Cost comparison
API or your own GPU?
For models with open weights, the answer depends on volume. Below a certain throughput the API wins — you pay nothing while idle. Above it, the rented card's fixed cost pays off.
155
Break-even is arithmetic: the hourly cost of the cheapest card that holds the model, divided by what the same tokens cost on the API. The blend assumes 3 tokens in for every 1 out, which is typical of chat.
What this does not tell you is whether your setup reaches that throughput. Tokens per hour depend on quantisation, context and batching — that is measurement, and we have not measured it.
| Model | API | Renting wins above | |||
|---|---|---|---|---|---|
| DeepSeek: DeepSeek V4.1 Flash 1017 | 763.2 | $0.262 | MI350X ×6 | 125 | |
| Reflection: Beam 1.0 501B A23B | 501.3 | $0.475 | MI300X ×6 | 30 | |
| Moondream: Moondream 3 Preview | 9.3 | $0.85 | RTX 3090 | 0.14 | |
| Moondream: Moondream 3.1 | 9.3 | $0.85 | RTX 3090 | 0.14 | |
| Meta: Muse Spark Open | 822.0 | $2.00 | MI350X ×6 | 16 | |
| LiquidAI: LFM2.5-24B-A2B | 24.0 | $0.213 | A800 PCIE | 2 | |
| StepFun: Step 5 Preview | 604.3 | $1.425 | MI300X ×7 | 12 | |
| Xiaomi: MiMo-V2.6-Flash | 159.4 | $0.175 | Q RTX 8000 ×8 | 12 | |
| Xiaomi: MiMo-V2.6-Pro | 524.1 | $0.544 | MI300X ×6 | 26 | |
| DeepSeek: DeepSeek V4.1 Flash | 763.2 | $0.072 | MI350X ×6 | 457 | |
| inclusionAI: Ling 3.0 Flash VL | 124.8 | $0.09 | MI350X | 61 | |
| DeepSeek: DeepSeek V4.1 Flash (batch) | 763.2 | $0.075 | MI350X ×6 | 439 | |
| Nex AGI: Nex-N2.5-Mini | 35.1 | $0.044 | A800 PCIE | 11 | |
| Nex AGI: Nex-N2.5-Pro | 396.8 | $0.119 | MI300X ×5 | 101 | |
| IBM: Granite 4.2 8B | 8.8 | $0.108 | RTX 3090 | 1 | |
| Tencent: Hy4 preview | 780.0 | $1.126 | MI350X ×6 | 29 | |
| Z.ai: GLM 5.3 Flash (batch) | 321.3 | $0.083 | RTX PRO 6000 Max-Q ×8 | 97 | |
| Qwen: Qwen3.8 Flash | 180.0 | $0.23 | A800 PCIE ×5 | 10 | |
| Z.ai: GLM 5.3 Flash | 321.3 | $0.065 | RTX PRO 6000 Max-Q ×8 | 123 | |
| DeepSeek: DeepSeek V4 Flash Vision Exp | 304.6 | $0.323 | RTX PRO 6000 Max-Q ×7 | 22 | |
| Tencent: Hy-MT2-30B-A3B | 30.1 | $0.129 | A800 PCIE | 4 | |
| Tencent: Hy-MT2-1.8B | 2.0 | $0.077 | Tesla P100 | 0.35 | |
| Tencent: Hy-MT2-7B | 8.0 | $0.129 | RTX 3090 | 0.95 | |
| Z.ai: GLM 5.3 (batch) | 753.3 | $0.838 | MI350X ×6 | 39 | |
| Z.ai: GLM 5.3 | 753.3 | $0.409 | MI350X ×6 | 80 | |
| Qwen: Qwen3.8 27B | 27.8 | $0.368 | A800 PCIE | 1 | |
| Qwen: Qwen3.8 2.4T A95B | 2446.2 | $3.00 | |||
| DeepSeek: DeepSeek V4 Pro 0813 | 1650.5 | $0.99 | |||
| NVIDIA: Nemotron 3.5 Lightning | 31.6 | $0.103 | A800 PCIE | 5 | |
| Meta: Muse Glimmer 30B | 29.8 | $0.525 | A800 PCIE | 0.89 | |
