Anthropic said Claude Fable 5.2 scored 69.9% on FrontierSWE v2, outperforming GPT-6 Astra at 65.5% and Gemini 4 Carbon at 61.2%