AI can now design circuit boards. EEBench proves which models actually work.
OpenAI demoed GPT-6 Astra doing hardware design. EEBench is testing if these AI-generated circuits survive real-world physics.
Tools · Source: Hacker News
What happened
OpenAI recently showcased GPT-6 Astra designing a circuit board in KiCad on their launch post front page. It is exciting to see hardware design enter the mainstream AI conversation. But watching an AI click around a graphical user interface does not prove the resulting hardware actually works. Graphical CAD tools force agents to waste context on coordinates and menus. The team at atopile built EEBench to solve this. They measure if AI models can actually engineer electronics by forcing them to write declarative code instead of drawing lines in a GUI.
The benchmark tests real physics and messy real-world constraints. One task asks the AI to build a residential energy meter power supply. When the five volt supply drops, the circuit must keep the processor alive for twenty milliseconds. Models often suggest adding a capacitor. But real capacitors lose capacitance under voltage bias and have strict tolerances. EEBench cuts the power in a SPICE simulation to see if the circuit survives. It also grades the AI on whether it picked real, cost-effective parts from actual manufacturer datasheets. Cost efficiency only matters if the circuit actually functions.
The September leaderboard results are surprising. Anthropic is currently dominating the hardware space. Claude Opus 5 scored 61.6 percent across thirteen tasks. Grok 4.6 took second place at 57.1 percent. xAI even included EEBench in their official model card for Grok 4.6, proving labs care about this category. OpenAI models are lagging behind right now. GPT-5.5 scored just 42.3 percent, and GPT-5.6 Sol scored 39.4 percent. The industry is now waiting to see how GPT-6 Astra will perform on these objective tests.
Key facts
- 61.6% — Claude Opus 5 score on EEBench, taking first place on the leaderboard
- 57.1% — Grok 4.6 score on EEBench, securing second place
- 39.4% — GPT-5.6 Sol score, showing OpenAI models currently lagging in this benchmark
- 20 ms — Time the energy meter circuit must keep the processor alive after a power loss
Why it matters
Hardware engineering is finally becoming a software problem. For people building with AI, this means you can start treating circuit design like code generation. Models are moving past spitting out textbook theory and ideal math. They are learning to balance electrical performance, component availability, and unit economics. This lowers the barrier to entry for hardware startups. You no longer need a massive team of electrical engineers to prototype a basic device. You just need an agent that understands declarative hardware languages and can iterate through simulation failures.
The second-order effect is a massive shift in how frontier models are trained. EEBench is not just a static test. It is a reinforcement learning environment. Failed SPICE simulations give models exact data on which voltage missed limits or which operating corner failed. Labs will use these simulation engines to train the next generation of engineering models. We are moving from models that say a schematic looks plausible to models that mathematically prove a circuit will not catch fire.
For builders
Train models on declarative hardware code
Stop forcing agents to click through graphical CAD tools. Use text-based hardware languages like atopile to give models direct control over components and constraints. This saves context window space, reduces visual processing errors, and lets the agent run simulations without leaving the project.
Sell simulation environments to AI labs
Frontier labs desperately need domain-specific reinforcement learning environments. If you can build programmatic grading systems for niche engineering fields, labs will pay you to help train their models. Atopile is already working directly with frontier labs using this exact business model.
Prepare for Grok 4.7 engineering capabilities
Elon Musk reportedly plans to release Grok 4.7 in weeks. It is reportedly trained on a massive collection of SpaceX data to make it exceptional at engineering. Hardware founders should watch this release closely. It could become the default model for physical engineering tasks if it delivers on these promises.
My take
Hardware is hard because the real world is messy. I love that EEBench forces AI to deal with component tolerances and actual supply chain costs instead of ideal textbook math. If your AI cannot pass a SPICE simulation, it is just hallucinating expensive garbage. We are not ready to let AI design pacemakers blindly, but this benchmark proves we are getting close to reliable automated hardware design.
Original reporting: Hacker News. This is my rewrite and opinion.