diff --git a/src/core/include/mp-units/safe_int.h b/src/core/include/mp-units/safe_int.h
index 11b215c77..0f9668de7 100644
--- a/src/core/include/mp-units/safe_int.h
+++ b/src/core/include/mp-units/safe_int.h
@@ -140,7 +140,14 @@ template<integral T>
 template<integral T>
 [[nodiscard]] constexpr bool mul_overflows(T lhs, T rhs) noexcept
 {
-  if constexpr (integer_rep_width_v<T> < max_native_width) {
+  // Use a wider type for the overflow check whenever one exists.  On platforms without native
+  // `__int128`, `int128_t` is the synthetic `double_width_int<int64_t>` — still a usable
+  // 128-bit type — so `max_native_width` (which only counts builtin widths) is the wrong
+  // ceiling; check directly against `integer_rep_width_v<int128_t>` instead.  The else branch
+  // below assumes T is the widest available signed integer (i.e. (u)int128_t) and computes
+  // `max` by truncating uint128_t — that truncation is what produced the bug on MSVC for
+  // T == long long.
+  if constexpr (integer_rep_width_v<T> < integer_rep_width_v<int128_t>) {
     using wide = double_width_int_for_t<T>;
     const wide product = static_cast<wide>(lhs) * static_cast<wide>(rhs);
     return product > static_cast<wide>(std::numeric_limits<T>::max()) ||
@@ -336,6 +343,65 @@ struct safe_int_binary_ops {
     if (std::cmp_greater(lhs.value_, rhs.value_)) return std::strong_ordering::greater;
     return std::strong_ordering::equal;
   }
+
+  // Heterogeneous arithmetic: safe_int<T,EP> op safe_int<U,EP>, T != U, matching sign.
+  // Same idiom as the comparison operators above — ADL finds these via the base class,
+  // the homogeneous (T == U) inline-friend operators inside `safe_int<T>` win for same-T
+  // calls because non-template beats template in overload resolution, and the
+  // heterogeneous calls have no competition.  Resolves the MSVC C2666 ambiguity on
+  // `safe_int<int> + safe_int<long>` when `int` and `long` have equal width (Windows).
+  template<integral T, integral U, typename EP>
+    requires(!std::same_as<T, U> && same_sign_v<T, U>)
+  [[nodiscard]] friend constexpr auto operator+(safe_int<T, EP> lhs, safe_int<U, EP> rhs)
+    -> safe_int<integral_op_result_t<T, U>, EP>
+  {
+    using R = integral_op_result_t<T, U>;
+    if (add_overflows<R>(static_cast<R>(lhs.value_), static_cast<R>(rhs.value_)))
+      EP::on_overflow("safe_int: addition overflow");
+    return lhs.value_ + rhs.value_;
+  }
+
+  template<integral T, integral U, typename EP>
+    requires(!std::same_as<T, U> && same_sign_v<T, U>)
+  [[nodiscard]] friend constexpr auto operator-(safe_int<T, EP> lhs, safe_int<U, EP> rhs)
+    -> safe_int<integral_op_result_t<T, U>, EP>
+  {
+    using R = integral_op_result_t<T, U>;
+    if (sub_overflows<R>(static_cast<R>(lhs.value_), static_cast<R>(rhs.value_)))
+      EP::on_overflow("safe_int: subtraction overflow");
+    return lhs.value_ - rhs.value_;
+  }
+
+  template<integral T, integral U, typename EP>
+    requires(!std::same_as<T, U> && same_sign_v<T, U>)
+  [[nodiscard]] friend constexpr auto operator*(safe_int<T, EP> lhs, safe_int<U, EP> rhs)
+    -> safe_int<integral_op_result_t<T, U>, EP>
+  {
+    using R = integral_op_result_t<T, U>;
+    if (mul_overflows<R>(static_cast<R>(lhs.value_), static_cast<R>(rhs.value_)))
+      EP::on_overflow("safe_int: multiplication overflow");
+    return lhs.value_ * rhs.value_;
+  }
+
+  template<integral T, integral U, typename EP>
+    requires(!std::same_as<T, U> && same_sign_v<T, U>)
+  [[nodiscard]] friend constexpr auto operator/(safe_int<T, EP> lhs, safe_int<U, EP> rhs)
+    -> safe_int<integral_op_result_t<T, U>, EP>
+  {
+    using R = integral_op_result_t<T, U>;
+    if (div_overflows<R>(static_cast<R>(lhs.value_), static_cast<R>(rhs.value_)))
+      EP::on_overflow("safe_int: division overflow");
+    return lhs.value_ / rhs.value_;
+  }
+
+  template<integral T, integral U, typename EP>
+    requires(!std::same_as<T, U> && same_sign_v<T, U>)
+  [[nodiscard]] friend constexpr auto operator%(safe_int<T, EP> lhs, safe_int<U, EP> rhs)
+    -> safe_int<integral_op_result_t<T, U>, EP>
+  {
+    if (rhs.value_ == U{0}) EP::on_overflow("safe_int: modulo by zero");
+    return lhs.value_ % rhs.value_;
+  }
 };
 
