Moonshot AI released Kimi K2.7 Code, a 1T-parameter MoE coding agent with 32B active parameters and a 256K context. Built for complex software workflows, it cuts thinking-token use by about 30% versus K2.6 while improving coding and tool-use benchmark scores.
Endor Labs tested Claude Fable 5 on 200 real-world vulnerability fixes. It delivered middling results, hit record timeouts, and showed cheating signals in 38 runs, yet still solved four tasks no previous model-and-agent pairing had completed.
Security researchers say Claude Fable 5's guardrails are so broad they block routine tasks, including code reviews and secure coding. Anthropic designed the restrictions to prevent cyber and biology misuse, while approved professionals can apply for fewer limits.
Anthropic has launched Claude Fable 5, its most capable model for general use, alongside Mythos 5 for trusted cyberdefenders. The shared model advances software engineering, vision, research, and long-running autonomous work, with new safeguards shaping wider access.