Same model, wildly different price
An open-weight model is one artifact, but many providers host it — and they don't agree on price. Across 84 models served by two or more providers, the median gap between the cheapest and priciest host is77%; at the extreme it reaches1347%. Same weights, same output — you just pay more for picking the wrong door. Every figure is computed from published prices.
deepseek-v3.2 is served by 13 providers. Cheapest: $0.518/1M blended (GMICloud). Priciest: $7.5(SambaNova) — 1347% more.
+$4.89 wasted per 1,000 chatbot requests
…if you route to the priciest host instead of the cheapest, for identical output.
The spread, drawn
Each bar is one model, from its cheapest host (green dot) to its priciest (red dot), on a log price scale. The bar's length is money left on the table — same weights, same output at both ends.
Blended $/1M tokens, log scale — hover a bar for both hosts, click for the model's full provider table. Models served by 3+ providers.
Widest spreads
Models where the choice of provider matters most. "Waste" = extra cost per 1,000 chatbot requests (1k in / 500 out) if you pick the priciest host over the cheapest.
| Model | Providers | Cheapest | Priciest | Spread | Waste /1k |
|---|---|---|---|---|---|
| deepseek-v3.2 | 13 | $0.518 GMICloud | $7.5 SambaNova | 1347% | $4.89 |
| deepseek-v4-flash | 22 | $0.15 Baidu | $1.76 Cloudflare | 1073% | $1.00 |
| llama-3.1-8b | 5 | $0.06 DeepInfra | $0.44 CoreWeave | 633% | $0.29 |
| llama-3.3-70b | 12 | $0.42 DeepInfra | $2.546 Cloudflare | 506% | $1.16 |
| gpt-5.6-sol | 3 | $6 OpenAI | $35 Azure OpenAI | 483% | $16.50 |
| gemma-4-31b | 12 | $0.44 CoreWeave | $2.48 Cerebras | 464% | $1.47 |
| gpt-oss-120b | 16 | $0.2 AkashML | $1.1 Cerebras | 450% | $0.61 |
| qwen3-coder-30b-a3b | 4 | $0.34 Novita AI | $1.755 Alibaba | 416% | $0.82 |
| qwen3-coder | 5 | $1.3 DeepInfra | $5.85 Alibaba | 350% | $2.61 |
| mistral-nemo | 4 | $0.049 DeepInfra | $0.21 Novita AI | 329% | $0.09 |
| qwen3-14b | 2 | $0.36 DeepInfra | $1.137 Alibaba | 216% | $0.44 |
| gpt-oss-20b | 12 | $0.12 AkashML | $0.375 Groq | 212% | $0.16 |
| minimax-m3 | 9 | $1.19 CoreWeave | $3 GMICloud | 152% | $1.09 |
| qwen3-235b-a22b | 9 | $0.438 GMICloud | $1.1 Google | 151% | $0.40 |
| deepseek-v4 | 18 | $2.323 Alibaba | $5.81 Phala | 150% | $2.18 |
| minimax-m2.5 | 7 | $1.22 Venice | $3 Minimax | 146% | $1.06 |
| kimi-k2.7-code | 11 | $4.08 DeepInfra | $9.9 Moonshot AI | 143% | $3.52 |
| llama-4-scout | 3 | $0.4 DeepInfra | $0.95 Google | 137% | $0.35 |
| minimax-m2.7 | 7 | $1.35 Novita AI | $3 SambaNova | 122% | $0.99 |
| gemma-3-27b | 4 | $0.24 DeepInfra | $0.53 Parasail | 121% | $0.14 |
Blended = input + output per 1M tokens (USD), from each provider's published rates. Verified 2026-09-05.
FAQ
Does the same open model cost the same on every provider?
No. Across 84 models hosted by two or more providers, the median price gap between the cheapest and priciest host is 77%, reaching 1347% at the extreme. Same weights, same output — the difference is purely which provider you route to.
How much can picking the wrong provider cost?
For deepseek-v3.2, served by 13 providers, the priciest host is 1347% dearer than the cheapest — about $4.89 extra per 1,000 chatbot requests, for identical output.
Why do prices differ if the model weights are identical?
Providers differ on hardware, margins, batching, and serving/quantization choices — not on the model itself. So for open-weight models it pays to compare hosts, because the quality you get is the same.