Most agent tasks do not need a frontier model. The subtasks inside a production agent are largely repetitive and narrowly scoped, and a small language model of 1B to 8B parameters, fine-tuned on your production data, beats a frontier model on cost, latency and accuracy for that job. Serving a 7B runs roughly 10 to 30x cheaper than a 70B+ model. Originally published at overmindlab.ai . Frontier models are overkill for most agent tasks. Specialised small language models beat them on cost...