diff --git a/src/cpp/src/tokenizer.cpp b/src/cpp/src/tokenizer.cpp
index a1a08710a4..5a97982411 100644
--- a/src/cpp/src/tokenizer.cpp
+++ b/src/cpp/src/tokenizer.cpp
@@ -176,10 +176,16 @@ class Tokenizer::TokenizerImpl {
 
     void setupTokenizer(const std::pair<std::shared_ptr<ov::Model>, std::shared_ptr<ov::Model>>& models,  const ov::AnyMap& properties) {
         auto [ov_tokenizer, ov_detokenizer] = models;
+        OPENVINO_ASSERT(ov_tokenizer || ov_detokenizer, "Neither tokenizer nor detokenzier models were provided");
 
-        m_older_than_24_5 = ov_tokenizer->get_rt_info().count("openvino_tokenizers_version") != 1;
         auto core = get_core_singleton();
         std::string device = "CPU"; // only CPU is supported for now
+        
+        std::string version_str;
+        utils::read_rt_info(ov_tokenizer != nullptr ? ov_tokenizer: ov_detokenizer , "openvino_tokenizers_version", version_str);
+        // Saving IR version was added only in 24.5, so if it's empty, then it's older than 24.5
+        m_older_than_24_5 = version_str.empty();
+
         if (ov_tokenizer) {
             ov::pass::Manager manager;
             manager.register_pass<MakeCombineSegmentsSatateful>();
@@ -207,7 +213,8 @@ class Tokenizer::TokenizerImpl {
         if (m_tokenizer) {
             // TODO CVS-150630: Empty strings sporadically can fail, therefore use nonempty string for warmup.
             encode("non empty string").input_ids;
-        if (m_detokenizer)
+        }
+        if (m_detokenizer) {
             decode({1, 33, 199, 42, 42});
         }
 
@@ -354,6 +361,9 @@ class Tokenizer::TokenizerImpl {
     }
 
     TokenizedInputs encode(std::string prompt, const ov::AnyMap& tokenization_params = {}) {
+        OPENVINO_ASSERT(m_ireq_queue_tokenizer, "Either openvino_tokenizer.xml was not provided or it was not loaded correctly. "
+                                                "Tokenizer::encode is not available");
+
         bool add_special_tokens_flag = true;
         ov::genai::utils::read_anymap_param(tokenization_params, add_special_tokens.name(), add_special_tokens_flag);
 
@@ -370,6 +380,8 @@ class Tokenizer::TokenizerImpl {
     }
 
     TokenizedInputs encode(std::vector<std::string>& prompts, const ov::AnyMap& tokenization_params = {}) {
+        OPENVINO_ASSERT(m_ireq_queue_tokenizer, "Either openvino_tokenizer.xml was not provided or it was not loaded correctly. "
+                                                "Tokenizer::encode is not available");
         TokenizedInputs unpadded;
         {
             bool add_special_tokens_flag = true;
diff --git a/tests/python_tests/test_generate_api.py b/tests/python_tests/test_generate_api.py
index 715f9b97a3..859e2a7586 100644
--- a/tests/python_tests/test_generate_api.py
+++ b/tests/python_tests/test_generate_api.py
@@ -779,7 +779,8 @@ def run_perf_metrics_collection(model_descr, generation_config: Dict, prompt: st
 def test_perf_metrics(model_descr, generation_config, prompt):
     import time
     start_time = time.perf_counter()
-    perf_metrics = run_perf_metrics_collection(read_model(model_descr), generation_config, prompt)
+    # To check prefill exclusion we need long initial prompt.
+    perf_metrics = run_perf_metrics_collection(read_model(model_descr), generation_config, prompt * 1000)
     total_time = (time.perf_counter() - start_time) * 1000
     
     # Check that load time is adequate.