Report: Opus 5.5 drafts permit-ready CAD drawings, builders say
A self-hosted model-to-AutoCAD loop reportedly drafts sealable 2D sets. Plus: independent benches deflate Le Chonk, and your JS framework carries a token bill.
Copy markdown
Your CAD stack, driven headless
Developers report wiring Claude Opus 5.5 into AutoCAD headless with a firm's own templates to produce permit-level 2D drawings; one practicing engineer called it 'a game changer for drafting,' roughly a junior drafter's output that a licensed pro still has to seal. The unlock for you: a self-built model-to-CAD pipeline can reach sealable sets without a $3,200-$8,500/month vertical tool like Blueprints AI. It's unverified beyond demos and firsthand accounts, so test your own templates before trusting a set.
Update: Le Chonk's coding scores don't reproduce
Mistral's 1T open-weight Large 4 'Le Chonk' (noted at 5AM) claimed 41% on Terminal-Bench 4.0 and 19.7% on FrontierCode 2.0; independent testers (Artificial Analysis, Vals.ai) measured roughly 25% and 15%, and put DeepSeek V4.1 Flash ahead on Terminal-Bench (33.8%), Vibe Code Bench (87.7% vs 81.6%) and the Vals Index (61.3% vs 56.4%). Le Chonk's real edge is efficiency, about 47% fewer tokens per task, plus human preference on code style. If you were about to repoint a coding agent, DeepSeek V4.1 Flash is still the stronger open coder on shared tests.
Your JS framework has a token bill
Nolan Lawson had an agent build the same content site in several frameworks and tallied the cost: Astro reportedly finished fastest with near-lowest spend and zero client-side JS, while a 'vanilla JS + Vite' baseline trailed several frameworks. The takeaway for agent-built sites: stable, well-documented framework APIs let the model work in fewer tokens, so framework choice is now a cost lever. Results are for content-driven apps; benchmark your own for heavily interactive UIs.