Pathfinder reads your real Gemini usage, replays it against the 3.x models inside your own project, and tells you exactly where every workload should land — and what it will actually cost. Measured, not guessed.
Each stage does one job and hands the next a cleaner artifact. Stages marked ● need a human decision — pick any to see what it takes to run it.
These came out of a real run against real Gemini calls — and each one contradicts the assumption everyone starts with.
The "5× price shock" was supposed to multiply the bill. Measured, moving everything to 3.5 Flash cut it a quarter — because 2.5 Flash was burning 96% of its output on hidden reasoning that 3.5 Flash skips by default.
Nine in ten of one workload's responses were hitting the token cap and returning invalid JSON — today, on 2.5 Flash, before any migration. The thinking tokens were eating the output budget.
The frontier model a heavy workload needs doesn't exist in the project's regions. Pathfinder prices it, flags it modelled, and routes it to a human — it never quietly emits a number it couldn't measure.
Here, a cost is never just a number. It carries its own evidence — and a total is only as trustworthy as the weakest figure inside it.
The clock is the only thing not negotiable.