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

The team behind the atopile tool has launched the EEBench benchmark to answer a surprisingly hard question: can artificial intelligence design electronic circuits — and how do you even measure whether the resulting boards are any good? The prompt was a demo of GPT-6 Astra working on a board in KiCad, which OpenAI put on the front page of its announcement.

Code instead of mouse clicks

The authors note that models know far more about electronics than they let on in graphical CAD editors. When driving a visual interface, an agent spends most of its context on coordinates, menus and the state of the application. That's why EEBench uses atopile, where a schematic is described in declarative code: the agent works directly with components, connections and electrical constraints, and it can rebuild the project, run a simulation and figure out exactly what went wrong.

The real world doesn't forgive idealizations

One of the public tasks is built around a household electricity meter: when the 5 V supply drops out, the circuit must keep the processor above its 3.0 V shutdown threshold for another 20 ms so it has time to save its readings. Almost every model intuitively arrives at the right idea — add a capacitor. But a real ceramic capacitor under voltage delivers noticeably less than its rated capacitance, parts have tolerances, and excess capacitance costs money and slows recovery. A circuit that works at nominal values can fail with real components.

EEBench models the loss of power and measures the voltage throughout the entire dip, the effective capacitance at the operating point, the recovery, and compliance with constraints on package, dielectric and price. More complex analog tasks require, for example, synthesizing a low-pass filter on an op-amp and keeping the gain, cutoff frequency and Q factor within tolerance across the worst-case combinations of parameters. All the checks are deterministic: the system builds a graph of the circuit and a specification, runs SPICE simulations and compares every measurement against a limit. The key point is the use of real manufacturer components, which must actually exist, be orderable and fit within budget. That trade-off between performance, price and availability, in the authors' view, is exactly what engineering comes down to.

What the leaderboard shows

The authors call the results as of September 1 encouraging. Across the 13 tasks of EEBench V1, Claude Opus 5 leads with 61.6%, followed by Grok 4.6 (57.1%) and Claude Fable 5.1 (56.4%). OpenAI's models are lower so far: GPT-5.5 scored 42.3%, GPT-5.6 Sol scored 39.4%, and there's no result yet for GPT-6 Astra. Tellingly, xAI included EEBench in the model card for Grok 4.6.

The benchmark doubles as a reinforcement-learning environment: a failed run points to which voltage went out of bounds and whether the solution turned out to be needlessly expensive. So, can AI design boards? For a growing set of tasks — yes, already, the authors conclude, though you shouldn't blindly trust it to develop a pacemaker just yet.

Source: Hacker News