scaledown/extract vs ultra-nvfp4
For 100,000 calls of 2,000 in and 500 out, scaledown/extract is cheaper: $8.00 a month at OpenRouter, against $240.00 for ultra-nvfp4 at Fireworks AI — 30.0× the cost.
Cheapest on each
scaledown/extract
$8.00/mo
cheapest, at OpenRouter — $0.04 in, $0 out per million
ultra-nvfp4
$240.00/mo
cheapest, at Fireworks AI — $0.6 in, $2.4 out per million
Aligned facts
Same rows on both sides, including the ones we cannot fill.
| Attribute | scaledown/extract | ultra-nvfp4 |
|---|---|---|
| Context window | 1M | ○ no answer |
| Accepts | Aa · IMG · FILE | ○ no answer |
| Sellers | 1 | 1 |
| Cheapest input rate | $0.04 / M | $0.6 / M |
| Cached input rate | ○ no answer | ○ no answer |
No quality or benchmark row appears here. We hold no throughput or accuracy measurement for either model, and a cost-per-quality figure built from a lab’s own claim would look authoritative and mean nothing.
Cost for this workload — both models, every seller
100,000 calls a month · 2,000 input and 500 output tokens each. Sellers of both models lead, cheapest combined cost first.
| Seller | scaledown/extract /mo | ultra-nvfp4 /mo |
|---|---|---|
| OpenRouter | $8.002026-10-27 | ○ no answer |
| Fireworks AI | ○ no answer | $240.002026-10-22 |
Cheapest tier per seller, this workload only. In and output rates, and a cached prefix, are broken out in the cost calculator.