Appreciate the shout-out to Parametric CAD Bench in CDFAM’s AI to CAD Benchmarks 2026 report. The report calls out a couple important findings from our Parametric CAD Bench cadbench.ai. First, our bench evaluates editability and parametric structure, reflecting a key principle in Physical AI: looking right isn’t enough—it must have real-world engineering usability. Second, the agent harness matters: with the same underlying model, changing the agent harness can shift benchmark performance by roughly 10% in either direction, showing that engineering AI progress depends on more than foundation models—it also depends on agents, tools, representations. What metrics are you paying attention to when evaluating engineering AI? Drop your thoughts below! 👇 #AI #GenerativeAI #gNucleusAI #CAD #EngineeringAI #ParametricCAD #AICAD #ParametricCADBench #AIAgents #FreeCAD #Benchmark #CDFAM

See you at the Forum!

We are honored and excited to be heading to the Gemini Startup Forum! 🚀✨ Hosted by @GoogleStartups in partnership with @GoogleDeepMind and @googlecloud , this unique two-day event brings together founders and innovators from all over the world. For us, the real magic is the opportunity to engage in direct, forward-thinking dialogue with the Founders and Google's leading AI experts, top researchers, and engineers. Working side-by-side with these brilliant minds and gaining access to world-class technical resources—is an incredible catalyst as we continue to grow the future of AI for engineering teams. Anyone else attending the forum? We’d love to connect! ☕🤝 startup.google.com/programs/… PS: We're hiring. #GeminiStartupForum #GoogleDeepMind #GoogleCloud #EngineeringAI #GenerativeAI #gNucleusAI #Startups #TechLeadership

V3 now includes Opus5.5: xitter.cf/gNucleusAI/status/2103…

🚀 Parametric CAD Bench V3 is here. 100 new tasks — significantly more complex than V2 — across Create, Create + Edit & Drawing → CAD. 🏆 Astra — 56.87% | $317.81 🥈 Fable 5.1 — 56.75% | $1,056.04 🥉 Opus 5 — 50.86% | $1,012.21 A few insights: ✏️ All 6 tasks unsolved by every model were Create + Edit. 📐 Gemini 3.8 & Grok 4.7 are relatively close to the leaders on Text → CAD, but fall much further behind on Drawing → CAD — making multimodal engineering understanding a key differentiator. 💰 Astra costs ~3.3× more/token than Kimi K3, yet total benchmark cost was almost identical ($317.81 vs. $314.55) — showing that for harder agentic tasks, model efficiency can matter more than token price. 🤖 Agent framework also matters: Gemini 3.8 Flash scored 36.19% with mini-swe-agent (w/ multimodal bridge) vs. 27.56% with Antigravity. 📊 Leaderboard: cadbench.ai/ 📄 V3 Report: cadbench.ai/news/cad-bench-v… #CADBench #AICAD #ParametricCADBench #EngineeringAI #AIAgents #FreeCAD #Benchmark #ClaudeFable51 #GPT6Astra #Grok47

As promised, we tested Claude Opus 5.5 — and it’s now 🚀#1 on Parametric CAD Bench V3. 🏆 Opus 5.5 + Claude Code — 61.03% | $573.94 🥈 Opus 5.5 + mini-swe-agent — 57.96% | $498.28 🥉 GPT-6 Astra + Codex — 56.87% | $317.81 A few highlights: 📐 Drawing → CAD improved significantly: 62.64% → 84.66% vs. Opus 5. 💰 Better and more efficient: 41.6% fewer tokens and 43.3% lower cost than Opus 5. 🤖 Agent still matters: the same Opus 5.5 scores 61.03% vs. 57.96% with different agents. The frontier in multimodal Engineering AI is moving fast. 📊 cadbench.ai/ 📄 cadbench.ai/news/cad-bench-v… #CADBench #ParametricCADBench #AICAD #EngineeringAI #AIAgents #MultimodalAI #FreeCAD #Benchmark #ClaudeOpus55 #GPT6Astra

The San Francisco 3DEXPERIENCE Regional User Group (SF3DXRUG) meeting is happening tomorrow, September 23rd, on Google Sunnyvale Campus, and our team at gNucleus AI is excited to be attending! As the local CATIA and 3DEXPERIENCE community comes together in the Bay Area, we’re looking forward to connecting with fellow engineers, designers, and innovation leads to discuss the future of Engineering AI. If you’re attending tomorrow’s event, let’s sync up! We’d love to meet you, exchange insights, and discuss where engineering design is headed next. See you there! 📍 #SF3DXRUG #3DEXPERIENCE #CATIA #EngineeringAI #GenerativeAI #CAD #DesignEngineering #BayAreaEngineering #gNucleusAI #AICAD #PhysicsAI

We will be there!

