Cascade routing simulator
See what routing a share of simple prompts to the cheaper sibling saves, and how the parity breaker protects quality.
Routing 40% of gpt-6-astra prompts to its sibling gpt-6.1-sol on a 2,000-in / 300-out call saves $1,120 a month at 100,000 calls (32%); moving every call would save $2,800. The classifier decides per prompt, shadow evaluation re-runs a sample on gpt-6-astra, and a pair under 99.8% parity is switched off automatically.
· Source: pricing table 2026-10-26 and the gateway's own tokenizer and benchmark files; your inputs never leave the browser except token counts, which are not stored
At a glance
- Model
- gpt-6-astra
- Cascade sibling
- gpt-6.1-sol
- Per call: model / sibling
- $0.03500 / $0.00700 (computed)
- Saving at 40% simple share
- $1,120/month (32%) (computed)
- Saving if every call moved
- $2,800 (computed)
- Parity floor
- 99.8% rolling, 20-sample minimum
- Breaker
- Clopper-Pearson upper bound on disagreement; conformal alpha 0.05
- Scope
- same provider only; advisory by default (X-TokenZero-Cascade: advise)
- Reasoning guard
- caps reasoning effort to low on simple prompts (X-TokenZero-Reasoning-Cap)
gpt-6-astra: 2,000-in / 300-out call, 100,000 calls per month, table 2026-10-26
| Simple-prompt share | Per month | Saved | Saving |
|---|---|---|---|
| 0% | $3,500 | $0 | 0% |
| 20% | $2,940 | $560 | 16% |
| 40% | $2,380 | $1,120 | 32% |
| 60% | $1,820 | $1,680 | 48% |
| 80% | $1,260 | $2,240 | 64% |
| 100% | $700 | $2,800 | 80% |
Simulator
Turn it on: X-TokenZero-Cascade: auto per request or cascade: auto on the workspace; shadowEval: true adds the parity monitor. The recommendation is always returned in X-TokenZero-Cascade-Recommendation, even in advise mode.
Questions people ask
What counts as a simple prompt?
A deterministic classifier scores each request on length, structure, tool use, requested output shape and the presence of reasoning cues; only prompts it scores simple are eligible, and the score is returned in the recommendation header so you can audit it.
What stops quality from dropping?
Shadow evaluation re-runs about 1% of routed requests on the original model with your key and compares outputs; a pair whose rolling parity falls under 99.8% is auto-disabled for the workspace, and the conformal upper bound on disagreement is checked before every route.
Does it route across providers?
No. The sibling is always the same provider's cheaper model, so tools, keys and request shape never change; cross-provider moves only happen on failover when you name the fallback.
What does TokenZero cost?
Free up to $500 per month of routed spend; then 15% of verified savings with a 3x-ROI-or-$0 rule. Gateway overhead is 1.52 ms p50, measured.