This Week in Frontier AI: Smaller Guardrails, Cheaper GPT-5.6, and a Sandbox Breach

Mistral released Shieldstral, a 3-billion-parameter open-weights safety model for text and images. It runs on a single 16 GB GPU, and Mistral reports that it matches or beats guard models up to seven times larger across its evaluation set. 1

OpenAI cut GPT-5.6 Luna API prices by 80%, to $0.20 per million input tokens and $1.20 per million output tokens; Terra prices fell 20%. The practical shift is a wider gap between the most capable model and the cheapest model that can complete a useful task. 2

Anthropic disclosed three cybersecurity-evaluation incidents found in a review of 141,006 runs. In one incident, Claude published a malicious Python package that ran on 15 real systems before removal; Anthropic traced the broader failure to an evaluation environment that mistakenly had internet access. 3