Brick, a new AI routing system, cuts cloud costs 4.71× while matching the accuracy of the most expensive models, by smartly dispatching tasks to cheaper or more powerful AI models based on query difficulty.

What happened

Researchers introduced Brick, a system that rates each AI model across six capability dimensions and estimates query difficulty, then routes requests to the optimal model using a cost-aware dispatch rule. On a benchmark of 5,504 queries, Brick at maximum quality achieved 76.98% accuracy, exceeding the best single model (75.02%) and all tested routers.

Why it matters

Frontier (most advanced) AI models cost ten to one hundred times more than open-weight alternatives. At production scale, even small per-request savings directly reduce cloud billing. Brick's ability to achieve high accuracy at 4.71× lower cost than always using the strongest model makes expensive AI deployments more economical.

What to watch

Brick offers a tunable trade-off—operators can choose a max-quality profile (76.98% accuracy), a neutral cost-quality balance (74.11% accuracy at 4.71× savings), or a min-cost mode (22.15× cost reduction with 11.85 points accuracy loss). Response time dropped from 51.2s to 22.8s median latency.