9b897c4..83fda10 100644
--- a/neuralnetworks/aidl/utils/src/Conversions.cpp
+++ b/neuralnetworks/aidl/utils/src/Conversions.cpp
@@ -614,7 +614,7 @@
     using Ts::operator()...;
 };
 template <class... Ts>
-overloaded(Ts...)->overloaded<Ts...>;
+overloaded(Ts...) -> overloaded<Ts...>;
 
 #ifdef __ANDROID__
 nn::GeneralResult<common::NativeHandle> aidlHandleFromNativeHandle(
@@ -1190,4 +1190,8 @@
     return std::vector<int32_t>(vec.begin(), vec.end());
 }
 
+std::vector<uint8_t> toVec(const std::array<uint8_t, IDevice::BYTE_SIZE_OF_CACHE_TOKEN>& token) {
+    return std::vector<uint8_t>(token.begin(), token.end());
+}
+
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/src/Device.cpp b/neuralnetworks/aidl/utils/src/Device.cpp
index f3f4fdb..b64a40d 100644
--- a/neuralnetworks/aidl/utils/src/Device.cpp
+++ b/neuralnetworks/aidl/utils/src/Device.cpp
@@ -229,7 +229,6 @@
     const auto aidlDeadline = NN_TRY(convert(deadline));
     auto aidlModelCache = NN_TRY(convert(modelCache));
     auto aidlDataCache = NN_TRY(convert(dataCache));
-    const auto aidlToken = NN_TRY(convert(token));
 
     const auto cb = ndk::SharedRefBase::make<PreparedModelCallback>(kFeatureLevel);
     const auto scoped = kDeathHandler.protectCallback(cb.get());
@@ -240,12 +239,13 @@
         const auto ret = kDevice->prepareModelWithConfig(
                 aidlModel,
                 {aidlPreference, aidlPriority, aidlDeadline, std::move(aidlModelCache),
-                 std::move(aidlDataCache), aidlToken, std::move(aidlHints),
+                 std::move(aidlDataCache), token, std::move(aidlHints),
                  std::move(aidlExtensionPrefix)},
                 cb);
         HANDLE_ASTATUS(ret) << "prepareModel failed";
         return cb->get();
     }
+    const auto aidlToken = NN_TRY(convert(token));
     const auto ret = kDevice->prepareModel(aidlModel, aidlPreference, aidlPriority, aidlDeadline,
                                            aidlModelCache, aidlDataCache, aidlToken, cb);
     HANDLE_ASTATUS(ret) << "prepareModel failed";
diff --git a/neuralnetworks/aidl/utils/src/InvalidDevice.cpp b/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
index 33270ff..44f8ea9 100644
--- a/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
+++ b/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
@@ -189,7 +189,8 @@
         }
     }
     return prepareModel(model, config.preference, config.priority, config.deadlineNs,
-                        config.modelCache, config.dataCache, config.cacheToken, callback);
+                        config.modelCache, config.dataCache, utils::toVec(config.cacheToken),
+                        callback);
 }
 
 ndk::ScopedAStatus InvalidDevice::prepareModelFromCache(
diff --git a/neuralnetworks/aidl/vts/functional/Utils.h b/neuralnetworks/aidl/vts/functional/Utils.h
index 4e0a4aa..ccb0778 100644
--- a/neuralnetworks/aidl/vts/functional/Utils.h
+++ b/neuralnetworks/aidl/vts/functional/Utils.h
@@ -21,6 +21,7 @@
 #include <gtest/gtest.h>
 
 #include <algorithm>
+#include <array>
 #include <iosfwd>
 #include <string>
 #include <utility>
@@ -47,6 +48,7 @@
 inline constexpr int64_t kOmittedTimeoutDuration = -1;
 inline constexpr int64_t kNoDuration = -1;
 inline const std::vector<uint8_t> kEmptyCacheToken(IDevice::BYTE_SIZE_OF_CACHE_TOKEN);
+inline const std::array<uint8_t, IDevice::BYTE_SIZE_OF_CACHE_TOKEN> kEmptyCacheTokenArray{};
 
 // Returns the amount of space needed to store a value of the specified type.
 //
diff --git a/neuralnetworks/aidl/vts/functional/ValidateModel.cpp b/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
index 931ba25..060434e 100644
--- a/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
+++ b/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
@@ -85,7 +85,7 @@
     std::shared_ptr<PreparedModelCallback> preparedModelCallback =
             ndk::SharedRefBase::make<PreparedModelCallback>();
     const auto prepareLaunchStatus = device->prepareModelWithConfig(
-            model, {preference, priority, kNoDeadline, {}, {}, kEmptyCacheToken, {}, {}},
+            model, {preference, priority, kNoDeadline, {}, {}, kEmptyCacheTokenArray, {}, {}},
             preparedModelCallback);
     ASSERT_FALSE(prepareLaunchStatus.isOk());
     ASSERT_EQ(prepareLaunchStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
diff --git a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
index 51b4805..bf87f15 100644
--- a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
+++ b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
@@ -72,7 +72,7 @@
                                                 kNoDeadline,
                                                 {},
                                                 {},
-                                                kEmptyCacheToken,
+                                                kEmptyCacheTokenArray,
                                                 {},
                                                 {}},
                                                preparedModelCallback);