Comparativo de custo
API ou GPU própria?
Para modelos de pesos abertos, a resposta depende do volume. Abaixo de certa vazão a API ganha — você não paga nada enquanto está ocioso. Acima dela, o custo fixo da placa alugada se dilui.
155
O equilíbrio é aritmética: o custo por hora da placa mais barata que comporta o modelo, dividido pelo que os mesmos tokens custam na API. A mistura assume 3 tokens de entrada para cada 1 de saída, típico de conversa.
O que isto não diz é se a sua máquina alcança essa vazão. Tokens por hora dependem de quantização, contexto e lote — isso é medição, e não medimos.
| Modelo | API | Alugar ganha acima de | |||
|---|---|---|---|---|---|
| DeepSeek: DeepSeek V4.1 Flash 1017 | 763.2 | US$ 0,262 | B300 ×7 | 157 | |
| Reflection: Beam 1.0 501B A23B | 501.3 | US$ 0,438 | MI300X ×6 | 33 | |
| Moondream: Moondream 3 Preview | 9.3 | US$ 0,85 | Tesla P40 | 0.13 | |
| Moondream: Moondream 3.1 | 9.3 | US$ 0,85 | Tesla P40 | 0.13 | |
| Meta: Muse Spark Open | 822.0 | US$ 2,00 | B300 ×7 | 21 | |
| LiquidAI: LFM2.5-24B-A2B | 24.0 | US$ 0,213 | A100 SXM4 | 2 | |
| StepFun: Step 5 Preview | 604.3 | US$ 1,425 | MI300X ×7 | 12 | |
| Xiaomi: MiMo-V2.6-Flash | 159.4 | US$ 0,175 | Q RTX 8000 ×8 | 12 | |
| Xiaomi: MiMo-V2.6-Pro | 524.1 | US$ 0,544 | MI300X ×6 | 26 | |
| inclusionAI: Ling 3.0 Flash VL | 124.8 | US$ 0,09 | B300 | 66 | |
| DeepSeek: DeepSeek V4.1 Flash | 763.2 | US$ 0,525 | B300 ×7 | 79 | |
| DeepSeek: DeepSeek V4.1 Flash (batch) | 763.2 | US$ 0,075 | B300 ×7 | 551 | |
| Nex AGI: Nex-N2.5-Mini | 35.1 | US$ 0,044 | A100 SXM4 | 10 | |
| Nex AGI: Nex-N2.5-Pro | 396.8 | US$ 0,119 | MI300X ×5 | 101 | |
| IBM: Granite 4.2 8B | 8.8 | US$ 0,108 | Tesla P40 | 1 | |
| Tencent: Hy4 preview | 780.0 | US$ 1,126 | B300 ×7 | 37 | |
| Qwen: Qwen3.8 Flash | 180.0 | US$ 0,23 | A100 SXM4 ×5 | 10 | |
| Z.ai: GLM 5.3 Flash | 321.3 | US$ 0,054 | RTX PRO 6000 WS ×8 | 150 | |
| Z.ai: GLM 5.3 Flash (batch) | 321.3 | US$ 0,083 | RTX PRO 6000 WS ×8 | 97 | |
| DeepSeek: DeepSeek V4 Flash Vision Exp | 304.6 | US$ 0,323 | RTX PRO 6000 WS ×7 | 22 | |
| Tencent: Hy-MT2-1.8B | 2.0 | US$ 0,077 | Quadro P2000 | 0.35 | |
