The most interesting model release of Q1 2026 wasn't from a frontier lab. It was the 8B-parameter reasoning models — DeepSeek-R2, Qwen3-Think, and a quietly competent batch of open-source distillations — that suddenly score within ten points of GPT-class models on math, code, and structured tool use.
That changes the economic shape of every agent product. Routing the easy 70% of requests to an 8B model running on a single H100 cuts inference cost by an order of magnitude. The remaining 30% — the genuinely hard, multi-hop, ambiguous-context tasks — still need frontier models, and that's fine. They're a smaller bill.
The architecture shift is a small-model-first router with a confidence-based escalation policy. The small model attempts the task, emits a confidence signal (or you score the output with a cheap classifier), and only failures get escalated. We've seen 60–80% cost reduction with no measurable quality regression on customer-facing chat, classification, and routine code edits.
The trap is benchmarking on the easy tier and assuming parity everywhere. Small models still collapse on long-horizon planning, ambiguous user intent, and any task requiring the model to know what it doesn't know. Score the escalation rate, not the average quality.
If your inference bill is the second-largest line on your infra spreadsheet, you have a routing problem, not a model problem.
