Updated weekly · checked
Cheapest DeepSeek V4.1 Flash, per task
36 providers ranked by what a real agent task costs, not by input price. Low-bit quants and shaky uptime are flagged.
Fastest live: baseten at 104 tok/s, 4841 ms to first token (Hugging Face router, $0.3 in / $1.2 out).
Price-skew trap: the lowest input price is OpenInference ($0.011 in), but with output at $1.8 it costs $18.34 per 1,000 tasks, vs $4.53 at Sail Research.
| Provider | Per 1,000 tasks | Blended $/1M | $/1M in / out | Quant | Uptime 24h |
|---|---|---|---|---|---|
| Sail Researchsail-research/fp4/us | $4.53 | $0.1133 | $0.101 / $0.15 | fp4 | 100.0% |
| Sail Researchsail-research/fp4 | $4.53 | $0.1133 | $0.101 / $0.15 | fp4 | 100.0% |
| DeepInfradeepinfra/fp8/flex | $4.70 | $0.1176 | $0.078 / $0.235 | fp8 | 63.9% |
| Morphmorph/fp8 | $5.15 | $0.1288 | $0.044 / $0.383 | fp8 | 100.0% |
| DeepInfradeepinfra/fp8 | $5.88 | $0.147 | $0.098 / $0.294 | fp8 | 100.0% |
| Rekareka | $6.20 | $0.155 | $0.04 / $0.5 | unknown | 96.5% |
| Waferwafer/mxfp4 | $7.10 | $0.1775 | $0.07 / $0.5 | mxfp4 | 100.0% |
| Relacerelace | $7.10 | $0.1775 | $0.07 / $0.5 | unknown | 100.0% |
| InferenceNetinference-net/us | $7.90 | $0.1975 | $0.08 / $0.55 | fp8 | 100.0% |
| InferenceNetinference-net/fp8 | $7.90 | $0.1975 | $0.08 / $0.55 | fp8 | 100.0% |
| DeepInfradeepinfra/fp8/priority | $8.82 | $0.2205 | $0.147 / $0.441 | fp8 | 100.0% |
| Ionstreamionstream | $9.00 | $0.225 | $0.15 / $0.45 | unknown | 99.9% |
| Cloudflarecloudflare | $9.00 | $0.225 | $0.1 / $0.6 | unknown | 99.4% |
| Decartdecart/fp4 | $9.03 | $0.2258 | $0.129 / $0.516 | fp4 | 99.7% |
| AtlasCloudatlas-cloud/fp8 | $9.45 | $0.2363 | $0.135 / $0.54 | fp8 | 100.0% |
| StreamLakestreamlake/fp8 | $10.50 | $0.2625 | $0.15 / $0.6 | fp8 | 100.0% |
| DeepSeekdeepseek | $10.50 | $0.2625 | $0.15 / $0.6 | unknown | 100.0% |
| DigitalOceandigitalocean | $12.60 | $0.315 | $0.18 / $0.72 | unknown | 100.0% |
| GMICloudgmicloud/fp8 | $12.60 | $0.315 | $0.18 / $0.72 | fp8 | 100.0% |
| Novitanovita/fp8 | $13.65 | $0.3412 | $0.195 / $0.78 | fp8 | 100.0% |
| Fireworksfireworks/spot | $14.70 | $0.3675 | $0.21 / $0.84 | unknown | 100.0% |
| Makoramakora/fp8 | $16.50 | $0.4125 | $0.15 / $1.2 | fp8 | 99.9% |
| Modalmodal | $17.50 | $0.4375 | $0.25 / $1 | unknown | 99.4% |
| Phalaphala | $18.00 | $0.45 | $0.2 / $1.2 | unknown | 99.9% |
| OpenInferenceopen-inference/fp4 | $18.34 | $0.4586 | $0.011 / $1.8 | fp4 | 99.9% |
| Baidubaidu/fp8 | $19.70 | $0.4925 | $0.28 / $1.13 | fp8 | 100.0% |
| Crusoecrusoe/fp8 | $20.70 | $0.5175 | $0.29 / $1.2 | fp8 | 100.0% |
| Nebiusnebius/fp8 | $21.00 | $0.525 | $0.3 / $1.2 | fp8 | 99.4% |
| SiliconFlowsiliconflow/fp8 | $21.00 | $0.525 | $0.3 / $1.2 | fp8 | 99.7% |
| Togethertogether | $21.00 | $0.525 | $0.3 / $1.2 | unknown | 100.0% |
| Fireworksfireworks | $21.00 | $0.525 | $0.3 / $1.2 | unknown | 89.8% |
| PrimeIntellectprimeintellect | $21.00 | $0.525 | $0.3 / $1.2 | unknown | 100.0% |
| BaseTenbaseten | $21.00 | $0.525 | $0.3 / $1.2 | unknown | 100.0% |
| Parasailparasail/fp8 | $21.00 | $0.525 | $0.3 / $1.2 | fp8 | 99.9% |
| Alibabaalibaba | $23.10 | $0.5775 | $0.33 / $1.32 | unknown | 99.5% |
| Fireworksfireworks/us | $31.50 | $0.7875 | $0.45 / $1.8 | unknown | 100.0% |
One task = 30,000 input + 10,000 output tokens (3:1, a typical agent step with context). Blended = (3 × input + output) ÷ 4. Amber = low-bit quantization (fp4/int4, can change outputs) or under 95% uptime in the last 24 hours. Prices are OpenRouter's per-provider list prices; cache discounts not included.
Source: OpenRouter provider endpoints · Hugging Face router · checked
FAQ
What's the cheapest way to run DeepSeek V4.1 Flash?
On 2026-10-28, Sail Research at $0.101 per 1M input and $0.15 per 1M output tokens (fp4), about $4.53 per 1,000 agent tasks of 30,000 in + 10,000 out tokens.
Why cost per task and not input price?
Agents send more input than output, but output tokens usually cost several times more. A provider with the lowest input price can still be the more expensive one per task, so we rank by a 3:1 blend.
Is the cheapest provider the same model?
Same weights, but not always the same precision. Check the quant column: Sail Research serves fp4. Lower-bit quantization is cheaper and can change outputs.
How big is the price spread?
36 providers on OpenRouter, from $0.1133 to $0.7875 per 1M blended tokens.
Which provider is fastest?
baseten: 104 tokens per second on the Hugging Face router's live data, 2026-10-28.
Same data as the Agent Stack Index (JSON). More: Free LLM APIs · OpenRouter alternatives