Go with custom AI model. I've spent weeks testing 8B–30B coding models on Azure A100s. The model that transformed our architecture? 137M parameters-trained in 3 minutes 35 seconds on an M2 Pro Mac mini. Microsoft’s FastContext revealed the bottleneck: repository exploration can consume 46.5% of a coding agent’s tokens. Their published experiments cut main-agent usage by up to 60.3%. I took the idea further in DeltaCode: → Go enumerates the legal repository context → Our custom 137M ranker handles routine selection → FastContext 4B handles only ambiguous READ/GLOB/GREP exploration → Custom Qwen3 4B performs coding preflight → Expensive models receive compressed evidence-not an entire repo → Compilers and tests retain final authority My custom ranker achieved: • 57.97% sealed Recall@5 • +5.80 points over the stock model • +14.49 points in Recall@1 • 73.33% Recall@5 on routed external cases • 0 execution authority The efficiency difference is enormous: • 96.6% fewer parameters than a 4B explorer • 99.5% fewer than a 27B–30B model • Only 0.3225% of its parameters were trained • Therefore 99.68% of the base remained frozen • 100% of A100 expense was avoided for this adapter • Estimated 35–55% reduction in cloud-context tokens once the cascade handles routine repository searches locally* The larger models did not fail because they lacked intelligence. They failed because we asked one model to explore, reason, format, code, operate tools, and verify itself. The smaller models won because each had one measurable job. I've used expensive A100 experiments to learn how not to require expensive compute in production. The future of coding agents may not be one enormous LLM. It may be a team of tiny specialists-each trained for one task, each independently measured, and none trusted with final authority.