Model choice trades quality for cost and speed

Frontier models (GPT, Claude, Gemini) deliver the best quality but at the highest cost and latency. Open models served via providers (Together, Groq, Fireworks, DeepInfra) offer 5 to 20x cost reductions at a quality discount that narrows each quarter. Local models (Ollama, llama.cpp) eliminate per-token cost entirely but require hardware investment. That quality discount has effectively closed for coding tasks: open models like Ornith 1.0 now match or beat frontier models on SWE-bench and Terminal-Bench while running on a laptop with no API key.

The practical pattern is multi-model routing: use a cheap, fast model for simple tasks (classification, routing, formatting) and a frontier model for complex reasoning. But the threshold for “complex enough to need frontier” keeps rising as open models improve. OpenRouter (see: an LLM gateway sits between your app and providers) and similar aggregators make multi-model routing trivial by exposing all models behind one API. The tradeoff is never just quality versus cost; it is quality versus cost versus latency versus privacy, and the right answer changes with each model release. Inference is profitable (see: inference is profitable training is what costs), which means open-model providers will keep driving prices down.