Frontier models from the major providers, chosen per task and upgraded as the frontier moves. The model is deliberately swappable: your advantage lives in the brain, not the weights, so Heidi always rides the best available intelligence.
Different work wants different models: fast cheap ones for routine extraction, frontier ones for judgment. Heidi routes accordingly, and the 5 percent token cost line on your bill is exactly what that usage cost, transparently.
The strategic point matters more than the vendor list: because context and model are separated, every model upgrade anywhere in the industry makes your Heidi better the week it ships, with no migration.