Anthropic claims that Claude Sonnet 4.5 scored 77.2 percent on the SWE bench benchmark, beating GPT-5 and Gemini 2.5 Pro.