The frontier-vs-open debate is usually argued as identity. It should be argued per use case. The right model is the one that clears the quality bar at acceptable cost, latency, and governance for that specific job.
The decision framework
- Use frontier when capability is the bottleneck — genuinely hard reasoning where smaller models fail, and cost is justified by value.
- Use open/tuned when control is the bottleneck — regulated data, cost ceilings, latency needs, or the requirement to host in your own boundary.
- Tune a smaller model on your data when the task is narrow and repetitive; it often beats a generic frontier model on your specific domain.
- Don't marry one provider. Abstract the model layer so you can route per task and swap without a rewrite.
Bottom line: match the model to the job — frontier for hard reasoning, deployable open models for control, and route per use case rather than per trend.