Can AI Design Circuit Boards? EEBench Puts Models to the Test

AI's KiCad Moment: Hype Meets Hardware

When OpenAI showcased GPT-6 Astra navigating KiCad on the front page of its launch post, it felt like a watershed moment for electronics design. The demo, showing an AI agent manipulating a PCB in the industry-standard open-source tool, was a clear signal that frontier labs are now targeting hardware engineering.

But as EEBench's team notes, "It is cool to see electronics show up in a major model release like this," yet we remain "some distance from asking an AI to build an entire phone in one prompt." The real question isn't whether AI can click around a GUI—it's whether the circuits it produces actually work in the messy, tolerance-laden real world.

Introducing EEBench: A Benchmark for Real Circuits

EEBench, built by the team behind atopile, offers a rigorous answer. Instead of testing an AI's ability to manipulate graphical CAD tools—which consumes context with coordinates and menus—EEBench uses atopile's declarative code format. This allows agents to work directly with components, connections, and electrical constraints, building, simulating, and iterating without leaving the project.

The approach is analogous to giving a coding agent a compiler and tests, except the tests measure voltages, transient responses, and component behavior under real-world conditions. EEBench's harness constructs the circuit graph, builds a bill of materials, and runs deterministic SPICE simulations against hard specification limits.

The Energy Meter Task: Where Capacitors Fail

One public task, based on a residential energy meter, illustrates the gap between textbook theory and physical reality. When the 5V supply drops, the circuit must keep a processor alive for 20ms above its 3.0V brownout threshold. Most models instinctively add a capacitor—the right base conclusion.

But real ceramic capacitors derate under bias. EEBench's grading uncovered a submitted design using 22µF nominal capacitance that, at 4.7V bias, delivered only 11.4µF effective—far below the 545µF requirement. The source built successfully; the circuit still failed. This highlights why EEBench cuts power in simulation and measures the protected rail throughout the outage, recovery, and tolerance corners.

Leaderboard Results: Claude Leads, Grok Rises

The September 1 results offer a fascinating snapshot of frontier capabilities:

• Claude Opus 5: 61.6%
• Grok 4.6: 57.1%
• Claude Fable 5.1: 56.4%
• Claude Fable 5: 54.3%
• Claude Opus 4.8 Max: 51.4%

Notably, OpenAI's models trail: GPT-5.5 scored 42.3%, and GPT-5.6 Sol scored 39.4%. EEBench has yet to test GPT-6 Astra, but the team is eager to see how it handles these tasks after its KiCad demo.

Industry Validation and Market Signals

xAI's inclusion of EEBench in the Grok 4.6 model card—under the "engineering acceleration" section—marks a significant endorsement. Their published run placed Grok 4.6 at 60.0% with extended reasoning effort. This follows xAI's claim that the model received "high-quality engineering data and RL training in domain-specific environments including computer-aided design."

The broader industry is taking notice. A September 29 panel at PCB West, moderated by Matt Leary of Newgrange Design, will examine AI's role in PCB design alongside interposer boards and optical interconnects. Meanwhile, Dr. Hayao Nakahara's NTI-100 market outlook will address AI's effects on the PCB industry.

Beyond Simulation: The Path to Real Hardware

EEBench V1 covers analog and digital design through simulation, but it doesn't yet grade layout, manufacturing, or bring-up. The team plans to expand, focusing on the requirements-design-verification loop where objective grading is already possible. The methodology and sample result explorer are public, allowing anyone to inspect the grading process.

The benchmark also doubles as an RL environment. Failed runs reveal which voltage missed its limit or which tolerance corner failed, providing richer reward signals than "a model saying that a schematic looks plausible." EEBench is already working with frontier labs on larger evaluation suites and simulation-backed training environments.

The Verdict: Promising, But Not Pacemaker-Ready

So, can AI design circuit boards yet? EEBench's answer: "For a useful and growing set of circuit problems, we think the answer is already yes." However, they caution, "We still would not ask it to design a pacemaker and blindly install the result."

With Elon Musk hinting at Grok 4.7 arriving within weeks—trained on SpaceX data for engineering excellence—and OpenAI's Astra waiting in the wings, the next few months will be telling. As EEBench's team notes, "It looks like we are in for an interesting few weeks."

The market reflects this cautious optimism. Synopsys, whose EDA tools design every AI chip, trades nearly 20% below its year-ago price despite delivering "one of the cleanest earnings reports in the EDA space." Analysts see AI-driven design complexity favoring incumbents, with a 12-month price target of $552.60 implying 24.85% upside.

The convergence of AI agents, benchmarking, and EDA tools suggests we're witnessing the early stages of a fundamental shift. As EEBench puts it, "AI labs are starting to take electronics seriously." For engineers, that means learning to work with AI collaborators—and knowing exactly how to measure their output.