Anthropic opens its frontier systems to outside evaluators as local open models reach Macs

October 16, 2026

Anthropic began giving six researchers from the independent evaluator METR employee-like access to its systems, including hardware, office space and internal tools, so they can examine its frontier models and safety practices more deeply.

This appears to move external model oversight from published test results toward supervised access to the inside of a lab, which matters beyond Anthropic because stronger AI governance increasingly depends on outsiders being able to inspect how systems are built and controlled rather than only what they output. For Microsoft and other cloud and model vendors, the pressure is likely to rise to offer comparable ways for customers, regulators or evaluators to verify safety claims without having to trust a provider's own summaries.

Meta published Muse Spark Open, an open-weights multimodal conversational and coding model, on Hugging Face, and Apple and Meta said developers will be able to run it locally on high-end Macs with Apple silicon. Open weights means the model parameters are released for others to download and run themselves, rather than only through the provider's own service. This likely shifts some advanced AI work another step away from rented cloud access and toward ordinary software distribution on customer-controlled machines, where cost, latency and data handling can be set by the user instead of the model vendor. It also sharpens the choice that Microsoft now faces across Windows PCs and its cloud: if capable third-party models arrive on mainstream hardware with first-party support from platform owners, cloud control and local control start to compete inside the same enterprise and developer workflows.

Bloomberg made its ASKB conversational AI interface generally available on the Bloomberg Terminal. OpenAI introduced the Rosalind Academic Plan for eligible U.S. life-sciences labs. Amazon unveiled Bee Frontier and Bee Eclipse personal AI devices that continuously listen to conversations and generate summaries and reminders.