💰
LLM Economics
LLM Economics
Compare hardware, models and inference setups by operating cost, token output and total cost of ownership.
💸 Operating cost🖥️ TCO⚡ Tokens / euro🔋 Tokens / kWh
Settings are stored in your browser and shared via the URL. Power draw estimated from GPU TDP + CPU (idle + 15 %) + board. Methodology →
Model economics
67 Models
Most economical setup per model. No quality judgement – only measured cost per token. Number after the model = runs · HW configs.
| # | Model | Gen tok/s i | EUR / 1M tok i | Tok/€ | Tok/kWh | cheapest HW |
|---|---|---|---|---|---|---|
| 1 | Nemotron-3-Nano-4B4BiNVIDIA ↗ Source | 579.8 | EUR 0.022 | 46.33M | 13.90M | NVIDIA GeForce RTX 5070 Ti |
| 2 | gemma-4-E4B-it8BiGoogle ↗ Source | 436.1 | EUR 0.024 | 41.90M | 12.57M | NVIDIA GeForce RTX 5070 Ti |
| 3 | gemma-4-26B-A4B-it-assistant26BiGoogle ↗ Source | 58.3 | EUR 0.13 | 7.45M | 2.24M | AMD Radeon 8060S Graphics |
| 4 | Gemma-4-26B-A4B-it26BiGoogle ↗ Source | 388.8 | EUR 0.026 | 39.06M | 11.72M | NVIDIA GeForce RTX 5070 Ti |
| 5 | Qwen2.5-3B-Instruct3BiQwen (Alibaba) ↗ Source | 170.3 | EUR 0.12 | 8.69M | 2.61M | 2× NVIDIA GeForce RTX 2060 |
| 6 | Gemma-4-26B-A4B26BiGoogle ↗ Source | 155.4 | EUR 0.12 | 8.54M | 2.56M | NVIDIA RTX A6000 |
| 7 | Llama-3.2-3B-Instruct3BiMeta ↗ Source | 157.6 | EUR 0.13 | 7.58M | 2.27M | 2× NVIDIA GeForce RTX 2060 |
| 8 | Ornith-1.5-35B-A3B35BiDeepReinforce ↗ Source | 331.3 | EUR 0.089 | 11.20M | 3.36M | 2× NVIDIA RTX A6000 |
| 9 | glm-4.7-flashiZ.ai (Zhipu) ↗ Source | 48.9 | EUR 0.24 | 4.19M | 1.26M | NVIDIA GB10 (DGX Spark) |
| 10 | Ornith-1.0-9B9BiDeepReinforce ↗ Source | 330.0 | EUR 0.029 | 34.20M | 10.26M | NVIDIA GeForce RTX 5070 Ti |
| 11 | Meta-Llama-3.1-8B-Instruct8BiMeta ↗ Source | 335.8 | EUR 0.028 | 35.36M | 10.61M | NVIDIA GeForce RTX 5070 Ti |
| 12 | Laguna-XS-2.1iPoolside ↗ Source | 607.8 | EUR 0.029 | 34.08M | 10.22M | NVIDIA GeForce RTX 5090 |
| 13 | Nemotron-Cascade-2-30B-A3B30BiNVIDIA ↗ Source | 380.8 | EUR 0.030 | 33.19M | 9.96M | NVIDIA GeForce RTX 5090 |
| 14 | Qwen2.5-7B-Instruct7BiQwen (Alibaba) ↗ Source | 108.7 | EUR 0.19 | 5.36M | 1.61M | 2× NVIDIA GeForce RTX 2060 |
| 15 | Ornith-1.5-9B9BiDeepReinforce ↗ Source | 246.6 | EUR 0.12 | 8.44M | 2.53M | 2× NVIDIA RTX A6000 |
| 16 | gemma-4-12B-it12BiGoogle ↗ Source | 262.8 | EUR 0.038 | 26.23M | 7.87M | NVIDIA GeForce RTX 5070 Ti |
| 17 | gemma-4-E2B-it5BiGoogle ↗ Source | 517.5 | EUR 0.013 | 77.43M | 23.23M | NVIDIA GB10 (DGX Spark) |
| 18 | Nemotron-3-Nano-Omni-30B-A3B-Reasoning30BiNVIDIA ↗ Source | 407.2 | EUR 0.030 | 33.39M | 10.02M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 19 | North-Mini-Code-1.0iCohere ↗ Source | 401.4 | EUR 0.033 | 30.22M | 9.07M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 20 | Ministral-3-14B-Reasoning-251214BiMistral AI ↗ Source | 247.5 | EUR 0.044 | 22.84M | 6.85M | NVIDIA GeForce RTX 5070 Ti |
| 21 | Qwen3-Omni-30B-A3B-Thinking30BiQwen (Alibaba) ↗ Source | 383.1 | EUR 0.031 | 32.31M | 9.69M | NVIDIA GeForce RTX 5090 |
