DeepSeek V4.1 Flash 1017 服务商比价
同一个模型，不同服务商的价格可以差 18 倍。下表按「输入+输出」合计价格从低到高排列。
服务商数量
45 家
官方输入价
$0.150
每百万 tokens
最低输入价
$0.025
Particle AI
价差
18 倍
价格最低的几家里，有 7 家使用 FP4/MXFP4 低精度量化。低精度会牺牲输出质量，适合分类、抽取、格式转换这类容错任务；涉及长链推理、代码生成、专业判断时，建议选 FP8 以上的服务商。
全部 45 家服务商
单位：美元 / 百万 tokens。按输入+输出合计排序。「相对官方」以官方输入价为基准。
| 服务商 | 输入 | 输出 | 合计 | 量化 | 上下文 | 相对官方 |
|---|---|---|---|---|---|---|
| Sail Research | $0.101 | $0.150 | $0.251 | fp4 | 1,048,576 | 官方 67.3% |
| Sail Research | $0.101 | $0.150 | $0.251 | fp4 | 1,048,576 | 官方 67.3% |
| Particle AI | $0.025 | $0.250 | $0.275 | unknown | 1,048,576 | 官方 16.7% |
| DeepInfra | $0.078 | $0.235 | $0.314 | fp8 | 1,048,576 | 官方 52.3% |
| Morph | $0.035 | $0.315 | $0.350 | fp8 | 1,048,576 | 官方 23.3% |
| DeepInfra | $0.098 | $0.294 | $0.392 | fp8 | 1,048,576 | 官方 65.3% |
| Relace | $0.035 | $0.500 | $0.535 | fp4 | 1,048,576 | 官方 23.3% |
| Wafer | $0.038 | $0.500 | $0.538 | mxfp4 | 1,048,576 | 官方 25.3% |
| OpenInference | $0.038 | $0.500 | $0.538 | fp4 | 1,048,576 | 官方 25.7% |
| Particle AI | $0.050 | $0.500 | $0.550 | unknown | 1,048,576 | 官方 33.3% |
| DeepInfra | $0.147 | $0.441 | $0.588 | fp8 | 1,048,576 | 官方 98.0% |
| Decart | $0.123 | $0.492 | $0.615 | fp4 | 1,048,576 | 官方 82.0% |
| InferenceNet | $0.080 | $0.550 | $0.630 | fp8 | 1,040,000 | 官方 53.3% |
| InferenceNet | $0.080 | $0.550 | $0.630 | fp8 | 1,040,000 | 官方 53.3% |
| Wafer | $0.160 | $0.500 | $0.660 | mxfp4 | 1,048,576 | 高 1.1 倍 |
| Ionstream | $0.165 | $0.495 | $0.660 | unknown | 1,048,576 | 高 1.1 倍 |
| Cloudflare | $0.100 | $0.600 | $0.700 | unknown | 1,048,576 | 官方 66.7% |
| StreamLake | $0.144 | $0.576 | $0.720 | fp8 | 1,024,000 | 官方 96.0% |
| StepFun | $0.150 | $0.600 | $0.750 | fp8 | 1,000,000 | 同價 |
| DeepSeek | $0.150 | $0.600 | $0.750 | unknown | 1,048,576 | 同價 |
| Baidu | $0.150 | $0.605 | $0.754 | fp8 | 1,048,576 | 官方 99.9% |
| Alibaba | $0.165 | $0.660 | $0.825 | unknown | 1,000,000 | 高 1.1 倍 |
| ModelRun | $0.170 | $0.670 | $0.840 | fp4 | 1,048,576 | 高 1.1 倍 |
| Cohere | $0.195 | $0.660 | $0.855 | unknown | 1,048,576 | 高 1.3 倍 |
| DigitalOcean | $0.180 | $0.720 | $0.900 | unknown | 1,048,576 | 高 1.2 倍 |
| GMICloud | $0.180 | $0.720 | $0.900 | fp8 | 1,048,575 | 高 1.2 倍 |
| AtlasCloud | $0.195 | $0.780 | $0.975 | fp8 | 1,048,576 | 高 1.3 倍 |
| Novita | $0.195 | $0.780 | $0.975 | fp8 | 1,048,576 | 高 1.3 倍 |
| Fireworks | $0.210 | $0.840 | $1.05 | unknown | 1,048,576 | 高 1.4 倍 |
| Relace | $0.300 | $0.900 | $1.20 | fp4 | 1,048,576 | 高 2.0 倍 |
| Modal | $0.250 | $1.00 | $1.25 | unknown | 1,048,576 | 高 1.7 倍 |
| Reka | $0.160 | $1.10 | $1.26 | unknown | 1,048,576 | 高 1.1 倍 |
| PrimeIntellect | $0.156 | $1.20 | $1.36 | unknown | 1,048,576 | 高 1.0 倍 |
| Phala | $0.200 | $1.20 | $1.40 | unknown | 1,048,576 | 高 1.3 倍 |
| Crusoe | $0.290 | $1.20 | $1.49 | fp8 | 1,048,576 | 高 1.9 倍 |
| Near AI | $0.300 | $1.20 | $1.50 | unknown | 1,048,576 | 高 2.0 倍 |
| Venice | $0.300 | $1.20 | $1.50 | fp8 | 1,000,000 | 高 2.0 倍 |
| Mistral | $0.300 | $1.20 | $1.50 | fp8 | 1,048,576 | 高 2.0 倍 |
| Nebius | $0.300 | $1.20 | $1.50 | fp8 | 1,048,576 | 高 2.0 倍 |
| SiliconFlow | $0.300 | $1.20 | $1.50 | fp8 | 1,048,576 | 高 2.0 倍 |
| Together | $0.300 | $1.20 | $1.50 | unknown | 1,048,576 | 高 2.0 倍 |
| Fireworks | $0.300 | $1.20 | $1.50 | unknown | 1,048,576 | 高 2.0 倍 |
| BaseTen | $0.300 | $1.20 | $1.50 | unknown | 1,048,576 | 高 2.0 倍 |
| Parasail | $0.300 | $1.20 | $1.50 | fp8 | 1,048,576 | 高 2.0 倍 |
| Fireworks | $0.450 | $1.80 | $2.25 | unknown | 1,048,576 | 高 3.0 倍 |
最划算的三家
| # | 服务商 | 输入 | 输出 | 合计 | 精度 |
|---|---|---|---|---|---|
| 1 | Sail Research | $0.101 | $0.150 | $0.251 | FP4，精度有損，適合非關鍵任務 |
| 2 | Sail Research | $0.101 | $0.150 | $0.251 | FP4，精度有損，適合非關鍵任務 |
| 3 | Particle AI | $0.025 | $0.250 | $0.275 | 未標示，建議先小量測試 |
怎么选
只看价格会踩坑。建议按这个顺序判断：先看量化精度是否满足任务要求，再看上下文是否够用，最后才比价格。同一个模型在不同服务商之间可以随时切换，配置里换个 base_url 和 key 即可，不需要改代码逻辑。
官方 API 的优势是稳定性与文档完整，第三方服务商的优势是价格。对成本敏感的批量任务，用第三方通常能省一半以上；对稳定性要求高的线上服务，官方或大厂云服务更稳。