PCBBench

PCBBench

AI agents design PCBs in KiCad from a prompt. We check each board with scripts, price it and have an AI judge review its layout.

Tasks
1
Runs
7
Harnesses
4
Models
3
Updated
Harness
Model
Change the mix

Each group's weight sets its share of the 100 points. The defaults are the official score.

  • 20%
  • 20%
  • 30%
  • 20%
  • 10%
Rank
1Opus 5.5Anthropic, medium reasoningSprocketSprocket81.2$53.314 h 51 min90.1M$73.36
2GPT-6.1 SolOpenAI, medium reasoningSprocketSprocket71.3$53.721 h 36 min185.1M$148.08
3Opus 5.5Anthropic, medium reasoningBackplaneBackplane69.9$56.772 h 50 min124.9M$51.57
4Fable 5.2Anthropic, medium reasoningBackplaneBackplane68.4$49.093 h 45 min55.6M$52.68
5GPT-6.1 SolOpenAI, medium reasoningCodex desktop + text-to-cadCodex desktop + text-to-cad56.4$14.361 h 36 min52.4M$87.53
6GPT-6.1 SolOpenAI, medium reasoningBackplaneBackplane29.5−20, gate failed: Fatal review finding$5.811 h 38 min28.6M$181.34
7GPT-6.1 SolOpenAI, medium reasoningCodex desktopCodex desktop25.9−20, gate failed: Fatal review finding$6.411 h 18 min21.0M$77.64

Board price is for one assembled board when ordering 5 from PCBWay, without shipping.

Scores are out of 100 and cover the board only. How boards are scored and what each task asks for.