Beating GPT-4o and Claude 3.5 on SWE-bench Lite through repeated sampling