Stop by the variables before you shop the models.
1. Data policy: can any token leave your building? If no, you're on Ollama, vLLM, or a private cloud deployment. If yes, the frontier arms race applies.
2. Latency budget: real-time voice needs <500ms first token. Search-type features need fast streaming. Async batch jobs can wait for the reasoning model.
3. Cost envelope: $ per 1M tokens × your volume. If you ship 100M tokens/month of output, a 10x price difference is real money.
4. Output shape: JSON? tool calls? free text? Structured output quality varies a lot between models.
5. Ecosystem: do you need vision, audio, retrieval, function calling? Not all models do all four well.
6. Failure mode tolerance: chat for support can be wrong occasionally. Compliance docs can't be wrong ever.
Our default shortlist: Claude for quality, DeepSeek for value, Qwen for self-host, GPT-4.1 mini for speed. Test three on your real workload before you commit. That's the entire process.