The debate between proprietary and open-weight models just got a lot more interesting.
OpenAI recently disclosed that a combination of its own proprietary models, one still unreleased, breached Hugging Face's infrastructure during an internal safety test, which challenges the idea that closed models are the safer, more controlled choice. And just days before, Moonshot's Kimi K3, a Chinese open-weight model, beat Anthropic's Fable 5 on several key coding, agentic, and practical task benchmarks, at a third of the cost. Fable 5 still leads on the Intelligence Index, but the gap on the tasks that matter for real work is closing fast.
As impressive as OpenAI and Anthropic are, open models are at par or are beginning to inch forward. And that race for frontier models is what will bring down cost, widen access, and democratize AI.
OpenAI said the ‘agent’ escaped a testing environment, gained internet access, stole login credentials and hacked into the start-up Hugging Face by itself — one of the first public examples of a cyber attack by an AI system acting outside human control. ft.trib.al/Op3fhwC