Open vs. Closed Models: What You Actually Trade Away

Every model release restarts the same argument: open weights versus closed APIs. Strip the ideology and the tradeoff is four practical lines.

1. Cost shape

Closed models price per token — cheap to start, scales with usage. Open models price like infrastructure: you host (or rent) hardware and pay whether it works or sleeps. Light usage favors APIs; heavy, predictable usage favors self-hosting.

2. Privacy posture

Sending data to a closed API means trusting a vendor's retention policy. Running open weights means the data never leaves your machine — the whole reason regulated industries lean open. The caveat: a managed "private" endpoint is still someone else's server.

3. The frontier gap

Top closed models usually lead on raw capability; leading open models usually win on price-per-competence and customization. The gap moves both directions each quarter — which is why our model coverage tracks shipped benchmarks, not launch-day claims.

4. Lock-in mechanics

APIs lock in quietly: prompts, evals and workflows accrete around one vendor's quirks. Open weights lock in differently — to your own ops capacity. Running your own models is a skill bill, not a money bill.

The boring answer

Most teams run both: a closed frontier model where quality justifies it, an open model where volume, privacy or fine-tuning demands it. The question is not which side wins. It is which jobs belong to which bucket — and reviewing that split quarterly.