diff --git a/lib/std/Io/Threaded.zig b/lib/std/Io/Threaded.zig
index 999d7b93f503729e0f83f4c890ce6d809da5b21d..53d03354518e8bef833c8b248de7a1f23a402cc7 100644
--- a/lib/std/Io/Threaded.zig
+++ b/lib/std/Io/Threaded.zig
@@ -695,38 +695,38 @@ fn groupAsync(
 
     t.mutex.lock();
 
+    if (t.available_thread_count == 0) {
+        if (t.cpu_count != 0 and t.threads.items.len >= t.cpu_count) {
+            t.mutex.unlock();
+            gc.deinit(gpa);
+            return start(group, context.ptr);
+        }
+
+        t.threads.ensureUnusedCapacity(gpa, 1) catch {
+            t.mutex.unlock();
+            gc.deinit(gpa);
+            return start(group, context.ptr);
+        };
+
+        const thread = std.Thread.spawn(
+            .{ .stack_size = t.stack_size },
+            worker,
+            .{t},
+        ) catch {
+            t.mutex.unlock();
+            gc.deinit(gpa);
+            return start(group, context.ptr);
+        };
+
+        t.threads.appendAssumeCapacity(thread);
+    } else {
+        t.available_thread_count -= 1;
+    }
+
     // Append to the group linked list inside the mutex to make `Io.Group.async` thread-safe.
     gc.node = .{ .next = @ptrCast(@alignCast(group.token)) };
     group.token = &gc.node;
 
-    if (t.available_thread_count == 0) {
-        if (t.cpu_count != 0 and t.threads.items.len >= t.cpu_count) {
-            t.mutex.unlock();
-            gc.deinit(gpa);
-            return start(group, context.ptr);
-        }
-
-        t.threads.ensureUnusedCapacity(gpa, 1) catch {
-            t.mutex.unlock();
-            gc.deinit(gpa);
-            return start(group, context.ptr);
-        };
-
-        const thread = std.Thread.spawn(
-            .{ .stack_size = t.stack_size },
-            worker,
-            .{t},
-        ) catch {
-            t.mutex.unlock();
-            gc.deinit(gpa);
-            return start(group, context.ptr);
-        };
-
-        t.threads.appendAssumeCapacity(thread);
-    } else {
-        t.available_thread_count -= 1;
-    }
-
     t.run_queue.prepend(&gc.closure.node);
 
     // This needs to be done before unlocking the mutex to avoid a race with