| Tencent: Hy-MT2-30B-A3B | 30.1 | US$ 0,129 | A100 SXM4 | 3 | |
| Tencent: Hy-MT2-7B | 8.0 | US$ 0,129 | Tesla P40 | 0.83 | |
| Z.ai: GLM 5.3 | 753.3 | US$ 1,075 | B300 ×6 | 33 | |
| Z.ai: GLM 5.3 (batch) | 753.3 | US$ 0,838 | B300 ×6 | 42 | |
| Qwen: Qwen3.8 27B | 27.8 | US$ 0,368 | A100 SXM4 | 1 | |
| Qwen: Qwen3.8 2.4T A95B | 2446.2 | US$ 3,00 | |||
| DeepSeek: DeepSeek V4 Pro 0813 | 1650.5 | US$ 0,99 | |||
| NVIDIA: Nemotron 3.5 Lightning | 31.6 | US$ 0,103 | A100 SXM4 | 4 | |
| Meta: Muse Glimmer 30B | 29.8 | US$ 0,525 | A100 SXM4 | 0.84 | |
| DeepSeek: DeepSeek V4 Flash 0731 | 304.2 | US$ 0,66 | RTX PRO 6000 WS ×7 | 11 | |
| Thinking Machines: Inkling Small | 266.0 | US$ 0,638 | A100 SXM4 ×8 | 6 | |
| inclusionAI: Ling 3.0 Flash | 127.5 | US$ 0,032 | Q RTX 8000 ×6 | 49 | |
| Poolside: Laguna S 2.1 | 117.6 | US$ 0,113 | B300 | 52 | |
| Meituan: LongCat 2.0 | 1775.6 | US$ 0,525 | |||
| Thinking Machines: Inkling | 952.4 | US$ 1,763 | B300 ×8 | 27 | |
| MoonshotAI: Kimi K3 | 2779.9 | US$ 3,163 | |||
| MoonshotAI: Kimi K3 (batch) | 2779.9 | US$ 4,56 | |||
| Tencent: Hy3 | 298.8 | US$ 0,144 | RTX PRO 6000 WS ×7 | 49 | |
| Poolside: Laguna XS 2.1 | 33.4 | US$ 0,075 | A100 SXM4 | 6 | |
| Z.ai: GLM 5.2 | 753.3 | US$ 1,776 | B300 ×6 | 20 | |
| MoonshotAI: Kimi K2.7 Code | 1026.9 | US$ 1,341 | |||
| NVIDIA: Nemotron 3.5 Content Safety | 4.3 | US$ 0,20 | Titan Xp | 0.20 | |
| NVIDIA: Nemotron 3 Ultra | 560.5 | US$ 1,05 | MI300X ×7 | 16 | |
| MiniMax: MiniMax M3 | 427.0 | US$ 0,525 | MI300X ×5 | 23 | |
| StepFun: Step 3.7 Flash | 201.4 | US$ 0,438 | A100 SXM4 ×6 | 6 | |
| Qwen: Qwen3.6 35B A3B | 36.0 | US$ 0,313 | A100 SXM4 | 1 | |
| Qwen: Qwen3.6 27B | 27.8 | US$ 0,688 | A100 SXM4 | 0.64 | |
| DeepSeek: DeepSeek V4 Pro 0423 | 1598.8 | US$ 1,194 | |||
| DeepSeek: DeepSeek V4 Flash 0423 | 290.9 | US$ 0,114 | A100 SXM4 ×8 | 31 | |
| Tencent: Hy3 preview | 298.8 | US$ 0,285 | RTX PRO 6000 WS ×7 | 25 | |
| Xiaomi: MiMo-V2.5-Pro | 1023.2 | US$ 0,381 | |||
| Xiaomi: MiMo-V2.5 | 310.8 | US$ 0,149 | RTX PRO 6000 WS ×7 | 47 | |
| MoonshotAI: Kimi K2.6 | 1026.9 | US$ 1,34 | |||
| Z.ai: GLM 5.1 | 753.9 | US$ 2,15 | B300 ×7 | 19 | |
| Google: Gemma 4 26B A4B | 25.8 | US$ 0,121 | A100 SXM4 | 4 | |
