Right now the leading labs genuinely lead. The frontier is real, and you should use the best model you can for the work in front of you.
But look five years out. It may not matter which lab you started with. New entrants, open-source models and fine-tuned variants keep arriving, and the lead changes hands more often than procurement cycles do.
A component, not a commitment
So we treat the model as a component. You pick the best one for each job, and you change it the moment a better one ships, without rebuilding the work around it.
The opposite is quiet lock-in. When your capability is welded to one provider, their release schedule, their pricing and their limits slowly become yours.
One job, one tile
Each agent does one job, which we call a tile. Forecasting, collections, board reporting: each is its own tile, with its own definition of done.
A tile does not care which model sits underneath it. The work, the memory and the guardrails belong to the tile. You can move from a frontier model to an open one, and the tile still runs.
The model is a component, not a commitment.
Cost is an architecture choice
Running the same work for far fewer tokens is not a discount you negotiate. It is an architecture: send only the context that matters, check memory before re-reading, and route volume to smaller models.
Same work, a fraction of the tokens. That advantage is structural, and it holds whichever model is best this month.
Bring your own API keys
Because the model is just a component, you use your own accounts and your own negotiated rates. We never mark up your tokens, so you are never paying a margin on someone else's relationship with a lab.
What to keep portable
Keep three things in your hands and the model stops mattering: the work, the memory, and the right to swap. Hold those, and every new model on the frontier is an upgrade you opt into, not a migration you survive.
