diff --git a/esphome/core/scheduler.cpp b/esphome/core/scheduler.cpp
index 3ff8fa4475..c783a788cc 100644
--- a/esphome/core/scheduler.cpp
+++ b/esphome/core/scheduler.cpp
@@ -545,7 +545,7 @@ void HOT Scheduler::call(uint32_t now) {
     // during the function call and know if we were cancelled.
     auto executed_item = this->pop_raw_locked_();
 
-    if (executed_item->remove) {
+    if (this->is_item_removed_locked_(executed_item.get())) {
       // We were removed/cancelled in the function call, recycle and continue
       this->to_remove_--;
       this->recycle_item_main_loop_(std::move(executed_item));
@@ -605,7 +605,7 @@ size_t HOT Scheduler::cleanup_() {
   LockGuard guard{this->lock_};
   while (!this->items_.empty()) {
     auto &item = this->items_[0];
-    if (!item->remove)
+    if (!this->is_item_removed_locked_(item.get()))
       break;
     this->to_remove_--;
     this->recycle_item_main_loop_(this->pop_raw_locked_());
diff --git a/esphome/core/scheduler.h b/esphome/core/scheduler.h
index cf9a2a3660..020d822a0a 100644
--- a/esphome/core/scheduler.h
+++ b/esphome/core/scheduler.h
@@ -470,8 +470,12 @@ class Scheduler {
   // IMPORTANT: Caller must hold the scheduler lock before calling this function.
   bool is_item_removed_locked_(SchedulerItem *item) const {
 #ifdef ESPHOME_THREAD_MULTI_ATOMICS
-    // Lock already held - relaxed is sufficient, mutex provides ordering
-    return item->remove.load(std::memory_order_relaxed);
+    // Lock already held - relaxed is sufficient, mutex provides ordering.
+    // Use GCC __atomic_load_n builtin instead of std::atomic::load() because
+    // GCC for Xtensa emits std::atomic<bool>::load() as an out-of-line
+    // libstdc++ call, adding function call overhead that exceeds the memw
+    // barrier savings this optimization aims to eliminate.
+    return __atomic_load_n(reinterpret_cast<const bool *>(&item->remove), __ATOMIC_RELAXED);
 #else
     return item->remove;
 #endif