diff --git a/demos/benchmark/gen_benchmark.py b/demos/benchmark/gen_benchmark.py
new file mode 100644
index 00000000..adbe44fc
--- /dev/null
+++ b/demos/benchmark/gen_benchmark.py
@@ -0,0 +1,47 @@
+#!/usr/bin/env python
+#
+# A simple benchmark of the tornado.gen module.
+# Runs in two modes, testing new-style (@coroutine and Futures)
+# and old-style (@engine and Tasks) coroutines.
+
+from timeit import Timer
+
+from tornado import gen
+from tornado.options import options, define, parse_command_line
+
+define('num', default=10000, help='number of iterations')
+
+# These benchmarks are delicate.  They hit various fast-paths in the gen
+# machinery in order to stay synchronous so we don't need an IOLoop.
+# This removes noise from the results, but it's easy to change things
+# in a way that completely invalidates the results.
+
+@gen.engine
+def e2(callback):
+    callback()
+
+@gen.engine
+def e1():
+    for i in range(10):
+        yield gen.Task(e2)
+
+@gen.coroutine
+def c2():
+    pass
+
+@gen.coroutine
+def c1():
+    for i in range(10):
+        yield c2()
+
+def main():
+    parse_command_line()
+    t = Timer(e1)
+    results = t.timeit(options.num) / options.num
+    print 'engine: %0.3f ms per iteration' % (results * 1000)
+    t = Timer(c1)
+    results = t.timeit(options.num) / options.num
+    print 'coroutine: %0.3f ms per iteration' % (results * 1000)
+
+if __name__ == '__main__':
+    main()