---
title: "What the savings figures are, and what they were computed on"
canonical: "https://router.xark.io/savings"
description: "Rate-card savings of 36–42%, rising to 65–68% on a repeated-prefix agent workload, with the arithmetic and the workload stated."
section: "core"
updated: "2026-09-02"
source: "https://router.xark.io/savings.md"
---

# What the savings figures are, and what they were computed on

Every open-weight model here is priced 36–42% below what its own publisher charges, rising to 65–68% on agent and chat workloads where a repeated prompt prefix is billed at the cached rate — and the publisher's rate is printed beside ours on every row so the figure can be checked rather than believed.

We guarantee at least 20% below the model publisher's own rate on every model we serve. The smallest discount in the catalogue today is 36%, so the guarantee has room in it by design.

**Availability:** 7 of the 24 catalogued models are callable today. The rest are listed at their intended rate but no configured upstream serves them, so a request naming one returns `no available channel`. Every table below marks which is which.

## The three numbers

| Figure | Value | What it assumes |
| --- | --- | --- |
| Guarantee | at least 20% below the publisher's own rate | A floor written well under the worst discount in the catalogue, so it survives a rate change. |
| Rate card | 36–42% | No caching. Blended across input and output, weighted 1:3 towards output. |
| Agent workload | 65–68% | 400M input tokens a month with a 60% repeated prefix, and 8M output |

## Per model

| Model | Rate-card saving | Agent-workload saving | Callable |
| --- | --- | --- | --- |
| Kimi K2.6 | 42% | 68% | Yes |
| Kimi K3 | 40% | 65% | Yes |
| GLM-5.2 | 42% | 68% | Yes |
| GLM-5.2 Air | 43% | 67% | No |
| DeepSeek V4 Pro | 37% | 66% | Yes |
| DeepSeek V4 Flash | 36% | 65% | Yes |
| Qwen3 Max Instruct | 43% | 68% | No |
| Qwen3 235B A22B | 42% | 67% | No |
| MiniMax M2 | 40% | 65% | No |
| Llama 4 Maverick | 39% | 65% | No |
| Mistral Large 3 | 40% | 67% | No |
| GPT-OSS 120B | 40% | 67% | No |
| Wan 2.2 T2V A14B | 42% | 42% | No |
| Wan 2.2 I2V A14B | 42% | 42% | No |
| LTX-2.5 | 40% | 40% | No |
| HunyuanVideo 1.5 | 40% | 40% | No |
| CogVideoX-5B | 43% | 43% | No |
| Mochi 1 | 40% | 40% | No |
| FLUX.2 [schnell] | 40% | 40% | No |
| Qwen-Image | 40% | 40% | Yes |
| Stable Diffusion 3.5 Large | 40% | 40% | No |
| HiDream-I1 | 40% | 40% | No |
| Qwen3 Embedding 8B | 40% | 40% | Yes |
| BGE-M3 | 40% | 40% | No |

The publisher's own rate is printed beside ours on every row of [the pricing table](https://router.xark.io/pricing.md), so each figure above can be checked rather than believed.