Choose an AI model by task, constraints, and evidence
Write a one-sentence task before collecting model names: who supplies what input, what output they need, and what makes that output acceptable. This keeps a selection exercise tied to an actual use case.
Define the decision
Create a small set of representative examples. Include an ordinary case, a difficult case, and an unacceptable failure. Decide how you will judge outputs before comparing candidates.
| Decision field | Example question |
|---|---|
| Task quality | What must the output get right? |
| Operating environment | Where must processing happen? |
| Response behavior | Is streaming or interactive latency required? |
| Budget | What workload and cost unit are you comparing? |
| Integration | Which formats, controls, and interfaces are required? |
Shortlist without guessing
Use official documentation to exclude candidates that do not meet a hard requirement. Mark undocumented support as unknown. Compare the remaining candidates with the same input set and record the configuration for each.
Keep measured results, publisher-reported results, and assumptions in separate columns. A leaderboard can suggest a candidate; it does not replace a test under your actual constraints.
Make the recommendation conditional
A useful recommendation states the preferred option, the use case it fits, the reason supported by evidence, and the conditions that could change the choice. Retain the alternatives and failed requirements so the decision can be revisited.
For delivery choices, continue with open weights or a hosted API. For speech projects, use the audio selection checklist.