Shopify reports 90% lower serving cost for specialized GraphQL model

Shopify says a specialized open model serving its Sidekick commerce assistant handles GraphQL tasks at about 90% lower serving cost than a frontier model. In replayed production traffic, the post-trained Nemotron 3.5 Lightning model returned complete responses in a median 0.87 seconds, compared with 6.50 seconds for the frontier API. Shopify trains the specialist using repaired production conversations and evaluates it with a judge calibrated against human labels. After six training loops, its internal score reached 74.5 versus the frontier model's 73. The frontier model still handles merchant conversations, planning and tool selection.