We tested GPT-6 Astra x-high + ChatGPT Work on FreeCAD 1.1.0. We’re excited to see frontier models making clear progress on engineering tasks, while recognizing that gaps still remain before they can be reliably used in real production environments. Evaluation Results: 1. Car Transmission — Demo Reproduction • Prompt: generate the car transmission as the Astra CAD demo in FreeCAD • ✅ Successful STEP assembly in ~7 min (picture1) • ~400 LOC of generated FreeCAD code 2. Car Transmission — with Feature-Based Modeling • Prompt: generate the car transmission as the Astra CAD demo in FreeCAD using Part Design features, sketches, and constraints • ⚠️ Partially successful: ~60 min, ~7 FreeCAD crashes (picture2) • Gear teeth generated with polylines rather than smooth B-splines(picture3) 3. Motion Simulation • Prompt: generate the car transmission motion simulation as the Astra CAD demo in FreeCAD • ❌ Unsuccessful 4. Unseen CAD data — our own in-house Hairpin Winding parts • Prompt: generate the hairpin windings with detailed specs and dimensions provided in a document • ⚠️ More limited performance: generated simplified geometry with limited detail(picture4) TL;DR → Still mostly a coding agent, not Computer Use for CAD → Good at approximate geometry, but sketches, constraints, dimensions, and features remain challenging → Performance drops on unseen CAD and simulation tasks #GPT6 #Astra #ChatGPTWork #FreeCAD #CAD #EngineeringAI

Parametric CAD Bench V2 is live! 🚀🚀🚀 We’re releasing a major update to our benchmark for AI agents building parametric CAD models. What’s new in V2: • 10 frontier models: 8 closed-source models and 2 open-source models • FreeCAD 1.1.0: a more stable environment with better support for geometry, assemblies, and simulation • Grader improvements: tighter spec checks and better handling of reward hacking Leading benchmark scores: 🥇 Claude Fable 5.1 — 84.81% | $198.58 🥈 GPT-6 Astra — 84.78% | $135.62 🥉 Grok 4.6 — 82.21% | $75.02 Fable 5.1 and Astra are statistically tied on our 100-task benchmark, while Grok 4.6 is the most cost-efficient among the top three. Parametric CAD Bench V3 is coming soon with more types of parametric CAD tasks and more complex tasks. 📖 Read the V2 report: cadbench.ai/news/cad-bench-v… 📊 Explore the CAD Bench V2 leaderboard: cadbench.ai/ 🤝 Contributors: cadbench.ai/contributors #CADBench #AICAD #ParametricCAD #EngineeringAI #AIAgents #FreeCAD #Benchmarking #ClaudeFable51 #GPT6Astra #Grok46

We were surprised to see how quickly interest in Engineering AI is growing. 🚀 Our CAD datasets and benchmarks have reached 20,000 downloads on Hugging Face this month alone 📈, and we’ve also become one of the top 3 most-downloaded CAD datasets on Hugging Face. 🏆 It’s exciting to see more AI researchers and teams exploring CAD, simulation, and engineering tasks. 🛠️🤖 Engineering is quickly becoming one of the next major frontiers for AI. 🔥 Check out our datasets and benchmarks here: huggingface.co/gnucleus-ai #EngineeringAI #AICAD #AI #HuggingFace #gNucleusAI #ManufacturingAI

Excited to see gNucleus AI’s Parametric CAD Bench featured in the Grok 4.6 Model Card from @grok ! 🚀 It was great to see the strong gains Grok 4.6 has made in engineering and CAD, especially its impressive balance of performance and cost efficiency. ⚙️📈 We’ve also updated the Parametric CAD Bench leaderboard with more model–agent combinations, giving a broader view of how different models and agent harnesses perform on real parametric CAD tasks. We’ll keep building high-quality benchmarks that reflect real-world engineering tasks and pushing Engineering AI forward. 🛠️🤖 Grok 4.6 Model Card: media.x.ai/v1/website/card-4… Parametric CAD Bench: cadbench.ai/ #Grok46 #EngineeringAI #AICAD