Frontier AI
OpenTelemetry tool_decision events now carry classifier_verdict and permission_mode, plus classifier_category when OTEL_LOG_TOOL_DETAILS=1 is set.
Why it matters. Security teams can audit, in their own observability backend, the classifier verdict and permission mode behind each tool decision.
Source: Claude Code releases
The release adds speed and usage.speed to Messages, and returns a Console organization's usage-tier spend cap as a read-only spend limit.
Why it matters. Python apps can set speed and read the spend cap in code, though some models still need a beta header to use speed.
Source: Anthropic Python SDK releases
The open model uses native 2-bit quantization to fit in 24GB of memory, and is on Hugging Face with llama.cpp, Ollama and vLLM support.
Why it matters. Teams can run agent workloads locally on RTX PCs, DGX Spark and RTX PRO workstations alongside other applications.
Source: NVIDIA blog
The AWS AI/ML skill lets Kiro or another coding agent select, train, optimize and deploy models on Amazon SageMaker AI, including JumpStart deployments.
Why it matters. It costs nothing beyond the AWS resources your agents use, and comes preinstalled in SageMaker Studio JupyterLab.
Source: AWS What's New