| 22 | gpt-oss-20b20BiOpenAI ↗ Source | 445.7 | EUR 0.021 | 46.85M | 14.05M | NVIDIA GeForce RTX 5070 Ti |
| 23 | Qwen3.6-35B-A3B35BiQwen (Alibaba) ↗ Source | 254.6 | EUR 0.036 | 27.76M | 8.33M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 24 | Qwen3-Coder-30B-A3B-Instruct30BiQwen (Alibaba) ↗ Source | 334.2 | EUR 0.030 | 33.14M | 9.94M | NVIDIA GeForce RTX 5090 |
| 25 | Qwen3-30B-A3B-Thinking-250730BiQwen (Alibaba) ↗ Source | 338.6 | EUR 0.031 | 32.51M | 9.75M | NVIDIA GeForce RTX 5090 |
| 26 | Qwen3-30B-A3B-Instruct-250730BiQwen (Alibaba) ↗ Source | 317.0 | EUR 0.031 | 32.26M | 9.68M | NVIDIA GeForce RTX 5090 |
| 27 | Qwen-AgentWorld-35B-A3B35BiQwen (Alibaba) ↗ Source | 341.7 | EUR 0.036 | 27.79M | 8.34M | NVIDIA GeForce RTX 5090 |
| 28 | Qwen3-VL-30B-A3B-Instruct30BiQwen (Alibaba) ↗ Source | 334.3 | EUR 0.031 | 31.97M | 9.59M | NVIDIA GeForce RTX 5090 |
| 29 | Nemotron-3-Nano-30B-A3B30BiNVIDIA ↗ Source | 359.6 | EUR 0.030 | 33.25M | 9.97M | NVIDIA GeForce RTX 5090 |
| 30 | Ornith-1.0-35B35BiDeepReinforce ↗ Source | 446.8 | EUR 0.033 | 30.06M | 9.02M | NVIDIA GeForce RTX 5090 |
| 31 | Qwen3.5-35B-A3B35BiQwen (Alibaba) ↗ Source | 327.9 | EUR 0.035 | 28.34M | 8.50M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 32 | Nemotron-3.5-Lightning-30B-A3B30BiNVIDIA ↗ Source | 186.2 | EUR 0.36 | 2.74M | 822.17K | 4× AMD Radeon AI PRO R9700 |
| 33 | Devstral-Small-2-24B-Instruct-251224BiMistral AI ↗ Source | 156.1 | EUR 0.074 | 13.53M | 4.06M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 34 | Devstral-Small-250724BiMistral AI ↗ Source | 150.3 | EUR 0.083 | 12.07M | 3.62M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 35 | Qwen3.6-27B27BiQwen (Alibaba) ↗ Source | 139.6 | EUR 0.097 | 10.28M | 3.08M | NVIDIA GeForce RTX 5090 |
| 36 | Codestral-22B-v0.122BiMistral AI ↗ Source | 129.0 | EUR 0.081 | 12.38M | 3.72M | NVIDIA GeForce RTX 5090 |
| 37 | Magistral-Small-250924BiMistral AI ↗ Source | 172.9 | EUR 0.070 | 14.23M | 4.27M | NVIDIA GeForce RTX 5090 |
| 38 | Ling-3.0-flash127.5BiinclusionAI ↗ Source | 85.8 | EUR 0.046 | 21.74M | 6.52M | NVIDIA GB10 (DGX Spark) |
| 39 | Qwen3-Coder-NextiQwen (Alibaba) ↗ Source | 292.0 | EUR 0.043 | 23.36M | 7.01M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 40 | gpt-oss-120b120BiOpenAI ↗ Source | 309.3 | EUR 0.040 | 24.90M | 7.47M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 41 | Laguna-S-2.1-INT4iPoolside ↗ Source | 287.6 | EUR 0.064 | 15.57M | 4.67M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 42 | gemma-4-31B-it31BiGoogle ↗ Source | 79.0 | EUR 0.22 | 4.64M | 1.39M | NVIDIA GeForce RTX 3090 Ti |
| 43 | Mistral-Small-3.1-24B-Instruct-250324BiMistral AI ↗ Source | 63.2 | EUR 0.17 | 5.92M | 1.77M | AMD Radeon PRO W7900 Dual Slot |
| 44 | Qwen3.8-27B27BiQwen (Alibaba) ↗ Source | 103.5 | EUR 0.13 | 7.63M | 2.29M | 2× NVIDIA RTX A6000 |
| 45 | OpenReasoning-Nemotron-32B32BiNVIDIA ↗ Source | 114.9 | EUR 0.11 | 9.29M | 2.79M | NVIDIA GeForce RTX 5090 |
