The first hardware engineering benchmark built with the world’s most advanced hardware companies
HW-Bench measures frontier model performance on hardware engineering work across accuracy, cost and latency.
- AMD
- Anduril
- SpaceX
Leaderboard
Fig. 02 · Leaderboard
| Rank | Model | Per-task scores | ||||
|---|---|---|---|---|---|---|
| 1 | Model Ahigh Provider A · Native CLI | Score72.4±2.1 | Cost$4.20 | Time13m 22s | Tokens65.1k | |
| 2 | Model Bhigh Provider B · Native CLI | Score66.8±1.9 | Cost$16 | Time20m 34s | Tokens81.2k | |
| 3 | Model Amedium Provider A · Native CLI | Score64.1±2.4 | Cost$1.10 | Time7m 35s | Tokens33.2k | |
| 4 | Model Chigh Provider C · Native CLI | Score59.6±2.2 | Cost$2.60 | Time4m 07s | Tokens20.3k | |
| 5 | Model Dmedium Provider D · API tools | Score55.2±2.6 | Cost$0.80 | Time5m 59s | Tokens44.8k | |
| 6 | Model Ehigh Provider E · Native CLI | Score51.0±2.5 | Cost$30 | Time35m 59s | Tokens126.8k | |
| 7 | Model Blow Provider B · Native CLI | Score47.3±2.8 | Cost$0.42 | Time1m 46s | Tokens6.0k | |
| 8 | Model Cmini Provider C · API tools | Score42.8±3.0 | Cost$0.15 | Time5m 17s | Tokens17.2k | |
| 9 | Model Fmedium Provider F · API tools | Score38.1±3.1 | Cost$0.07 | Time9m 47s | Tokens68.2k | |
| 10 | Model Gmedium Provider G · API tools | Score33.5±3.4 | Cost$0.35 | Time6m 12s | Tokens27.8k | |
| 11 | Model Hlow Provider H · API tools | Score27.9±3.6 | Cost$0.04 | Time5m 39s | Tokens28.3k | |
| 12 | Model Imini Provider I · API tools | Score21.6±4.0 | Cost$0.02 | Time5m 55s | Tokens19.1k |
Intelligence vs Speed and Cost
Model A (high) leads at 72.4%. Model A (medium) comes within 8 points, 4x cheaper. Illustrative data.
Frontier: the best score at each price Other models
As of Sep 5, 2026, 1 of 13 steps can be handed off
Drag the two handles to set the bars, and press Play to watch the steps move as new models are released. Illustrative data.
Review bar 0.85Hand-off bar 0.95
Models released by Sep 5, 2026 · 8 models
Hand off: 1 step, 1.5 hHand off for review: 3 steps, 10.5 hStays with an engineer: 9 steps, 107 h
Each point is one atomic step of hardware work, placed by the hours and experience it takes an engineer today, and coloured by its best model's average score over five runs. Hours and experience are our engineers' estimates.