At Build 2026 on June 2, Microsoft announced two in-house AI models developed without OpenAI data or infrastructure (CNBC): MAI-Thinking-1, its first flagship reasoning model (TechTimes), and MAI-Code-1-Flash, a 5-billion-parameter coding model built for GitHub Copilot (Microsoft AI). MAI-Code-1-Flash applies adaptive reasoning - allocating more compute to complex requests while remaining concise on simple ones - and was trained directly on production Copilot tooling and licensed data to improve performance on real agentic coding workflows (Microsoft AI). Microsoft reports the model outperforms Claude Haiku 4.5 across four core coding benchmarks, with a 16-point lead on SWE-Bench Pro (Microsoft AI). Both models are rolling out to all GitHub Copilot plans via the model picker in Visual Studio Code (Thurrott).
Anthropic announced an expansion of Project Glasswing - its controlled vulnerability-discovery program using Claude Mythos - to roughly 150 additional organizations across more than 15 countries, now covering power, water, healthcare, communications, and hardware sectors not represented in the initial cohort (TechCrunch). Across all scanned targets to date, Mythos has identified more than 10,000 high- or critical-severity vulnerabilities, including a flaw in the wolfSSL cryptography library - used by billions of devices - that would permit certificate forgery for banking and email domains (Help Net Security). In open-model releases, NVIDIA debuted Nemotron 3 Nano Omni on June 1, a 30-billion-parameter mixture-of-experts model that activates roughly 3 billion parameters per inference and combines vision, audio, and language processing in a single architecture for agentic workloads (NVIDIA Blog); NVIDIA reports 9x higher throughput than comparable open omni models, with weights available on Hugging Face and as an NVIDIA NIM microservice (The Next Web).