单任务成本:智能体编程会话
一次长时间的工具调用循环:多轮共输入 60,000 个 token,输出 4,000 个 token。成本 = 输入 token 数 × 输入单价 + 输出 token 数 × 输出单价。“使用缓存”假定四分之三的输入按提供商的缓存读取价格从提示缓存读取。
截至 2026年11月16日,BenchLeader 在“智能体编程会话成本”榜上将 Gemma 4 E4B 排在第一($0.0016)。其后为 Qwen3.5-4B($0.0024)和 Llama 3.2 1B Instruct($0.0024)。
255 个模型 · 数据截至 2026年11月16日
- 1Gemma 4 E4B (thinking)$0.0016
- 2Qwen3.5-4B (thinking)$0.0024
- 3Llama 3.2 1B Instruct$0.0024
- 4Gemma 4 E2B (thinking)$0.0027
- 5Llama 3.1 8B$0.0027
- 6Nova Micro$0.0030
- 7Qwen2.5 7B Instruct$0.0032
- 8Mistral Small 3$0.0033
- 9Gemma 3 4B$0.0034
- 10Gemma 3 12B$0.0036
- 11Llama 3-8B$0.0036
- 12Nemotron 3 Nano 30B A3B (thinking)$0.0038
- 13DeepSeek-R1-0528-Qwen3-8B$0.0040
- 14Nemotron 3.5 Lightning$0.0044
- 15Llama 3.2 3B Instruct$0.0044
255 / 255
| # | ||||
|---|---|---|---|---|
| 1 | $2.52 | – | 63.1 | |
| 2 | $2.04 | – | 42.3 | |
| 3 | $2.04 | – | 39.5 | |
| 4 | $1.52 | – | 55.9 | |
| 5 | $1.38 | – | 59.6 | |
| 6 | $1.20 | $0.593 | 58.1 | |
| 7 | $1.20 | $0.593 | 55.7 | |
| 8 | $1.20 | – | 38.8 | |
| 9 | $1.14 | $0.802 | 52.5 | |
| 10 | $0.800 | $0.395 | 67.3 | |
| 11 | $0.800 | $0.361 | 66.7 | |
| 12 | $0.800 | $0.395 | 65.4 | |
| 13 | $0.720 | – | 41.1 | |
| 14 | 新 | $0.640 | $0.289 | 72.7 |
| 15 | $0.420 | $0.217 | 64.1 | |
| 16 | $0.420 | $0.217 | 56.9 | |
| 17 | $0.400 | $0.198 | 65.0 | |
| 18 | $0.400 | $0.198 | 63.1 | |
| 19 | $0.400 | $0.198 | 62.5 | |
| 20 | $0.400 | $0.198 | 61.1 | |
| 21 | $0.400 | $0.198 | 59.7 | |
| 22 | $0.320 | $0.149 | 68.4 | |
| 23 | $0.320 | $0.158 | 65.3 | |
| 24 | $0.320 | $0.149 | 64.4 | |
| 25 | $0.288 | $0.117 | 40.5 | |
| 26 | $0.240 | $0.118 | 60.7 | |
| 27 | $0.240 | – | 58.1 | |
| 28 | $0.240 | $0.118 | 57.9 | |
| 29 | $0.240 | $0.118 | 57.6 | |
| 30 | $0.240 | $0.118 | 53.8 | |
| 31 | $0.240 | – | 50.9 | |
| 32 | $0.240 | – | 46.8 | |
| 33 | $0.210 | $0.109 | 63.1 | |
| 34 | $0.200 | $0.116 | 45.1 | |
| 35 | $0.190 | $0.134 | 42.8 | |
| 36 | $0.190 | – | 42.1 | |
| 37 | $0.180 | $0.090 | 60.0 | |
| 38 | $0.168 | $0.087 | 60.4 | |
| 39 | $0.168 | $0.087 | 60.1 | |
| 40 | $0.161 | $0.090 | 62.9 | |
| 41 | $0.161 | $0.090 | 59.9 | |
| 42 | $0.161 | – | 58.3 | |
| 43 | $0.161 | $0.090 | 57.6 | |
| 44 | $0.160 | $0.074 | 68.9 | |
| 45 | $0.160 | $0.074 | 66.6 | |
| 46 | $0.160 | $0.074 | 65.3 | |
| 47 | $0.160 | $0.074 | 64.2 | |
| 48 | $0.160 | $0.079 | 62.6 | |
| 49 | $0.160 | $0.079 | 61.8 | |
| 50 | $0.160 | $0.079 | 59.0 | |
| 51 | $0.152 | $0.085 | 60.4 | |
| 52 | $0.152 | $0.085 | 48.7 | |
| 53 | $0.152 | – | 42.7 | |
| 54 | $0.152 | – | 25.8 | |
| 55 | $0.144 | $0.076 | 60.6 | |
| 56 | $0.144 | $0.068 | 59.3 | |
| 57 | $0.144 | $0.065 | 58.7 | |
| 58 | $0.144 | $0.076 | 57.4 | |
| 59 | $0.144 | $0.065 | 51.4 | |
| 60 | $0.144 | $0.063 | 40.8 |
引用格式:BenchLeader,《单任务成本:智能体编程会话》,数据截至 2026年11月16日,https://www.benchleader.com/zh/leaderboards/cost-agent-session。