How close is AI to designing real circuit boards?

AI models are getting surprisingly good at electrical engineering, but they work much better when writing code than clicking through CAD software.

EEBench released a benchmark evaluating how well models design real electronics using atopile, a tool that represents circuits as code. Instead of forcing AI agents to navigate desktop interfaces like KiCad, EEBench lets them write declarative circuit logic, run SPICE simulations, and select real components from manufacturer datasheets. Across 13 tasks, Claude Opus 5 took first place with a 61.6% score, while Grok 4.6 took second at 57.1%.

Why it matters: Ideal textbook circuits work on paper, but real hardware fails when voltage drops, component tolerances, and manufacturing costs collide. EEBench tests whether models can navigate those real-world trade-offs. Version 1 only grades simulation and logic rather than physical board layout, but top models are already far ahead of where engineers expected them to be.

Here's the gist: the full methodology and result explorer are public, letting you inspect the exact SPICE run captures and component picks for each model.

We are still far from prompting an entire smartphone into existence, but the hardware design loop is shortening fast.