Claude Opus 5 tops EEBench, where circuits are graded by physics
A new benchmark grades AI-generated circuits the way hardware is actually graded — in a simulator, against real datasheets, at every tolerance corner. The early results are better than anyone expected and nowhere near good enough.
OpenAI put a printed circuit board in a GPT-6 Astra demo, and the team behind atopile answered with the missing half: EEBench, a benchmark that hands models real electrical engineering tasks and then grades the result by running it through SPICE against components pulled from actual manufacturer datasheets. The setup is deliberately unglamorous. Rather than asking a model to click around a graphical CAD tool — where most of its context gets eaten by coordinates, menus and window state — the agent writes the circuit in code, and the harness rebuilds the design, extracts the bill of materials and simulates it. One public task is based on a residential energy meter: when the 5 V supply drops, the circuit has to hold the processor above its 3.0 V brownout threshold for another 20 milliseconds so it can save the accumulated reading. Every model reaches for the same answer — add a capacitor — and then the task gets hard, because a real ceramic part delivers far less than its advertised capacitance once voltage is across it, tolerances stack, and extra capacitance costs money, board area and recharge time.
That is the point, and it is the reason this benchmark is worth more than its leaderboard position. Getting the equations right is table stakes; EEBench also makes the agent pick parts that exist, can be ordered and fit the budget, then re-runs the simulation at worst-case tolerance corners. Grading is fully deterministic, cost efficiency only counts once the circuit actually works, and a failure says which rail missed which limit rather than whether a schematic looked plausible. On the September 1 results across 13 tasks, Claude Opus 5 leads at 61.6%, Grok 4.6 is second at 57.1% and Claude Fable 5.1 third at 56.4%. The OpenAI models sit much further down, with GPT-5.5 at 42.3% and GPT-5.6 Sol at 39.4% — and there is no Astra result yet, which is the number the field is waiting for.
Two things follow. The first is that the same deterministic checks double as a reinforcement-learning environment: a failed run carries a usable gradient, which is how you turn "the schematic looks right" into "the circuit holds at 3.0 V through every corner." The second is the uncomfortable gap between the demo and the scoreboard. Astra routing a board in KiCad looks like mastery; a 61.6% top score on 13 tasks is a competent junior who still needs review. EEBench's own verdict is the honest one — for a useful and growing set of circuit problems the answer is already yes, and you still would not let one design a pacemaker and install the result blind. We covered the other end of this same push earlier in the week — AI designs, verifies and deploys a chip in two weeks — where an end-to-end system built an accelerator from a written spec. Between that and xAI publishing EEBench in a model card, hardware is quietly becoming the next serious eval surface.
Separately, Alabama lawmakers are working on AI rules aimed at keeping children safe, the latest state-level effort to arrive while Congress leaves the field empty. The shape is familiar: youth-facing protections, transparency and parental-control language rather than any attempt to govern frontier training runs. It lands in a week when the politicians who wrote the first wave of state AI statutes publicly asked frontier labs to slow down — a sign that the state-level center of gravity is shifting from disclosure paperwork to actual limits on deployment. Watch whether Alabama's bill follows the pattern of exportable model language or stalls in committee.
If a model can pass 6 of 13 physics-graded circuit tasks, would you ship its board without a human signing off on the bill of materials? Tell us in the comments.