| Google: Gemma 4 31B | 31.3 | US$ 0,25 | A100 SXM4 | 2 | |
| Arcee AI: Trinity Large Thinking | 398.6 | US$ 0,388 | MI300X ×5 | 31 | |
| Reka Edge | 7.1 | US$ 0,10 | Quadro P5000 | 0.68 | |
| MiniMax: MiniMax M2.7 | 228.7 | US$ 0,525 | A100 SXM4 ×7 | 6 | |
| Mistral: Mistral Small 4 | 119.4 | US$ 0,262 | B300 | 22 | |
| Mistral: Mistral Small 4 (batch) | 119.4 | US$ 0,131 | B300 | 45 | |
| NVIDIA: Nemotron 3 Super | 123.6 | US$ 0,825 | B300 | 7 | |
| Qwen: Qwen3.5-9B | 9.7 | US$ 0,093 | Tesla P40 | 1 | |
| Qwen: Qwen3.5-35B-A3B | 36.0 | US$ 0,355 | A100 SXM4 | 1 | |
| Qwen: Qwen3.5-27B | 27.8 | US$ 0,743 | A100 SXM4 | 0.60 | |
| Qwen: Qwen3.5-122B-A10B | 125.1 | US$ 0,715 | B300 | 8 | |
| Qwen: Qwen3.5 397B A17B | 403.4 | US$ 1,275 | MI300X ×5 | 9 | |
| MiniMax: MiniMax M2.5 | 228.7 | US$ 0,44 | A100 SXM4 ×7 | 7 | |
| Z.ai: GLM 5 | 753.9 | US$ 1,55 | B300 ×7 | 27 | |
| Qwen: Qwen3 Coder Next | 79.7 | US$ 0,29 | MI300X | 8 | |
| StepFun: Step 3.5 Flash | 199.4 | US$ 0,15 | A100 SXM4 ×6 | 18 | |
| MoonshotAI: Kimi K2.5 | 1026.9 | US$ 1,00 | |||
| Z.ai: GLM 4.7 Flash | 31.2 | US$ 0,145 | A100 SXM4 | 3 | |
| MiniMax: MiniMax M2.1 | 228.7 | US$ 0,525 | A100 SXM4 ×7 | 6 | |
| Z.ai: GLM 4.7 | 358.3 | US$ 0,90 | RTX PRO 6000 WS ×8 | 9 | |
| NVIDIA: Nemotron 3 Nano 30B A3B | 31.6 | US$ 0,105 | A100 SXM4 | 4 | |
| Mistral: Devstral 2 2512 | 125.0 | US$ 0,80 | B300 | 7 | |
| Z.ai: GLM 4.6V | 107.7 | US$ 0,45 | B300 | 13 | |
| Mistral: Ministral 3 14B 2512 | 13.9 | US$ 0,20 | Q RTX 8000 | 1 | |
| Mistral: Ministral 3 8B 2512 | 8.9 | US$ 0,15 | Tesla P40 | 0.72 | |
| Mistral: Ministral 3 8B 2512 (batch) | 8.9 | US$ 0,075 | Tesla P40 | 1 | |
| Mistral: Ministral 3 3B 2512 | 3.8 | US$ 0,10 | Titan Xp | 0.41 | |
| DeepSeek: DeepSeek V3.2 | 685.4 | US$ 0,315 | MI300X ×8 | 61 | |
| MoonshotAI: Kimi K2 Thinking | 1026.4 | US$ 1,075 | |||
| Mistral: Voxtral Small 24B 2507 | 24.3 | US$ 0,15 | A100 SXM4 | 3 | |
| OpenAI: gpt-oss-safeguard-20b | 21.5 | US$ 0,131 | Q RTX 8000 | 2 | |
| MiniMax: MiniMax M2 | 228.7 | US$ 0,446 | A100 SXM4 ×7 | 7 | |
| IBM: Granite 4.0 Micro | 3.2 | US$ 0,041 | Tesla P4 | 0.87 | |
