Full Summary
This Tuesday morning, OpenAI has launched its new GPT 5.5 Cyber model, designed specifically for cybersecurity work, and multiple sources including Digit.in and Decrypt confirm it outperforms Anthropic's Mythos 5 on key benchmarks. The updated GPT 5.5 Cyber scored 85.6% on CyberGym, a benchmark for reproducing software vulnerabilities, beating Mythos 5's 83.8%. OpenAI states this model is "more permissive" and "more capable for advanced, authorized cybersecurity work," intended for verified defenders. Meanwhile, Japanese startup Sakana AI has released its "Sakana Fugu" system, with Moneycontrol.com, News9live, and NewsBytes all reporting it outperforms Anthropic's Claude Fable 5 in coding and scientific reasoning. Sakana Fugu, which uses multiple AI models rather than a single one, scored 93.2 on the LiveCodeBench coding benchmark, while Claude Fable 5 scored 89.8. This multi-model approach aims to provide resilience against AI vendor restrictions and reduce dependence on a single provider. What nobody expected: China's GLM-5.2 model from Knowledge Atlas has surpassed all OpenAI and Google models on a key benchmark for real-world tasks. OfficeChai reports GLM-5.2 placed third on GDPval-AA v2, scoring 1524 Elo, higher than GPT-5.5's 1509. The Times of India and citybiz highlight GLM-5.2's focus on long-context reasoning and coding, with Caixin Global reporting its developer, Zhipu AI, has reached a HK$1 trillion market capitalization. Chinese models, including GLM-5.2 and Moonshot AI's Kimi K2.7 Code, are also significantly more cost-effective, with pricing about one-tenth that of Anthropic Fable 5, according to 디지털투데이. This means businesses and cybersecurity professionals now have more powerful and varied AI tools, but also face increasing competition and more affordable alternatives from new global players.