o10/frontier is a virtual model: try a cheap model, have the judge verify that specific answer against your cluster's quality bar, escalate only on failure. Routing is prediction; this is runtime quality assurance. We sell the verified outcome, not the model.
o10/frontier request
→ classify (band, cluster — unchanged)
→ generate on cheapest verified rung (~2% of frontier cost)
→ judge verdict vs cluster quality bar
pass + confident → STREAM IT. done.
fail / low conf → generate on mid rung (~15%)
→ judge again
pass → stream
fail → fulfill on the real frontier model
→ every response: x-o10-frontier-trace: slm:loss(0.81) → mid:win(0.92)x-o10-frontier-trace: slm:loss(0.81) → mid:win(0.92)
Rungs tried, verdicts, confidences. When the cascade ends frontier-direct, you still paid the same fixed price — the trace shows exactly what answered.
Frontier answers, 40% off, receipt included.
Most requests clear on a cheap rung; every cascade terminates at the real frontier model. You pay one fixed per-token price anchored at 60% of the frontier reference.
escalations add seconds; C4 goes straight to frontier