# Open-weight inference pricing: the cheapest provider per model (prices checked Nov 7, 2026)

URL: https://ernd.ai/inference-prices

> ernd partners with [Wafer](https://ernd.ai/partners/wafer). Prices are from OpenRouter's public listings.

The cheapest provider for each popular open-weight model right now, taken from OpenRouter's public listings. Open a model to see all of its providers, with speed numbers too.

| Model | Cheapest provider | Input /1M | Output /1M | Blended 3:1 | Listings |
| --- | --- | --- | --- | --- | --- |
| [DeepSeek V4.1 Flash 1017](https://ernd.ai/inference-prices/deepseek-v4-1-flash-1017) | Particle AI (US, Flex) | $0.025 | $0.25 | $0.0813 | 51 |
| [DeepSeek V4.1 Flash](https://ernd.ai/inference-prices/deepseek-v4-1-flash) | Baidu Qianfan | $0.0507 | $0.203 | $0.0887 | 39 |
| [GLM 5.3 Flash](https://ernd.ai/inference-prices/glm-5-3-flash) | Relace | $0.0335 | $0.083 | $0.0459 | 41 |
| [Hy4 preview](https://ernd.ai/inference-prices/hy4-preview) | Tencent Cloud | $0.751 | $2.25 | $1.13 | 7 |
| [MiniMax M3.1 Flash](https://ernd.ai/inference-prices/minimax-m3-1-flash) | MiniMax | $0.075 | $0.30 | $0.131 | 13 |
| [DeepSeek V4 Flash 0731](https://ernd.ai/inference-prices/deepseek-v4-flash-0731) | Baidu Qianfan | $0.0444 | $0.133 | $0.0667 | 29 |
| [GLM 5.3](https://ernd.ai/inference-prices/glm-5-3) | Wafer (US) | $0.0699 | $1.40 | $0.402 | 47 |
| [Kimi K3.1 Pro](https://ernd.ai/inference-prices/kimi-k3-1-pro) | Relace | $0.40 | $7.88 | $2.27 | 24 |
| [Nemotron 3 Ultra](https://ernd.ai/inference-prices/nemotron-3-ultra-550b-a55b) | DeepInfra (Flex) | $0.40 | $1.76 | $0.74 | 5 |
| [Hy3](https://ernd.ai/inference-prices/hy3) | Tencent Cloud | $0.0825 | $0.33 | $0.144 | 5 |
| [Nemotron 3.5 Super VL](https://ernd.ai/inference-prices/nemotron-3-5-super-vl) | Darkbloom | $0.06 | $0.35 | $0.133 | 10 |
| [MiMo-V2.6-Flash](https://ernd.ai/inference-prices/mimo-v2-6-flash) | GMICloud | $0.112 | $0.224 | $0.14 | 8 |
| [DeepSeek V4 Flash 0423](https://ernd.ai/inference-prices/deepseek-v4-flash) | StreamLake | $0.0435 | $0.0871 | $0.0544 | 14 |
| [Kimi K3](https://ernd.ai/inference-prices/kimi-k3) | Wafer (US) | $0.276 | $9.60 | $2.61 | 30 |
| [Nemotron 3.5 Lightning](https://ernd.ai/inference-prices/nemotron-3-5-lightning) | Darkbloom | $0.025 | $0.12 | $0.0488 | 5 |
| [MiMo-V2.6-Pro](https://ernd.ai/inference-prices/mimo-v2-6-pro) | DeepInfra | $0.387 | $0.783 | $0.486 | 5 |
| [GLM 5.2](https://ernd.ai/inference-prices/glm-5-2) | Baidu Qianfan (fp8) | $0.189 | $0.594 | $0.29 | 33 |
| [DeepSeek V4 Pro 0813](https://ernd.ai/inference-prices/deepseek-v4-pro-0813) | Alibaba Cloud Int. | $0.581 | $1.74 | $0.871 | 25 |
| [Qwen3.8 27B](https://ernd.ai/inference-prices/qwen3-8-27b) | Chutes | $0.039 | $1.25 | $0.342 | 24 |
| [gpt-oss-120b](https://ernd.ai/inference-prices/gpt-oss-120b) | Prime Intellect | $0.015 | $0.055 | $0.025 | 26 |

Blended is three parts input price to one part output, per million tokens. Latency is time to first token and throughput is output tokens per second, both averaged from OpenRouter's daily figures for the last 7 days. Cheapest and fastest only count listings that were up when checked.

## Questions

### What is the cheapest way to run open-weight models?

A hosted API, where you pay per token and never touch a GPU. Of the 20 models here, gpt-oss-120b was cheapest on November 7, 2026, at $0.025 per million tokens blended from Prime Intellect. The provider matters too. For DeepSeek V4.1 Flash 1017, the priciest listing costs about 52 times the cheapest.

### What is the cheapest DeepSeek API provider?

On November 7, 2026, StreamLake had the cheapest DeepSeek listing on OpenRouter: DeepSeek V4 Flash 0423 at $0.0544 per million tokens blended ($0.0435 input, $0.0871 output). For DeepSeek V4 Pro 0813, the cheapest was Alibaba Cloud Int. at $0.871 per million tokens blended. DeepSeek's own API can price differently, so check its pricing page too.

### Where do these prices come from?

OpenRouter's public model and provider listings, checked November 7, 2026. Each row is one listing as OpenRouter shows it. A provider's own site can charge something different.

### How is the blended price worked out?

Three times the input price plus the output price, divided by four. It's a rough stand-in for chat and agent traffic, which sends more tokens in than it gets back. Cached input is billed at its own rate and isn't in the blend.

### Can you get cash back on inference spend?

Yes, if the card pays more on AI vendors than on everything else. The ernd card is a credit card for US companies. It earns 4% back on AI vendor spend and 1.5% back on everything else. Wafer charges paid directly on the card earn 10% back. Inference bought through OpenRouter credits is a charge from OpenRouter, so a provider's own perk doesn't apply to it.

The ernd card is a credit card for US companies. It earns 4% back on AI vendor spend and 1.5% back on everything else. [Request the card at app.ernd.ai](https://app.ernd.ai/login?next=%2Fapp%2Fcard%2Fapply&utm_source=llms&utm_campaign=inference_prices).