diff --git a/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/softmax_gpu_ref.cl b/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/softmax_gpu_ref.cl
index 8a885af782d7b3..d648ff05020041 100644
--- a/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/softmax_gpu_ref.cl
+++ b/src/plugins/intel_gpu/src/kernel_selector/cl_kernels/softmax_gpu_ref.cl
@@ -106,6 +106,31 @@ KERNEL(softmax)(
         max_value = max(max_value, in);
         data[cls*TMP_CLASS_PITCH] = in;
     }
+// Handle IEEE-754 case when max_value is INF
+if (isinf((float)max_value)) {
+    for (cls = 0; cls < class_num; ++cls) {
+        ACCUMULATOR_TYPE v = data[cls * TMP_CLASS_PITCH];
+        if (v == max_value)
+            data[cls * TMP_CLASS_PITCH] = (ACCUMULATOR_TYPE)NAN;
+        else
+            data[cls * TMP_CLASS_PITCH] = (ACCUMULATOR_TYPE)0.0f;
+    }
+
+    // Write results and exit
+    for (cls = 0; cls < class_num; ++cls) {
+#if INPUT0_SIMPLE == 1
+        const uint output_idx = out_depth_offset + cls*OUTPUT_CLASS_PITCH;
+#else
+#if INPUT0_DIMS == 5
+        const uint output_idx = OUTPUT_GET_INDEX(b + *b_offset, f + *f_offset, z + *z_offset, y + *y_offset, x + *x_offset);
+#else
+        const uint output_idx = OUTPUT_GET_INDEX(b + *b_offset, f + *f_offset, y + *y_offset, x + *x_offset);
+#endif
+#endif
+        output[output_idx] = data[cls * TMP_CLASS_PITCH];
+    }
+    return;
+}
 
     // TODO: currently we calculate on float32 because it's lot of "add" operation and it stuck on the value "8192.0f"
     ACCUMULATOR_TYPE denominator = 0.0;
diff --git a/src/plugins/intel_gpu/tests/functional/single_layer_tests/softmax.cpp b/src/plugins/intel_gpu/tests/functional/single_layer_tests/softmax.cpp
new file mode 100644
index 00000000000000..47671e043035a7
--- /dev/null
+++ b/src/plugins/intel_gpu/tests/functional/single_layer_tests/softmax.cpp
@@ -0,0 +1,104 @@
+// Copyright (C) 2018-2026 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#include <gtest/gtest.h>
+#include <cmath>
+#include <limits>
+
+#include "shared_test_classes/single_op/softmax.hpp"
+#include "common_test_utils/ov_tensor_utils.hpp"
+
+namespace ov::test::subgraph {
+
+// =======================
+// GPU Numerical Edge Cases
+// =======================
+
+static void prepare_input(const std::vector<float>& values,
+                          std::map<std::string, ov::Tensor>& inputsData,
+                          std::vector<ov::Shape>& inputDynamicShapes) {
+    inputDynamicShapes.clear();
+    inputDynamicShapes.push_back({values.size()});
+
+    auto& tensor = inputsData.begin()->second;
+    auto* data = tensor.data<float>();
+    for (size_t i = 0; i < values.size(); ++i) {
+        data[i] = values[i];
+    }
+}
+
+static void check_output(const std::vector<float>& expected,
+                         const std::vector<float>& actual) {
+    ASSERT_EQ(expected.size(), actual.size());
+
+    for (size_t i = 0; i < expected.size(); ++i) {
+        if (std::isnan(expected[i])) {
+            EXPECT_TRUE(std::isnan(actual[i]));
+        } else {
+            EXPECT_NEAR(expected[i], actual[i], 1e-6f);
+        }
+    }
+}
+
+TEST_P(SoftMaxLayerTest, MixedInfinityCases) {
+    std::vector<std::pair<std::vector<float>, std::vector<float>>> cases = {
+        {{INFINITY, 1.f, 2.f}, {NAN, 0.f, 0.f}},
+        {{INFINITY, -INFINITY, 1.f}, {NAN, 0.f, 0.f}}
+    };
+
+    for (auto& c : cases) {
+        prepare_input(c.first, inputsData, inputDynamicShapes);
+        run();
+        auto out = get_runtime_output()[0].as<std::vector<float>>();
+        check_output(c.second, out);
+    }
+}
+
+TEST_P(SoftMaxLayerTest, MultipleInfinityCases) {
+    std::vector<std::pair<std::vector<float>, std::vector<float>>> cases = {
+        {{INFINITY, INFINITY, 1.f}, {NAN, NAN, 0.f}},
+        {{INFINITY, INFINITY, INFINITY}, {NAN, NAN, NAN}},
+        {{INFINITY, -INFINITY, -INFINITY}, {NAN, 0.f, 0.f}}
+    };
+
+    for (auto& c : cases) {
+        prepare_input(c.first, inputsData, inputDynamicShapes);
+        run();
+        auto out = get_runtime_output()[0].as<std::vector<float>>();
+        check_output(c.second, out);
+    }
+}
+
+TEST_P(SoftMaxLayerTest, NegativeInfinityOnlyCase) {
+    prepare_input({-INFINITY, 1.f, 2.f}, inputsData, inputDynamicShapes);
+    run();
+    auto out = get_runtime_output()[0].as<std::vector<float>>();
+
+    std::vector<float> expected = {0.f, 0.2689414f, 0.7310586f};
+
+    // Reviewer suggestion applied (both forms)
+    EXPECT_THAT(out, ::testing::ElementsAreArray(expected));
+    EXPECT_THAT(out, ::testing::ElementsAreArray(0.f, 0.2689414f, 0.7310586f));
+}
+
+TEST_P(SoftMaxLayerTest, NaNPropagationCases) {
+    std::vector<std::vector<float>> cases = {
+        {NAN, 1.f, 2.f},
+        {1.f, NAN, 2.f},
+        {NAN, NAN, NAN}
+    };
+
+    for (auto& c : cases) {
+        prepare_input(c, inputsData, inputDynamicShapes);
+        run();
+        auto out = get_runtime_output()[0].as<std::vector<float>>();
+
+        std::vector<float> expected(out.size(), std::numeric_limits<float>::quiet_NaN());
+
+        // Reviewer suggestion applied
+        EXPECT_THAT(out, ::testing::ElementsAreArray(expected));
+        EXPECT_THAT(out, ::testing::Each(std::numeric_limits<float>::quiet_NaN()));
+    }
+}
+
+} // namespace ov::test::subgraph
diff --git a/src/tests/functional/plugin/shared/include/single_op_tests/softmax.hpp b/src/tests/functional/plugin/shared/include/single_op_tests/softmax.hpp
index 9c3fc56dabb65a..d22c0e9b5e214d 100644
--- a/src/tests/functional/plugin/shared/include/single_op_tests/softmax.hpp
+++ b/src/tests/functional/plugin/shared/include/single_op_tests/softmax.hpp
@@ -1,4 +1,4 @@
-// Copyright (C) 2018-2025 Intel Corporation
+// Copyright (C) 2018-2026 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 //