| DeepSeek: DeepSeek V4 Flash 0731 | 304.2 | $0.66 | RTX PRO 6000 Max-Q ×7 | 11 | |
| Thinking Machines: Inkling Small | 266.0 | $0.638 | A800 PCIE ×8 | 6 | |
| inclusionAI: Ling 3.0 Flash | 127.5 | $0.032 | MI350X | 174 | |
| Poolside: Laguna S 2.1 | 117.6 | $0.113 | MI350X | 49 | |
| Meituan: LongCat 2.0 | 1775.6 | $0.525 | |||
| Thinking Machines: Inkling | 952.4 | $1.763 | MI350X ×7 | 22 | |
| MoonshotAI: Kimi K3 (batch) | 2779.9 | $4.56 | |||
| MoonshotAI: Kimi K3 | 2779.9 | $3.163 | |||
| Tencent: Hy3 | 298.8 | $0.144 | RTX PRO 6000 Max-Q ×7 | 49 | |
| Poolside: Laguna XS 2.1 | 33.4 | $0.075 | A800 PCIE | 6 | |
| Z.ai: GLM 5.2 | 753.3 | $1.07 | MI350X ×6 | 31 | |
| MoonshotAI: Kimi K2.7 Code | 1026.9 | $1.341 | MI350X ×8 | 33 | |
| NVIDIA: Nemotron 3 Ultra | 560.5 | $1.05 | MI300X ×7 | 16 | |
| NVIDIA: Nemotron 3.5 Content Safety | 4.3 | $0.20 | Tesla P100 | 0.14 | |
| MiniMax: MiniMax M3 | 427.0 | $0.525 | MI300X ×5 | 23 | |
| StepFun: Step 3.7 Flash | 201.4 | $0.438 | A800 PCIE ×6 | 6 | |
| Qwen: Qwen3.6 35B A3B | 36.0 | $0.313 | A800 PCIE | 1 | |
| Qwen: Qwen3.6 27B | 27.8 | $0.688 | A800 PCIE | 0.68 | |
| DeepSeek: DeepSeek V4 Pro 0423 | 1598.8 | $1.194 | |||
| DeepSeek: DeepSeek V4 Flash 0423 | 290.9 | $0.114 | A800 PCIE ×8 | 33 | |
| Xiaomi: MiMo-V2.5 | 310.8 | $0.175 | RTX PRO 6000 Max-Q ×7 | 40 | |
| Tencent: Hy3 preview | 298.8 | $0.285 | RTX PRO 6000 Max-Q ×7 | 25 | |
| Xiaomi: MiMo-V2.5-Pro | 1023.2 | $0.544 | MI350X ×8 | 81 | |
| MoonshotAI: Kimi K2.6 | 1026.9 | $1.34 | MI350X ×8 | 33 | |
| Z.ai: GLM 5.1 | 753.9 | $1.483 | MI350X ×6 | 22 | |
| Google: Gemma 4 26B A4B | 25.8 | $0.121 | A800 PCIE | 4 | |
| Google: Gemma 4 31B | 31.3 | $0.25 | A800 PCIE | 2 | |
| Arcee AI: Trinity Large Thinking | 398.6 | $0.388 | MI300X ×5 | 31 | |
| Reka Edge | 7.1 | $0.10 | Tesla P100 | 0.27 | |
| MiniMax: MiniMax M2.7 | 228.7 | $0.525 | A800 PCIE ×7 | 6 | |
| Mistral: Mistral Small 4 | 119.4 | $0.262 | MI350X | 21 | |
| Mistral: Mistral Small 4 (batch) | 119.4 | $0.131 | MI350X | 42 | |
| NVIDIA: Nemotron 3 Super | 123.6 | $0.173 | MI350X | 32 | |
| Qwen: Qwen3.5-9B | 9.7 | $0.113 | RTX 3090 | 1 | |
| Qwen: Qwen3.5-27B | 27.8 | $0.743 | A800 PCIE | 0.63 | |
| Qwen: Qwen3.5-35B-A3B | 36.0 | $0.355 | A800 PCIE | 1 | |
| Qwen: Qwen3.5-122B-A10B | 125.1 | $0.715 | MI350X | 8 | |
| Qwen: Qwen3.5 397B A17B | 403.4 | $0.878 | MI300X ×5 | 14 | |
| MiniMax: MiniMax M2.5 | 228.7 | $0.44 | A800 PCIE ×7 | 7 | |
| Z.ai: GLM 5 | 753.9 | $1.55 | MI350X ×6 | 21 | |
| Qwen: Qwen3 Coder Next | 79.7 | $0.29 | MI300X | 8 | |
| StepFun: Step 3.5 Flash | 199.4 | $0.15 | A800 PCIE ×6 | 19 | |
