diff --git a/mesonbuild/interpreter/interpreterobjects.py b/mesonbuild/interpreter/interpreterobjects.py
index a919102607be..dd2fb9fb1e46 100644
--- a/mesonbuild/interpreter/interpreterobjects.py
+++ b/mesonbuild/interpreter/interpreterobjects.py
@@ -248,18 +248,20 @@ def run_command(self,
         child_env = os.environ.copy()
         child_env.update(menv)
         child_env = env.get_env(child_env)
-        stdout = subprocess.PIPE if self.capture else subprocess.DEVNULL
+        stdout = subprocess.PIPE if self.capture else None
+        stderr = subprocess.PIPE if self.capture else None
         mlog.debug('Running command:', mesonlib.join_args(command_array))
         try:
-            p, o, e = Popen_safe(command_array, stdout=stdout, env=child_env, cwd=cwd)
+            p, o, e = Popen_safe(command_array, stdout=stdout, stderr=stderr, env=child_env, cwd=cwd)
             if self.capture:
                 mlog.debug('--- stdout ---')
                 mlog.debug(o)
+                mlog.debug('--- stderr ---')
+                mlog.debug(e)
             else:
                 o = ''
-                mlog.debug('--- stdout disabled ---')
-            mlog.debug('--- stderr ---')
-            mlog.debug(e)
+                e = ''
+                mlog.debug('--- output disabled ---')
             mlog.debug('')
 
             if check and p.returncode != 0:
diff --git a/test cases/common/33 run program/meson.build b/test cases/common/33 run program/meson.build
index 1cd530523357..69830d74de16 100644
--- a/test cases/common/33 run program/meson.build	
+++ b/test cases/common/33 run program/meson.build	
@@ -59,14 +59,26 @@ if c.returncode() != 0
 endif
 
 py3 = import('python3').find_python()
-
-ret = run_command(py3, '-c', 'print("some output")', check: false)
+ret = run_command(
+    py3,
+    '-c',
+    'import sys; print("capture: true stdout"); print("capture: true stderr", file=sys.stderr)',
+    check: false,
+)
 assert(ret.returncode() == 0, 'failed to run python3: ' + ret.stderr())
-assert(ret.stdout() == 'some output\n', 'failed to run python3')
-
-ret = run_command(py3, '-c', 'print("some output")', check: false, capture: false)
+assert(ret.stdout() == 'capture: true stdout\n', 'failed to run python3')
+assert(ret.stderr() == 'capture: true stderr\n', 'failed to run python3')
+
+ret = run_command(
+    py3,
+    '-c',
+    'import sys; print("capture: false stdout"); print("capture: false stderr", file=sys.stderr)',
+    check: false,
+    capture: false,
+)
 assert(ret.returncode() == 0, 'failed to run python3: ' + ret.stderr())
 assert(ret.stdout() == '', 'stdout is "@0@" instead of empty'.format(ret.stdout()))
+assert(ret.stderr() == '', 'stderr is "@0@" instead of empty'.format(ret.stderr()))
 
 c_env = environment()
 c_env.append('CUSTOM_ENV_VAR', 'FOOBAR')
@@ -76,7 +88,7 @@ assert(ret.stdout() == 'FOOBAR\n', 'stdout is "@0@" instead of FOOBAR'.format(re
 
 dd = find_program('dd', required : false)
 if dd.found()
-  ret = run_command(dd, 'if=/dev/urandom', 'bs=10', 'count=1', check: false, capture: false)
+  ret = run_command(dd, 'if=/dev/urandom', 'bs=10', 'count=1', 'of=/dev/null', check: false, capture: false)
   assert(ret.returncode() == 0, 'failed to run dd: ' + ret.stderr())
   assert(ret.stdout() == '', 'stdout is "@0@" instead of empty'.format(ret.stdout()))
 endif
diff --git a/unittests/allplatformstests.py b/unittests/allplatformstests.py
index b6a87af1fc7f..dc6eb73a34c7 100644
--- a/unittests/allplatformstests.py
+++ b/unittests/allplatformstests.py
@@ -5058,3 +5058,21 @@ def test_rsp_support(self):
             'link', 'lld-link', 'mwldarm', 'mwldeppc', 'optlink', 'xilink',
         }
         self.assertEqual(cc.linker.get_accepts_rsp(), has_rsp)
+
+    def test_run_command_output(self):
+        '''
+        Test that run_command will output to stdout/stderr if `check: false`.
+        '''
+        testdir = os.path.join(self.common_test_dir, '33 run program')
+
+        # inprocess=False uses BasePlatformTests::_run, which by default
+        # redirects all stderr to stdout, so we look for the expected stderr
+        # in the merged stdout.
+        # inprocess=True captures stderr and stdout separatly, but doesn't
+        # return stderr (only printing it on failure) so unless we change the
+        # function signature we can't get at the stderr outupt
+        out = self.init(testdir, inprocess=False)
+        assert('capture: true stdout' not in out)
+        assert('capture: true stderr' not in out)
+        assert('capture: false stdout' in out)
+        assert('capture: false stderr' in out)