Best cheap AI models for high-volume chat

What to weigh for high-volume chat

High-volume chat - a support desk, an in-app assistant, a classifier that runs on every message - is decided by the price column first. Only models under $1 per million tokens blended qualify here, and within that band the intelligence index separates the ones that answer well from the ones that merely answer cheaply.

The small models have closed much of the gap. On this list you will usually find models within ten index points of the frontier at a twentieth of the price, which is why routing saves what it does: the simple majority of requests do not need the frontier, and here is what they can go to instead.

The rule to write is the one that keeps everything else away from them: a complexity ceiling, a quality floor, and escalation when an outcome comes back bad. The router learns from those outcomes which cheap model holds up for which kind of request and shifts traffic accordingly - a classifier that misses on refund intents, say, loses those requests to the next model up while keeping the greetings and order lookups it answers well. The savings ledger then shows what the cheap tier saved against the model you would otherwise have sent everything to.