The most common question we get about mixture-of-agents runs is not "does it work" but "what does it cost." Fair. A panel consumes a multiple of a single call's tokens, and you should see that multiple before you commit to it. So let's work one example end to end.
All numbers below are illustrative: we use round token counts and a round frontier-model rate of $3 per million input tokens and $15 per million output tokens, which is in the neighborhood of published frontier pricing as of mid-2026. Your models, prompts, and rates will differ; the structureof the arithmetic won't.
The setup
One run: a 1,000-token prompt, a panel of 4 proposers that each write an ~800-token draft, one critique layer where each model revises after reading the others, and a judge that synthesizes the final answer.
Layer by layer
Propose. Four models each read the 1,000-token prompt and write an 800-token draft.
- Input: 4 × 1,000 = 4,000 tokens → $0.012
- Output: 4 × 800 = 3,200 tokens → $0.048
- Layer total: $0.060
Critique.This is where the token count jumps, and it's worth seeing why. Each critic re-reads the prompt plus all four drafts — 1,000 + 3,200 = 4,200 input tokens — and writes an ~800-token revised draft.
- Input: 4 × 4,200 = 16,800 tokens → $0.050
- Output: 4 × 800 = 3,200 tokens → $0.048
- Layer total: $0.098
Judge. One model reads the prompt plus the four revised drafts (4,200 tokens) and writes the ~800-token final answer.
- Input: 4,200 tokens → $0.013
- Output: 800 tokens → $0.012
- Layer total: $0.025
The totals
| Call | Input tokens | Output tokens | Cost |
|---|---|---|---|
| Single-model call | 1,000 | 800 | $0.015 |
| Full 4-model panel run | 25,000 | 7,200 | $0.183 |
So the panel run costs roughly 18 cents against 1.5 cents for the single call — about 12x. Two things are worth noticing. First, the multiple is bigger than 9 (the raw call count) because critique and judging re-read the drafts: input tokens, not output tokens, dominate a mixture's bill. Second, input tokens are 5x cheaper than output tokens at typical rates, which is the only reason the multiple isn't worse.
Routing changes the blended number
Nobody should send every prompt through a full panel. With smart routing, a classifier sends easy prompts to a single model and reserves the panel for prompts that warrant it. Suppose — again, illustrative — 70% of your traffic routes to a single model and 30% runs the full panel:
- Routed: 0.70 × $0.015 = $0.0105
- Full panel: 0.30 × $0.183 = $0.0549
- Blended provider cost per request: ~$0.065 — about 4x a single call, not 12x
On Moamao, routing also changes the orchestration side of the bill: a routed prompt counts as a quarter of a run, so the 70% of traffic that skips the panel is cheap on both meters. The blended cost tracks how hard your traffic actually is, not the worst case.
Why BYOK keeps this cheaper
Everything above was priced at provider list rates, and that's the point of bring-your-own-keys. With BYOK, the panel's model calls run on your own Anthropic, OpenAI, Google, and open-weights accounts at whatever rates you already have — including any negotiated discounts, committed-use pricing, or prompt-caching savings. Moamao charges a flat per-run orchestration fee for the fan-out, critique choreography, judging, and the stored trace. It never marks up tokens.
Compare that to a platform that resells model access through its own keys with a margin on tokens. A markup that looks tolerable on single calls compounds badly here, because a mixture multiplies token volume ~12x: at a 30% token markup, our example run costs $0.238 instead of $0.183 — you'd pay roughly 5.5 cents of margin per run, which at 100,000 panel runs a month is $5,500 of pure markup. Multi-model workloads are exactly where token margins hurt most, so the fee should sit on the orchestration, which is flat per run, not on the tokens, which scale with panel size.
Run the same arithmetic with your own prompt lengths and rates before adopting any mixture setup — ours included. If the 12x (or, blended, 4x) doesn't buy you answer quality you can measure, use one model. If it does, make sure you're paying list price for the tokens.