Short answer
On evidence. Every model is benchmarked on FirstCoreAI's own hardware against frontier APIs for the specific job before it goes into production.
In detail: How FirstCoreAI benchmarks and picks AI models
What each model is measured on:
- Accuracy for the task, such as extraction, transcription or coding
- Latency
- Throughput under real concurrency
- Cost per unit of work
Sometimes a self-hosted open-weight model wins outright, and sometimes a frontier API is worth the premium. FirstCoreAI, FirstNet's AI business unit, shows you the evidence before you commit and puts each workload where it belongs: sovereign by default, frontier where it earns its keep.
Most enterprise workloads, such as document processing, transcription, classification, retrieval and internal assistants, run well on open-weight models in FirstNet's Durban data centre. When a request would use an overseas frontier model, FirstCoreAI tells you, because that means data leaving the country.