Route each prompt to the cheapest Gemini model that can answer it, and prove it with numbers. Scored against a 300-prompt golden set with an LLM judge.
Run of 2026-09-11: 300 prompts through six strategies, 1,800 answers scored by Gemini 3.1 Pro at temperature 0.
Fixed lite scores 4.82 against Pro's 4.90, a 2% gap, at 3% of Pro's cost and a quarter of its latency. On the 75 prompts labeled pro, lite still scores 4.64.
Semantic routing and fixed standard both match Pro's score. Semantic does it at 39% of Pro's spend and standard at 16%, so the honest "same quality, cheaper" number is standard or semantic, not lite.
Gemini 3.1 Flash-Lite reports CONFIDENCE 95 to 100 on every prompt, including the proofs and 700-word stories, so a 70% threshold never fires. Cascade is fixed lite plus a longer prompt. A self-rated confidence gate is not a verifier.
With a 2,048-token budget, 67 Pro answers were cut off mid-sentence because thinking tokens count against it, and the judge scored the truncation. The gateway default is now 8,192. Ten long-form answers still hit that.
Log-scale cost (X-axis) vs LLM Judge Quality Score (Y-axis, 1-5 scale). The ideal strategy appears in the top-left quadrant (high quality, lowest cost).
| Strategy | Prompts | Judge Score (1-5) | Quality Retention | Cost / 1k Req | Cost vs Pro | Latency (p50 / p95) | Tier Mix (L / S / P) | Routing Acc | Escalation |
|---|
| Category | Evaluated Prompts | Mean Score | Cost / 1k Requests |
|---|
The judge scores every strategy above against the same 300 prompts: eight task categories, each labeled with the cheapest Gemini tier a reviewer judged sufficient, plus the rubric the judge grades against. Every prompt is hand-written with its source material inline, and the set is validated to a 40 / 35 / 25 lite / standard / pro split. Browse every prompt →