| 46 | Mistral-Small-4-119B-2603119BiMistral AI ↗ Source | 93.1 | EUR 0.056 | 17.89M | 5.37M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 47 | Seed-OSS-36B-Instruct36BiByteDance Seed ↗ Source | 110.6 | EUR 0.062 | 16.21M | 4.86M | NVIDIA GeForce RTX 5090 |
| 48 | Qwen2.5-32B-Instruct-AWQ32BiQwen (Alibaba) ↗ Source | 139.2 | EUR 0.11 | 9.51M | 2.85M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 49 | Muse-Glimmer-30B30BiMeta ↗ Source | 77.4 | EUR 0.53 | 1.88M | 563.91K | 4× AMD Radeon AI PRO R9700 |
| 50 | minimax-m2-awq230BiMiniMax ↗ Source | 114.8 | EUR 0.15 | 6.69M | 2.01M | 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
| 51 | Qwen3.5-122B-A10B122BiQwen (Alibaba) ↗ Source | 91.1 | EUR 0.15 | 6.90M | 2.07M | 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
| 52 | Qwen2.5-Coder-32B-Instruct32BiQwen (Alibaba) ↗ Source | 97.6 | EUR 0.11 | 9.33M | 2.80M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 53 | GLM-4.5-Air106BiZ.ai (Zhipu) ↗ Source | 132.8 | EUR 0.081 | 12.34M | 3.70M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 54 | Llama-3.3-Nemotron-Super-49B-v1.549BiNVIDIA ↗ Source | 75.9 | EUR 0.15 | 6.79M | 2.04M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 55 | Nemotron-3-Super-120B-A12B120BiNVIDIA ↗ Source | 85.5 | EUR 0.089 | 11.20M | 3.36M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 56 | Nemotron-H-47B-Reasoning-128K47BiNVIDIA ↗ Source | 45.1 | EUR 0.26 | 3.81M | 1.14M | NVIDIA GeForce RTX 5090 |
| 57 | MiniMax-M2.5230BiMiniMax ↗ Source | 87.1 | EUR 0.11 | 9.05M | 2.71M | NVIDIA GB10 (DGX Spark) |
| 58 | DeepSeek-V4-Flash290.9BiDeepSeek ↗ Source | 15.4 | EUR 5.14 | 194.49K | 58.35K | 4× AMD Radeon AI PRO R9700 |
| 59 | DeepSeek-V4-Flash-284B-A13B284BiDeepSeek ↗ Source | 65.6 | EUR 0.21 | 4.71M | 1.41M | 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
| 60 | Laguna-M.1iPoolside ↗ Source | 39.7 | EUR 0.24 | 4.12M | 1.24M | 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
| 61 | Qwen2.5-72B-Instruct72BiQwen (Alibaba) ↗ Source | 45.2 | EUR 0.22 | 4.46M | 1.34M | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 62 | command-a-reasoning-08-2025111BiCohere ↗ Source | 40.8 | EUR 0.34 | 2.93M | 877.76K | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 63 | Mistral-Medium-3.5-128B128BiMistral AI ↗ Source | 10.2 | EUR 0.38 | 2.64M | 791.69K | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 64 | MiniMax-M2.7230BiMiniMax ↗ Source | 64.4 | EUR 0.067 | 14.93M | 4.48M | NVIDIA GB10 (DGX Spark) |
| 65 | Llama-3.1-Nemotron-Ultra-253B-v1253BiNVIDIA ↗ Source | 4.3 | EUR 1.32 | 758.23K | 227.47K | 3× NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition |
| 66 | Devstral-2-123B-Instruct-2512123BiMistral AI ↗ Source | 22.1 | EUR 0.38 | 2.67M | 799.97K | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
| 67 | glm-5.2-colibri744BiZ.ai (Zhipu) ↗ Source | 0.0 | EUR 1,788.19 | 559 | 168 | NVIDIA RTX PRO 6000 Blackwell Workstation Edition |
Low token costs do not automatically mean high model quality. For quality see the harness benchmarks.