EEBench grades AI circuit designs with SPICE, and the best model scores 61.6%

EEBench published its September 1 results on 4 September, scoring frontier models on 13 analog and digital circuit design tasks. Claude Opus 5 led on 61.6%, followed by Grok 4.6 on 57.1% and Claude Fable 5.1 on 56.4%. The two OpenAI models tested sit lower: EEBench reports GPT-5.5 at 42.3% and GPT-5.6 Sol at 39.4%. There is no GPT-6 Astra result yet.

The interesting part is how the grading works. EEBench says its checks are fully deterministic. It builds the submitted design, constructs the circuit graph and bill of materials, then runs SPICE simulations and design checks. Every requirement produces a measurement against a limit: gain, thresholds, ripple, transient response, and behaviour at component tolerance corners. No model is asked whether the schematic looks reasonable.

The team also skipped the graphical CAD tool. Designs are written in declarative code using atopile, so an agent works on components, connections and constraints directly instead of spending its context on menus and screen coordinates. EEBench says this tests electronics rather than computer use.

Where a design failed

One published failure shows why simulation catches what inspection misses. A submitted design specified 22 µF of capacitance. At 4.7 V bias the grader measured only 11.4 µF of effective capacitance, against a 545 µF requirement. The source built without errors. In simulation, the protected rail fell below its 3 V threshold after 0.85 ms, where the task required the processor to stay alive for 20 ms after power loss.

That gap comes from real parts. A ceramic capacitor can deliver far less than its rated value once there is voltage across it, and EEBench uses manufacturer parts with specifications pulled from datasheets into the SPICE model. Cost only counts once the circuit works.

An outside lab is using it

xAI included EEBench in the Grok 4.6 model card, in a section on engineering acceleration alongside 3D modelling and parametric CAD evaluations. xAI's own published run put Grok 4.6 at 60.0% using xhigh reasoning effort, which is higher than the 57.1% on EEBench's leaderboard. Those are two different runs at different settings, and the vendor figure is xAI's own measurement.

EEBench is built and funded by the team behind atopile, which also sells larger evaluation suites and training environments to labs. The site says it pays for the public benchmark runs and does not sell benchmark scores. Version 1 covers design and verification through simulation, and does not cover layout, manufacturing or board bring-up.

What this means for a team building software

The method transfers even if nobody on your team touches a circuit. EEBench works because the grader runs the thing and measures a number against a limit, so a design that reads well and fails at 0.85 ms is scored as a failure. Most teams evaluating model output on their own domain still use a second model as the judge, which cannot catch the case where output is fluent and wrong.

The score range is the other useful signal. A leading model getting 61.6% on tasks with objective pass conditions is a long way from work you would ship unchecked, and it is measured rather than asserted.

Source

Can AI design circuit boards yet?, EEBench, 4 September 2026.