A quiet news weekend, dominated by a genuine first: Washington took a frontier AI model offline. Three stories, two threads to watch.

Late Friday the U.S. Commerce Department issued a national-security export-control directive barring Anthropic from serving Claude Fable 5 and Mythos 5 to any foreign national — including its own non-citizen staff. Unable to screen users in real time, Anthropic took both models fully offline, and they stayed dark through the weekend; Opus 4.8 and earlier Claude models are unaffected. The reported trigger was a "jailbreak," which Anthropic disputes as grounds for recall. It appears to be the first time a publicly deployed frontier model has been pulled by federal order. Anthropic → CNBC →

AEC angle: Provider availability is now a risk variable, not a constant. Any firm that wired Fable 5 into estimating, spec review, or QA lost it without warning. A March Logicalis survey found 16% of organizations have no continuity plan if a key AI provider goes dark — the multi-vendor, fallback-routing discipline we keep flagging just stopped being hypothetical. Logicalis →

Google's Gemini 3.5 Pro — announced at I/O in May with a 2-million-token context window and a "Deep Think" reasoning mode — is widely expected to reach general availability before June 30; only the Flash variant shipped at launch. Two million tokens is roughly an order of magnitude past Opus 4.8's and GPT-5.5's context ceilings. TechTimes →

AEC angle: Context length is the spec that matters most to document-heavy practice. 2M tokens is enough to hold an entire project record — full drawing set, specs, RFIs, and contract — in one session instead of chunking it. Watch the cost: Deep Think is gated to the $250/month tier, so the question is which workflows justify it.

Structured AI, an Oxford-founded, New York–based startup, closed a $4.2M seed (FCVC, with Y Combinator, 20VC, Cherry, Sequoia Scout), reaching $5M total. Its computer-vision agents run spell-check-style QA across entire drawing sets before engineer review, flag discrepancies against field work, and integrate with Revit; it's piloting with Syska Hennessy. The pitch: 70% of construction rework traces back to errors in the drawings. ENR → SaaS News →

AEC angle: Capital keeps concentrating at the front of the project lifecycle. It's the fourth preconstruction/QA raise we've logged in two weeks — after Endra (MEP), LightTable (drawing review), and Drafted (floor plans). The drawing set, not the jobsite, is where the AI money is going.