| MoonshotAI: Kimi K2.5 | 1026.9 | $1.00 | MI350X ×8 | 44 | |
| Z.ai: GLM 4.7 Flash | 31.2 | $0.145 | A800 PCIE | 3 | |
| MiniMax: MiniMax M2.1 | 228.7 | $0.525 | A800 PCIE ×7 | 6 | |
| Z.ai: GLM 4.7 | 358.3 | $0.90 | RTX PRO 6000 Max-Q ×8 | 9 | |
| NVIDIA: Nemotron 3 Nano 30B A3B | 31.6 | $0.105 | A800 PCIE | 4 | |
| Mistral: Devstral 2 2512 | 125.0 | $0.80 | MI350X | 7 | |
| Z.ai: GLM 4.6V | 107.7 | $0.45 | MI350X | 12 | |
| Mistral: Ministral 3 14B 2512 | 13.9 | $0.20 | Q RTX 8000 | 1 | |
| Mistral: Ministral 3 8B 2512 | 8.9 | $0.15 | RTX 3090 | 0.81 | |
| Mistral: Ministral 3 8B 2512 (batch) | 8.9 | $0.075 | RTX 3090 | 2 | |
| Mistral: Ministral 3 3B 2512 | 3.8 | $0.10 | Tesla P100 | 0.27 | |
| DeepSeek: DeepSeek V3.2 | 685.4 | $0.315 | MI300X ×8 | 61 | |
| MoonshotAI: Kimi K2 Thinking | 1026.4 | $1.075 | MI350X ×8 | 41 | |
| Mistral: Voxtral Small 24B 2507 | 24.3 | $0.15 | A800 PCIE | 3 | |
| OpenAI: gpt-oss-safeguard-20b | 21.5 | $0.131 | RTX A6000 | 2 | |
| MiniMax: MiniMax M2 | 228.7 | $0.446 | A800 PCIE ×7 | 7 | |
| IBM: Granite 4.0 Micro | 3.2 | $0.041 | Tesla P100 | 0.67 | |
| Qwen: Qwen3 VL 8B Instruct | 8.8 | $0.375 | RTX 3090 | 0.33 | |
| Qwen: Qwen3 VL 30B A3B Thinking | 31.1 | $0.467 | A800 PCIE | 1.00 | |
| Qwen: Qwen3 VL 30B A3B Instruct | 31.1 | $0.262 | A800 PCIE | 2 | |
| Z.ai: GLM 4.6 | 356.8 | $0.963 | RTX PRO 6000 Max-Q ×8 | 8 | |
| DeepSeek: DeepSeek V3.2 Exp | 685.4 | $0.305 | MI300X ×8 | 63 | |
| TheDrummer: Cydonia 24B V4.1 | 23.6 | $0.35 | A800 PCIE | 1 | |
| Qwen: Qwen3 VL 235B A22B Instruct | 235.7 | $0.633 | A800 PCIE ×7 | 5 | |
| DeepSeek: DeepSeek V3.1 Terminus | 684.5 | $0.453 | MI300X ×8 | 42 | |
| Qwen: Qwen3 Next 80B A3B Thinking | 81.3 | $0.413 | MI300X | 6 | |
| Qwen: Qwen3 Next 80B A3B Instruct | 81.3 | $0.35 | MI300X | 7 | |
| MoonshotAI: Kimi K2 0905 | 1026.5 | $1.075 | MI350X ×8 | 41 | |
| Nous: Hermes 4 405B | 405.9 | $1.50 | MI300X ×5 | 8 | |
| DeepSeek: DeepSeek V3.1 | 684.5 | $0.425 | MI300X ×8 | 45 | |
| Z.ai: GLM 4.5V | 107.7 | $0.90 | MI350X | 6 | |
| OpenAI: gpt-oss-120b | 116.8 | $0.262 | MI350X | 21 | |
| OpenAI: gpt-oss-120b (batch) | 116.8 | $0.056 | MI350X | 98 | |
| OpenAI: gpt-oss-20b | 20.9 | $0.036 | Q RTX 8000 | 7 | |
| OpenAI: gpt-oss-20b (batch) | 20.9 | $0.046 | Q RTX 8000 | 6 | |
| Qwen: Qwen3 Coder 30B A3B Instruct | 30.5 | $0.123 | A800 PCIE | 4 | |
| Qwen: Qwen3 30B A3B Instruct 2507 | 30.5 | $0.143 | A800 PCIE | 3 | |
| Z.ai: GLM 4.5 | 358.3 | $1.00 | RTX PRO 6000 Max-Q ×8 | 8 | |
| Z.ai: GLM 4.5 Air | 110.5 | $0.31 | MI350X | 18 | |
| Qwen: Qwen3 235B A22B Thinking 2507 | 235.1 | $1.213 | A800 PCIE ×7 | 3 | |
| Qwen: Qwen3 Coder 480B A35B | 480.2 | $0.475 | MI300X ×6 | 30 | |
