diff --git a/src/builtin_test.cpp b/src/builtin_test.cpp
index cac1167b2..7284034a3 100644
--- a/src/builtin_test.cpp
+++ b/src/builtin_test.cpp
@@ -648,7 +648,6 @@ static bool parse_number(const wcstring &arg, number_t *number, wcstring_list_t
     const wchar_t *argcs = arg.c_str();
     double floating = 0;
     bool got_float = parse_double(argcs, &floating);
-
     errno = 0;
     long long integral = fish_wcstoll(argcs);
     bool got_int = (errno == 0);
@@ -656,6 +655,7 @@ static bool parse_number(const wcstring &arg, number_t *number, wcstring_list_t
         // Here the value is just an integer; ignore the floating point parse because it may be
         // invalid (e.g. not a representable integer).
         *number = number_t{integral, 0.0};
+
         return true;
     } else if (got_float && errno != ERANGE) {
         // Here we parsed an (in range) floating point value that could not be parsed as an integer.
@@ -664,6 +664,7 @@ static bool parse_number(const wcstring &arg, number_t *number, wcstring_list_t
         double intpart = std::floor(floating);
         double delta = floating - intpart;
         *number = number_t{static_cast<long long>(intpart), delta};
+
         return true;
     } else {
         // We could not parse a float or an int.
@@ -835,7 +836,7 @@ int builtin_test(parser_t &parser, io_streams_t &streams, wchar_t **argv) {
     const wcstring_list_t args(argv + 1, argv + 1 + argc);
 
     if (argc == 0) {
-        return STATUS_CMD_ERROR;  // Per 1003.1, exit false.
+        return STATUS_INVALID_ARGS;  // Per 1003.1, exit false.
     } else if (argc == 1) {
         // Per 1003.1, exit true if the arg is non-empty.
         return args.at(0).empty() ? STATUS_CMD_ERROR : STATUS_CMD_OK;
@@ -858,11 +859,13 @@ int builtin_test(parser_t &parser, io_streams_t &streams, wchar_t **argv) {
 
     wcstring_list_t eval_errors;
     bool result = expr->evaluate(eval_errors);
-    if (!eval_errors.empty() && !should_suppress_stderr_for_tests()) {
-        streams.err.append(L"test returned eval errors:\n");
-        for (size_t i = 0; i < eval_errors.size(); i++) {
-            streams.err.append_format(L"\t%ls\n", eval_errors.at(i).c_str());
+    if (!eval_errors.empty()) {
+        if (!should_suppress_stderr_for_tests()) {
+            for (size_t i = 0; i < eval_errors.size(); i++) {
+                streams.err.append_format(L"\t%ls\n", eval_errors.at(i).c_str());
+            }
         }
+        return STATUS_INVALID_ARGS;
     }
     return result ? STATUS_CMD_OK : STATUS_CMD_ERROR;
 }
diff --git a/src/fish_tests.cpp b/src/fish_tests.cpp
index 02136d888..f60942542 100644
--- a/src/fish_tests.cpp
+++ b/src/fish_tests.cpp
@@ -2155,7 +2155,12 @@ static bool run_one_test_test(int expected, wcstring_list_t &lst, bool bracket)
     argv[i + 1] = NULL;
     io_streams_t streams(0);
     int result = builtin_test(parser, streams, argv);
+
+    if (expected != result)
+        err(L"expected builtin_test() to return %d, got %d", expected, result);
+
     delete[] argv;
+
     return expected == result;
 }
 
@@ -2207,13 +2212,13 @@ static void test_test() {
     do_test(run_test_test(0, L"' 2' -eq 2"));
     do_test(run_test_test(0, L"'2 ' -eq 2"));
     do_test(run_test_test(0, L"' 2 ' -eq 2"));
-    do_test(run_test_test(1, L"' 2x' -eq 2"));
-    do_test(run_test_test(1, L"'' -eq 0"));
-    do_test(run_test_test(1, L"'' -ne 0"));
-    do_test(run_test_test(1, L"'  ' -eq 0"));
-    do_test(run_test_test(1, L"'  ' -ne 0"));
-    do_test(run_test_test(1, L"'x' -eq 0"));
-    do_test(run_test_test(1, L"'x' -ne 0"));
+    do_test(run_test_test(2, L"' 2x' -eq 2"));
+    do_test(run_test_test(2, L"'' -eq 0"));
+    do_test(run_test_test(2, L"'' -ne 0"));
+    do_test(run_test_test(2, L"'  ' -eq 0"));
+    do_test(run_test_test(2, L"'  ' -ne 0"));
+    do_test(run_test_test(2, L"'x' -eq 0"));
+    do_test(run_test_test(2, L"'x' -ne 0"));
     do_test(run_test_test(1, L"-1 -ne -1"));
     do_test(run_test_test(0, L"abc != def"));
     do_test(run_test_test(1, L"abc = def"));
@@ -2284,6 +2289,10 @@ static void test_test() {
     do_test(run_test_test(1, L"-4611686018427387904 -ge 4611686018427387904"));
     do_test(run_test_test(1, L"4611686018427387904 -gt 4611686018427387904"));
     do_test(run_test_test(0, L"4611686018427387904 -ge 4611686018427387904"));
+
+    // test out-of-range numbers
+    do_test(run_test_test(2, L"99999999999999999999999999 -ge 1"));
+    do_test(run_test_test(2, L"1 -eq -99999999999999999999999999.9"));
 }
 
 static void test_wcstod() {
diff --git a/tests/test4.out b/tests/test4.out
index e0d388e26..a7af43eef 100644
--- a/tests/test4.out
+++ b/tests/test4.out
@@ -38,7 +38,7 @@ Test 22 pass
 6 0
 7 4
 8 0
-9 121
+9 2
 10 0 A
 11 1 B