A compiler, but for voltages
EEBench, a benchmark built and funded by the team behind the hardware design tool atopile, published its first results on 4 September, from a run dated 1 September. Version 1 puts models through 13 analog and digital circuit design tasks and grades them without a human or a model in the loop.
The harness builds each submitted design, constructs the circuit graph and the bill of materials, then runs SPICE simulations and design rule checks against the task requirements. The maintainers compare it to handing a coding agent a compiler and a test suite, “except the tests are measuring voltages and component behavior.”
That determinism is the interesting part. Most evaluations of open-ended design work end up graded by a rubric, a human panel or another model, all of which drift. A circuit either meets the specification in simulation or it does not.
The scores
On the 1 September run, Anthropic’s Claude Opus 5 leads with 61.6 per cent across the 13 tasks. Grok 4.6 is second on 57.1 per cent, narrowly ahead of Claude Fable 5.1 on 56.4 per cent, Claude Fable 5 on 54.3 per cent and Claude Opus 4.8 Max on 51.4 per cent.
The spread is narrow — six points separate first from fifth — and every score is a long way from passing. A model that designs a working circuit three times in five is a model an engineer has to check every time, which is a different proposition from a tool that saves work.
EEBench also scores practical engineering trade-offs, not just whether the circuit functions: component cost and availability count. That matters, because a design that simulates perfectly from parts nobody stocks is not a design.
What it does not measure
The maintainers are direct about the limit. Simulation is not manufacture, and EEBench does not test whether a model can lay out a board, get it fabricated and bring up working hardware. Physical validation is stated as future work.
Two caveats belong with any early leaderboard like this one. Thirteen tasks is a small set, and a few unlucky items move a percentage a long way. And EEBench is built by atopile, a company selling tools in the field it is measuring — which is a reason to read the methodology rather than only the table.
What to watch is version 2, and specifically whether the layout and manufacturing stages arrive. Simulation scores tell you a model can reason about a circuit. They do not yet tell you anyone can ship the board.
