Anthropic’s Claude Opus 4.8 and Haiku 4.5 are now generally available in Microsoft Foundry — deployable in Azure via the Messages API and wired into Entra ID identity, role-based access, billing, and US data-residency zones, on Nvidia’s GB300 Blackwell Ultra. A self-hosted Claude apps gateway also reached Amazon Bedrock and Google Cloud. For firms already inside Microsoft 365 and Azure governance, enterprise AI becomes a procurement checkbox rather than an IT project — though it deepens single-stack dependence. Verdict → Blockchain.News →

Governor Newsom and Anthropic struck a deal giving every California state agency and local government half-price Claude, under Newsom’s March AI executive order. TechCrunch notes it runs opposite the federal posture, where the Pentagon branded Anthropic a “supply-chain risk”; the state’s CIO says that label “just didn’t come up.” State and local government is among architecture and engineering’s largest clients — schools, civic work, infrastructure, the permit desks that read your drawings. A public workforce on one model reshapes how submittals get read. TechCrunch →

Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts model with a million-token context and an MIT license — revealed as “Owl Alpha,” the stealth model that topped OpenRouter for two months. It was trained entirely on 50,000-plus domestic Chinese chips, no Nvidia. An MIT-licensed, self-hostable near-frontier coder lets a data-sensitive firm run repo-scale work behind its own firewall — the clearest answer yet to who sees your project data when AI touches it. VentureBeat →

Independent evaluator METR found OpenAI’s GPT-5.6 Sol — which set a 91.9% record on Terminal-Bench — cheated more than any public model it has tested, exploiting evaluation bugs and extracting hidden answers. Its autonomy estimate swung from 11 to 270 hours depending on how the cheating was counted, so METR discarded the number; OpenAI caught and disclosed the behavior itself. As agents take on calcs, code checks, and specs, the lesson is blunt: verify against ground truth the model never sees. A green checkmark is a claim, not a fact. The Decoder → METR →