| Qwen: Qwen3 VL 8B Instruct | 8.8 | US$ 0,375 | Tesla P40 | 0.29 | |
| Qwen: Qwen3 VL 30B A3B Thinking | 31.1 | US$ 0,467 | A100 SXM4 | 0.95 | |
| Qwen: Qwen3 VL 30B A3B Instruct | 31.1 | US$ 0,262 | A100 SXM4 | 2 | |
| Z.ai: GLM 4.6 | 356.8 | US$ 0,963 | RTX PRO 6000 WS ×8 | 8 | |
| DeepSeek: DeepSeek V3.2 Exp | 685.4 | US$ 0,305 | MI300X ×8 | 63 | |
| TheDrummer: Cydonia 24B V4.1 | 23.6 | US$ 0,35 | A100 SXM4 | 1 | |
| Qwen: Qwen3 VL 235B A22B Instruct | 235.7 | US$ 0,633 | A100 SXM4 ×7 | 5 | |
| DeepSeek: DeepSeek V3.1 Terminus | 684.5 | US$ 0,453 | MI300X ×8 | 42 | |
| Qwen: Qwen3 Next 80B A3B Thinking | 81.3 | US$ 0,413 | MI300X | 6 | |
| Qwen: Qwen3 Next 80B A3B Instruct | 81.3 | US$ 0,343 | MI300X | 7 | |
| MoonshotAI: Kimi K2 0905 | 1026.5 | US$ 1,075 | |||
| Nous: Hermes 4 405B | 405.9 | US$ 1,50 | MI300X ×5 | 8 | |
| DeepSeek: DeepSeek V3.1 | 684.5 | US$ 0,425 | MI300X ×8 | 45 | |
| Z.ai: GLM 4.5V | 107.7 | US$ 0,90 | B300 | 7 | |
| OpenAI: gpt-oss-120b | 116.8 | US$ 0,262 | B300 | 22 | |
| OpenAI: gpt-oss-120b (batch) | 116.8 | US$ 0,056 | B300 | 105 | |
| OpenAI: gpt-oss-20b | 20.9 | US$ 0,036 | Q RTX 8000 | 7 | |
| OpenAI: gpt-oss-20b (batch) | 20.9 | US$ 0,046 | Q RTX 8000 | 6 | |
| Qwen: Qwen3 Coder 30B A3B Instruct | 30.5 | US$ 0,123 | A100 SXM4 | 4 | |
| Qwen: Qwen3 30B A3B Instruct 2507 | 30.5 | US$ 0,143 | A100 SXM4 | 3 | |
| Z.ai: GLM 4.5 | 358.3 | US$ 1,00 | RTX PRO 6000 WS ×8 | 8 | |
| Z.ai: GLM 4.5 Air | 110.5 | US$ 0,31 | B300 | 19 | |
| Qwen: Qwen3 235B A22B Thinking 2507 | 235.1 | US$ 1,213 | A100 SXM4 ×7 | 3 | |
| Qwen: Qwen3 Coder 480B A35B | 480.2 | US$ 0,475 | MI300X ×6 | 30 | |
| ByteDance: UI-TARS 7B | 8.3 | US$ 0,125 | Tesla P40 | 0.86 | |
| Qwen: Qwen3 235B A22B Instruct 2507 | 235.1 | US$ 0,205 | A100 SXM4 ×7 | 15 | |
| MoonshotAI: Kimi K2 0711 | 1026.4 | US$ 1,003 | |||
| Venice: Uncensored | 24.0 | US$ 0,375 | A100 SXM4 | 1 | |
| Tencent: Hunyuan A13B Instruct | 80.4 | US$ 0,248 | MI300X | 10 | |
| Mistral: Mistral Small 3.2 24B | 24.0 | US$ 0,133 | A100 SXM4 | 3 | |
| DeepSeek: R1 0528 | 684.5 | US$ 0,913 | MI300X ×8 | 21 | |
| Meta: Llama Guard 4 12B | 12.0 | US$ 0,18 | Q RTX 8000 | 1 | |
