US Treasury Secretary Scott Bessent went on Fox Business yesterday and dropped a clear line: the US supports open source but will sanction Chinese models that steal IP. The comment landed right after weeks of tension around open-weight releases and export controls.

It’s the first explicit threat tying model distribution to IP enforcement at the government level. Expect open-source Chinese labs to face extra scrutiny on training data provenance.

OpenAI had a busy 24 hours too. They rolled out GPT-Live, a new generation of voice models that make ChatGPT conversations feel closer to talking with a human. Two variants — GPT-Live-1 and the lighter mini — are live for users globally now.

They also confirmed GPT-5.6 is the preferred model inside Microsoft 365 Copilot. That lines up with earlier reports of the 1.5M context long-horizon agentic focus.

Meanwhile the rumor mill hit Anthropic hard. A weekend story claiming they were acquiring robotics startup Physical Intelligence spread fast before getting denied. Acquisitions in the physical AI space are accelerating — this one just happened to catch fire on X.

On the agentic coding front, Anthropic’s own 2026 trends report keeps resurfacing in dev circles. The standout number: Rakuten engineers gave Claude Code a 12.5 million line vLLM codebase task. It finished the activation vector extraction work in a single seven-hour autonomous run at 99.9% accuracy.

That’s production-grade output, not a toy demo. The same report notes internal legal teams with zero coding background cut review cycles from days to 24 hours using the tools.

Open-source options are keeping pace. OpenCode (from the SST team) now supports 75+ providers including fully local Ollama setups with a clean terminal UI. GLM 5.2 from Zhipu dropped under MIT license with strong SWE-bench Pro scores and a built-in coding harness. Google’s Gemini CLI remains the free high-capability baseline anyone with a personal Google account can run today.

Google’s June recap also landed: Gemini 3.5 Live Translate, Android 17 local model features, and the new Google Home Speaker built around Gemini all point to the same direction — on-device and edge agentic capabilities moving out of research and into shipping products.

The through-line this week is the same one we’ve been tracking: frontier labs pushing long-horizon agentic workflows while open-source and smaller teams close the gap on practical tooling. Geopolitics is now an explicit variable in model access.

The next 30 days will show whether the sanctions language turns into actual enforcement actions or stays as signaling. Either way, the agentic coding stack just got another layer of real-world constraints.