Two years ago, choosing a closed frontier model was a no-brainer for serious work — the quality gap was real. In 2026 that gap has narrowed to the point where, for most production tasks, a good open-weight model is genuinely competitive. So the decision is no longer "which is smarter?" It's a trade-off across a few concrete dimensions.

The dimensions that decide it

Control and customization

Open weights let you fine-tune deeply, run on your own hardware, and pin a version that never changes under you. If your task is narrow and you'll tune it, or if "the model silently changed last Tuesday" is a problem you can't tolerate, open weights win.

Data and compliance

If data residency or "our prompts never leave our infrastructure" is a hard requirement — common in fintech, health, and government — self-hosting an open model is often the only option that clears legal. This alone decides many architectures.

Cost at scale

Hosted APIs are cheaper until you're not — at high, steady volume, self-hosting an open model on reserved GPUs can be dramatically cheaper per token. The crossover point depends on your utilization. Below it, pay the API; above it, run your own.

Latency and location

Self-hosting lets you put the model next to your data or your users — in-region, on-prem, or at the edge — which can matter more than raw quality for some products.

Closed models are a great default when you want capability without operational burden. Open models are the answer when control, compliance, cost-at-scale, or latency outweigh that convenience.

The pattern most teams land on

Not "pick one." Most serious teams in 2026 run a portfolio: a hosted frontier model for the hard, low-volume reasoning; a self-hosted open model for the high-volume, latency-sensitive, or compliance-bound bulk; and a routing layer that sends each request to the cheapest option that passes its eval. The model is a swappable part, and the architecture is built so swapping one is a config change, not a rewrite.

The takeaway

Don't religious-war this. Score your task on control, compliance, cost-at-scale, and latency, and let the answer fall out. And whatever you choose, build the seam so you can change your mind in six months when the landscape shifts again — because it will.