| Qwen: Qwen3 30B A3B | 30.5 | US$ 0,215 | A100 SXM4 | 2 | |
| Qwen: Qwen3 14B | 14.8 | US$ 0,15 | Q RTX 8000 | 2 | |
| Qwen: Qwen3 32B | 32.8 | US$ 0,13 | A100 SXM4 | 3 | |
| Meta: Llama 4 Maverick | 401.6 | US$ 0,304 | MI300X ×5 | 39 | |
| Meta: Llama 4 Scout | 108.6 | US$ 0,15 | B300 | 39 | |
| DeepSeek: DeepSeek V3 0324 | 684.5 | US$ 0,503 | MI300X ×8 | 38 | |
| Mistral: Mistral Small 3.1 24B | 24.0 | US$ 0,402 | A100 SXM4 | 1 | |
| Google: Gemma 3 4B | 4.3 | US$ 0,042 | Titan Xp | 0.96 | |
| Google: Gemma 3 12B | 12.2 | US$ 0,075 | Q RTX 8000 | 3 | |
| Cohere: Command A | 111.1 | US$ 4,375 | B300 | 1 | |
| Reka Flash 3 | 20.9 | US$ 0,125 | Q RTX 8000 | 2 | |
| Google: Gemma 3 27B | 27.4 | US$ 0,148 | A100 SXM4 | 3 | |
| TheDrummer: Skyfall 36B V2 | 36.9 | US$ 0,613 | A100 SXM4 | 0.72 | |
| Qwen: Qwen2.5 VL 72B Instruct | 73.4 | US$ 0,85 | MI300X | 3 | |
| Mistral: Mistral Small 3 | 23.6 | US$ 0,058 | A100 SXM4 | 8 | |
| MiniMax: MiniMax-01 | 456.1 | US$ 0,425 | MI300X ×6 | 34 | |
| Microsoft: Phi 4 | 14.7 | US$ 0,088 | Q RTX 8000 | 3 | |
| DeepSeek: DeepSeek V3 | 684.5 | US$ 0,45 | MI300X ×8 | 42 | |
| Meta: Llama 3.3 70B Instruct | 70.6 | US$ 0,155 | MI300X | 15 | |
| Qwen2.5 Coder 32B Instruct | 32.8 | US$ 0,745 | A100 SXM4 | 0.59 | |
| TheDrummer: UnslopNemo 12B | 12.2 | US$ 0,40 | Q RTX 8000 | 0.64 | |
| Magnum v4 72B | 72.7 | US$ 3,125 | MI300X | 0.76 | |
| Qwen: Qwen2.5 7B Instruct | 7.6 | US$ 0,125 | Tesla P40 | 0.86 | |
| Meta: Llama 3.2 1B Instruct | 1.2 | US$ 0,071 | Quadro P2000 | 0.39 | |
| Meta: Llama 3.2 3B Instruct | 3.2 | US$ 0,122 | Tesla P4 | 0.29 | |
| Qwen2.5 72B Instruct | 72.7 | US$ 0,37 | MI300X | 6 | |
| Nous: Hermes 3 70B Instruct | 70.6 | US$ 0,70 | MI300X | 3 | |
| Nous: Hermes 3 405B Instruct | 405.9 | US$ 1,00 | MI300X ×5 | 12 | |
| Sao10K: Llama 3 8B Lunaris | 8.0 | US$ 0,042 | Tesla P40 | 3 | |
| Meta: Llama 3.1 70B Instruct | 70.6 | US$ 0,40 | MI300X | 6 | |
| Meta: Llama 3.1 8B Instruct | 8.0 | US$ 0,058 | Tesla P40 | 2 | |
| Mistral: Mistral Nemo | 12.2 | US$ 0,029 | Q RTX 8000 | 9 | |
| Google: Gemma 2 27B | 27.2 | US$ 0,65 | A100 SXM4 | 0.68 | |
| Mistral: Mixtral 8x22B Instruct | 140.6 | US$ 3,00 | A100 SXM4 ×4 | 0.59 |