| ByteDance: UI-TARS 7B | 8.3 | $0.125 | RTX 3090 | 0.98 | |
| Qwen: Qwen3 235B A22B Instruct 2507 | 235.1 | $0.205 | A800 PCIE ×7 | 16 | |
| MoonshotAI: Kimi K2 0711 | 1026.4 | $1.003 | MI350X ×8 | 44 | |
| Venice: Uncensored | 24.0 | $0.375 | A800 PCIE | 1 | |
| Tencent: Hunyuan A13B Instruct | 80.4 | $0.248 | MI300X | 10 | |
| Mistral: Mistral Small 3.2 24B | 24.0 | $0.133 | A800 PCIE | 4 | |
| DeepSeek: R1 0528 | 684.5 | $0.913 | MI300X ×8 | 21 | |
| Meta: Llama Guard 4 12B | 12.0 | $0.18 | Q RTX 8000 | 1 | |
| Qwen: Qwen3 30B A3B | 30.5 | $0.215 | A800 PCIE | 2 | |
| Qwen: Qwen3 14B | 14.8 | $0.15 | Q RTX 8000 | 2 | |
| Qwen: Qwen3 32B | 32.8 | $0.13 | A800 PCIE | 4 | |
| Meta: Llama 4 Maverick | 401.6 | $0.304 | MI300X ×5 | 39 | |
| Meta: Llama 4 Scout | 108.6 | $0.15 | MI350X | 37 | |
| DeepSeek: DeepSeek V3 0324 | 684.5 | $0.503 | MI300X ×8 | 38 | |
| Mistral: Mistral Small 3.1 24B | 24.0 | $0.402 | A800 PCIE | 1 | |
| Google: Gemma 3 4B | 4.3 | $0.042 | Tesla P100 | 0.64 | |
| Google: Gemma 3 12B | 12.2 | $0.075 | Q RTX 8000 | 3 | |
| Cohere: Command A | 111.1 | $4.375 | MI350X | 1 | |
| Reka Flash 3 | 20.9 | $0.125 | Q RTX 8000 | 2 | |
| Google: Gemma 3 27B | 27.4 | $0.148 | A800 PCIE | 3 | |
| TheDrummer: Skyfall 36B V2 | 36.9 | $0.613 | A800 PCIE | 0.76 | |
| Qwen: Qwen2.5 VL 72B Instruct | 73.4 | $0.85 | MI300X | 3 | |
| Mistral: Mistral Small 3 | 23.6 | $0.058 | A800 PCIE | 8 | |
| MiniMax: MiniMax-01 | 456.1 | $0.425 | MI300X ×6 | 34 | |
| Microsoft: Phi 4 | 14.7 | $0.088 | Q RTX 8000 | 3 | |
| DeepSeek: DeepSeek V3 | 684.5 | $0.45 | MI300X ×8 | 42 | |
| Meta: Llama 3.3 70B Instruct | 70.6 | $0.155 | MI300X | 15 | |
| Qwen2.5 Coder 32B Instruct | 32.8 | $0.745 | A800 PCIE | 0.63 | |
| TheDrummer: UnslopNemo 12B | 12.2 | $0.40 | Q RTX 8000 | 0.63 | |
| Magnum v4 72B | 72.7 | $3.125 | MI300X | 0.76 | |
| Qwen: Qwen2.5 7B Instruct | 7.6 | $0.125 | RTX 3090 | 0.98 | |
| Meta: Llama 3.2 1B Instruct | 1.2 | $0.071 | Tesla P100 | 0.39 | |
| Meta: Llama 3.2 3B Instruct | 3.2 | $0.122 | Tesla P100 | 0.22 | |
| Qwen2.5 72B Instruct | 72.7 | $0.37 | MI300X | 6 | |
| Nous: Hermes 3 70B Instruct | 70.6 | $0.70 | MI300X | 3 | |
| Nous: Hermes 3 405B Instruct | 405.9 | $1.00 | MI300X ×5 | 12 | |
| Sao10K: Llama 3 8B Lunaris | 8.0 | $0.042 | RTX 3090 | 3 | |
| Meta: Llama 3.1 70B Instruct | 70.6 | $0.40 | MI300X | 6 | |
| Meta: Llama 3.1 8B Instruct | 8.0 | $0.058 | RTX 3090 | 2 | |
| Mistral: Mistral Nemo | 12.2 | $0.029 | Q RTX 8000 | 9 | |
| Google: Gemma 2 27B | 27.2 | $0.65 | A800 PCIE | 0.72 | |
| Mistral: Mixtral 8x22B Instruct | 140.6 | $3.00 | Q RTX 8000 ×7 | 0.59 |