AI API price vs measured latency — the same model, different hosts
The same model can cost 24.8× more depending on who hosts it. Qwen: Qwen3.8 27B is $0.040 per million input tokens at reka and $0.990 at cerebras — identical weights, 24.8× the price. Below, every multi-hosted model we track, with what each host charges and how fast we measured it answering.
Which models cost the most and least depending on the host?
| Model | Hosts | Cheapest | Dearest | Spread | Fastest measured |
|---|---|---|---|---|---|
| Qwen: Qwen3.8 27B | 4 | reka $0.040 | cerebras $0.990 | 24.8× | cerebras 78 ms |
| Z.ai: GLM 5.3 | 12 | inference-net $0.080 | glm $1.400 | 17.5× | fireworks 23 ms |
| MoonshotAI: Kimi K3 | 5 | inference-net $0.239 | baseten $3.000 | 12.6× | fireworks 23 ms |
| DeepSeek: DeepSeek V4 Flash 0731 | 7 | reka $0.050 | novita $0.409 | 8.2× | baseten 47 ms |
| DeepSeek: DeepSeek V4.1 Flash 1017 | 10 | deepinfra $0.078 | deepseek $0.300 | 3.8× | fireworks 23 ms |
| Z.ai: GLM 5.2 | 9 | deepinfra $0.450 | glm $1.400 | 3.1× | baseten 47 ms |
| DeepSeek: DeepSeek V4.1 Flash | 11 | inference-net $0.103 | siliconflow $0.300 | 2.9× | fireworks 23 ms |
| MiniMax: MiniMax M3 | 5 | deepinfra $0.224 | sambanova $0.600 | 2.7× | sambanova 24 ms |
| Z.ai: GLM 5.3 Flash | 11 | reka $0.080 | glm $0.150 | 1.9× | fireworks 23 ms |
| Reflection: Beam 1.0 501B A23B | 3 | nebius $0.250 | together $0.400 | 1.6× | baseten 47 ms |
| MiniMax: MiniMax M3.1 Flash | 3 | inference-net $0.100 | minimax $0.150 | 1.5× | inference-net 103 ms |
| NVIDIA: Nemotron 3.5 Super VL | 3 | together $0.150 | baseten $0.200 | 1.3× | baseten 47 ms |
| DeepSeek: DeepSeek V4 Pro 0813 | 5 | novita $0.990 | deepseek $1.320 | 1.3× | novita 80 ms |
| MoonshotAI: Kimi K2.7 Code | 3 | siliconflow $0.859 | nebius $0.950 | 1.1× | novita 80 ms |
| Xiaomi: MiMo-V2.6-Flash | 3 | deepinfra $0.133 | novita $0.140 | 1.1× | cohere 60 ms |
| Tencent: Hy4 preview | 4 | deepinfra $0.792 | siliconflow $0.834 | 1.1× | cohere 60 ms |
| StepFun: Step 5 Preview | 4 | inference-net $1.000 | stepfun $1.000 | 1.0× | inference-net 103 ms |
| Qwen: Qwen3.8 2.4T A95B | 3 | novita $2.000 | together $2.000 | 1.0× | novita 80 ms |
| Meta: Muse Spark Open | 4 | together $1.250 | baseten $1.250 | 1.0× | fireworks 23 ms |
Why compare hosts only within one model?
Because across models a price comparison means nothing. A cheap small model and an expensive large one are not competing for the same job, and a table that ranks them together ranks nothing. Held to a single model the comparison becomes exact: identical weights, identical task, and the only variables left are what the host charges and how quickly it answers.
Where do these two numbers come from?
They come from different places, and we keep them apart. Latency and uptime are measured by our own probes in 4 regions, every five minutes, against each host’s real API. Prices are not measured — providers publish them, and we read them from OpenRouter’s public catalogue. We do not blend the two into a single “value score”: that number would be our invention rather than either provider’s fact, and the weighting inside it would quietly decide the ranking. The two columns sit side by side and the trade-off stays yours.
Method: how latency is measured · Prices as published 2026-10-23 via OpenRouter · data & licence