# Inference hosts and what they charge

On October 20, 2026 we read 711 paid model prices from 9 inference hosts and 1 router, each from the host’s own pages. The same weights don’t cost the same: DeepSeek V4 Flash 0731 is 7.3× more per output token on Novita AI than on DeepInfra.

Prices read from each host's own pages; latest read 2026-10-20.

## Hosts

| Host | Kind | Models priced | Lowest output /1M | Top score served | Measured by us |
|---|---|---|---|---|---|
| [OpenRouter](https://benchlm.ai/hosts/openrouter) | Router | 422 | $0.03 | 88 (Claude Opus 5) | 2026-09-01 |
| [Groq](https://benchlm.ai/hosts/groq) | Inference host (LPU) | 4 | $0.30 | 73 (Qwen3.8-27B) | 2026-09-01 |
| [Cerebras](https://benchlm.ai/hosts/cerebras) | Inference host (wafer-scale) | 2 | $0.75 | 73 (Qwen3.8-27B) | 2026-09-01 |
| [Together AI](https://benchlm.ai/hosts/together) | Inference host | 31 | $0.14 | 79 (Kimi K3) | 2026-09-01 |
| [DeepInfra](https://benchlm.ai/hosts/deepinfra) | Inference host | 92 | $0.03 | 88 (Claude Opus 5) | 2026-09-01 |
| [Fireworks AI](https://benchlm.ai/hosts/fireworks) | Inference host | 18 | $0.20 | 81 (Qwen3.8 Max) | 2026-09-01 |
| [Baseten](https://benchlm.ai/hosts/baseten) | Inference host | 16 | $0.26 | 79 (Kimi K3) | Not yet |
| [SambaNova](https://benchlm.ai/hosts/sambanova) | Inference host (RDU) | 6 | $0.59 | 65 (MiniMax M3) | Not yet |
| [Nebius Token Factory](https://benchlm.ai/hosts/nebius) | Inference host | 27 | $0.24 | 79 (Kimi K3) | Not yet |
| [Novita AI](https://benchlm.ai/hosts/novita) | Inference host | 93 | $0.02 | 81 (Qwen3.8 Max) | Not yet |

## The same model on different hosts

OpenRouter is left out because it bills another host's rate.

| Model | Score | Hosts | Cheapest output /1M | Highest output /1M | Spread |
|---|---|---|---|---|---|
| [Qwen3.8 Max](https://benchlm.ai/models/qwen3-8-max) | 81 | 3 | $4.95 (DeepInfra) | $6.00 (Novita AI) | 1.2x |
| [Kimi K3](https://benchlm.ai/models/kimi-k3) | 79 | 6 | $14.25 (DeepInfra) | $15.00 (Together AI) | 1.1x |
| [Qwen3.7 Max](https://benchlm.ai/models/qwen3-7-max) | 76 | 3 | $3.75 (Novita AI) | $7.50 (Together AI) | 2x |
| [DeepSeek V4.1 Flash 1017](https://benchlm.ai/models/deepseek-v4-1-flash) | 73 | 6 | $0.29 (DeepInfra) | $1.20 (Together AI) | 4.1x |
| [Qwen3.8-27B](https://benchlm.ai/models/qwen3-8-27b) | 73 | 5 | $1.49 (Cerebras) | $4.00 (Groq) | 2.7x |
| [Muse Spark Open](https://benchlm.ai/models/muse-spark-open) | 72 | 3 | $4.25 (Baseten) | $4.25 (Nebius Token Factory) | 1x |
| [GLM-5.3-Flash](https://benchlm.ai/models/glm-5-3-flash) | 69 | 6 | $0.50 (Baseten) | $0.50 (Together AI) | 1x |
| [DeepSeek V4 Pro 0813](https://benchlm.ai/models/deepseek-v4-pro-0813) | 69 | 5 | $2.60 (DeepInfra) | $3.96 (Together AI) | 1.5x |
| [GLM-5.3](https://benchlm.ai/models/glm-5-3) | 68 | 6 | $2.50 (DeepInfra) | $4.40 (Together AI) | 1.8x |
| [Kimi K2.6](https://benchlm.ai/models/kimi-2-6) | 67 | 3 | $3.40 (Novita AI) | $4.00 (Nebius Token Factory) | 1.2x |
| [MiniMax M3](https://benchlm.ai/models/minimax-m3) | 65 | 6 | $1.10 (DeepInfra) | $2.40 (SambaNova) | 2.2x |
| [GLM-5.2](https://benchlm.ai/models/glm-5-2) | 65 | 5 | $1.80 (DeepInfra) | $4.40 (Together AI) | 2.4x |
| [DeepSeek V4 Flash 0731](https://benchlm.ai/models/deepseek-v4-flash-0731) | 62 | 5 | $0.18 (DeepInfra) | $1.32 (Novita AI) | 7.3x |
| [Llama 3.3 70B](https://benchlm.ai/models/llama-3-3-70b) | 61 | 4 | $0.32 (DeepInfra) | $1.20 (SambaNova) | 3.8x |
| [Gemma 4 31B](https://benchlm.ai/models/gemma-4-31b) | 56 | 4 | $0.34 (DeepInfra) | $1.15 (SambaNova) | 3.4x |

## FAQ

### What is an LLM inference provider?

An inference provider runs a model on its own hardware and sells access by the token through an API. Groq, Cerebras, Together AI, DeepInfra and Fireworks host open-weight models such as GPT-OSS and GLM. A router such as OpenRouter forwards your request to one of those hosts and bills you their rate.

### Which inference provider is cheapest?

It depends on the model. No single host had the lowest output rate on every model we matched on October 20, 2026, so the table below names the cheapest host per model. Our paid September 1, 2026 run found that caching and billing details can move the real bill away from the list price.

### Which inference provider is fastest?

In our paid September 1, 2026 run, Cerebras returned the fastest median, 0.35 seconds on GPT-OSS 120B, and Groq came next at 0.43 seconds on Qwen3.8 27B. Those were different models, and listed tokens per second on a host’s own page is a vendor figure we did not reproduce.

Canonical page: https://benchlm.ai/hosts