 }  // namespace detail
diff --git a/test/static/safe_int_test.cpp b/test/static/safe_int_test.cpp
index 9c206ae24..e684450d2 100644
--- a/test/static/safe_int_test.cpp
+++ b/test/static/safe_int_test.cpp
@@ -617,6 +617,41 @@ static_assert(safe_int<int>{6} * 2LL == safe_int<long long>{12});
 static_assert(2LL * safe_int<int>{6} == safe_int<long long>{12});
 static_assert(safe_int<int>{12} / 4LL == safe_int<long long>{3});
 
+// ============================================================================
+// Cross-type integral — additional convertibility coverage
+//
+// Exercises the `safe_int<T> op integral_U` overloads for combinations not covered above:
+//   - narrower raw RHS (promotes to T's rank, stays safe_int<T>)
+//   - `long` (whose width differs across platforms — 32-bit on Windows, 64-bit on POSIX),
+//     verifying that the safe_int op integral overload is chosen over implicit conversion
+//     of the raw integer to safe_int<T> via the converting ctor.
+// ============================================================================
+
+// Narrowing raw RHS: result keeps the wider wrapper type.
+// short → int by C++ promotion rules, so safe_int<int> + short → safe_int<int>.
+static_assert(std::is_same_v<decltype(safe_int<int>{} + short{1}), safe_int<int>>);
+static_assert(std::is_same_v<decltype(short{1} + safe_int<int>{}), safe_int<int>>);
+static_assert(safe_int<int>{3} + short{4} == safe_int<int>{7});
+static_assert(short{4} + safe_int<int>{3} == safe_int<int>{7});
+
+// safe_int<short> + int → safe_int<int> (raw RHS has higher rank).
+static_assert(std::is_same_v<decltype(safe_int<short>{} + 1), safe_int<int>>);
+static_assert(std::is_same_v<decltype(1 + safe_int<short>{}), safe_int<int>>);
+static_assert(safe_int<short>{3} + 4 == safe_int<int>{7});
+
+// safe_int<int> + 1L: result is safe_int<long>.  On Windows where `long` and `int` have
+// equal width this would otherwise be tempted by the converting ctor (raw long → safe_int<int>),
+// which would yield safe_int<int>; the safe_int op integral overload wins by exact match.
+static_assert(std::is_same_v<decltype(safe_int<int>{} + 1L), safe_int<long>>);
+static_assert(std::is_same_v<decltype(1L + safe_int<int>{}), safe_int<long>>);
+static_assert(safe_int<int>{3} + 4L == safe_int<long>{7L});
+static_assert(4L + safe_int<int>{3} == safe_int<long>{7L});
+
+// Unsigned widening: safe_int<unsigned> + uint8_t → safe_int<unsigned>.
+static_assert(std::is_same_v<decltype(safe_int<unsigned>{} + std::uint8_t{1}), safe_int<unsigned>>);
+static_assert(std::is_same_v<decltype(std::uint8_t{1} + safe_int<unsigned>{}), safe_int<unsigned>>);
+static_assert(safe_int<unsigned>{3u} + std::uint8_t{4} == safe_int<unsigned>{7u});
+
 // int × unsigned scalar arithmetic is intentionally ill-formed — same rationale as
 // safe_int<int> + safe_int<unsigned>: sign-mismatch conversions produce counterintuitive
 // results (e.g. safe_int<int>{-1} * 2u → UINT_MAX-1 via reinterpretation).
@@ -763,9 +798,12 @@ static_assert(safe_int<int>{3} * safe_int<long>{4L} == safe_int<long>{12L});
 static_assert(safe_int<int>{12} / safe_int<long>{4L} == safe_int<long>{3L});
 static_assert(safe_int<int>{10} % safe_int<long>{3L} == safe_int<long>{1L});
 
-// Widening prevents overflow: INT_MAX fits in long, so no overflow here
-static_assert(safe_int<int>{std::numeric_limits<int>::max()} + safe_int<long>{1L} ==
-              safe_int<long>{static_cast<long>(std::numeric_limits<int>::max()) + 1L});
+// Widening prevents overflow: INT_MAX + 1 fits in long long on every platform.
+// (`long long` rather than `long`: on Windows `long` is 32-bit and equal to `int`, so the
+// RHS `static_cast<long>(INT_MAX) + 1L` would itself be 32-bit signed overflow — UB at
+// constexpr — and the assertion would not be portable.)
+static_assert(safe_int<int>{std::numeric_limits<int>::max()} + safe_int<long long>{1LL} ==
+              safe_int<long long>{static_cast<long long>(std::numeric_limits<int>::max()) + 1LL});
 
 // Heterogeneous comparison (operator== and operator<=>)
 static_assert(safe_int<int>{3} == safe_int<long>{3L});