1. Name the job, not the vibe

Split work into buckets: classification and routing (cheap, high volume), retrieval-augmented answers (needs context window + faithfulness), long-form drafting (quality and tone), structured extraction (JSON schema discipline), embeddings and search (separate model family entirely). Each bucket has different tolerances for latency and price. Mixing them into one “do everything” prompt is how bills and incident reports grow together.

2. Default to smaller for triage, larger for synthesis

Ticket tagging, intent detection, first-pass moderation, and “should we escalate?” flows run well on compact chat or classifier endpoints with tight prompts. Reserve frontier-scale models for steps that truly need broad reasoning or long documents—and cap concurrency so a traffic spike does not become a finance spike.

3. Pin versions and measure on your data

In production, pin model IDs (or dated snapshots where vendors offer them) and upgrade on a schedule with evaluation—not because a changelog dropped. Maintain a small golden set of real prompts and expected shapes; regression-test before flipping defaults. Leaderboard scores rarely predict your enterprise PDFs or your users’ typos.

4. Embeddings are not interchangeable

Vector search quality depends on matching embedding models at index and query time, dimensionality, and distance metrics. Treat embedding choice as infrastructure: migrating later means re-embedding corpora. Pick a stable offering and document the pairing with your vector store.

5. Hosted API vs self-host

Use hosted APIs when you want SLAs, compliance attestations, and zero GPU ops. Consider self-hosted or dedicated open-weight models when data residency, air-gapped environments, or flat economics at huge token volume dominate—accepting that you now own capacity planning and security patching.

6. What I rarely prioritize in v1

  • The model that “feels smartest” in a one-off ChatGPT thread.
  • Multimodal stacks before text flows are measured and bounded.
  • Fine-tuning before prompt + RAG + eval discipline—most products are not there yet.

For business applications, the right model is the one your team can operate: predictable price per task, clear failure modes, and upgrades you control. Everything else is a lab exercise until those boxes are checked.