AA 每任务成本：V4.1 Flash 得分最高且第二便宜，$0.26/任务

💥 \newline @newlinedotco原推发布于 2026.10.21

在 X 上打开原推

Of the 4 open-weight releases from Oct 15 to 17, the one that scores highest is also the second cheapest to run per task: DeepSeek-V4.1-Flash-1017. Artificial Analysis (AA) prices each model per task: listed API price times the tokens the model spends on AA's Intelligence Index, a score averaged over 10 evaluations. DeepSeek spends the most tokens of the priced models (210M output tokens over the Index, vs 110M for Meta's Muse Spark Open and 79M for Reflection's Beam). Its price of $0.30 / $1.20 per 1M input/output tokens still puts it at $0.26 per task, against $0.57 for Beam and $0.84 for Muse, with an Index score of 45 to their 30 and 39. (That is DeepSeek's peak rate; off-peak is half.) The licenses vary just as much: MIT (DeepSeek), Apache 2.0 (Beam), OpenMDW-1.1 (NVIDIA's 2 Nemotron 3.5 models), and Meta's custom Muse Model License. That one requires a "Built with Muse Spark Open" label in your product or its docs, and a separate agreement with Meta if you sell model access through an API and your company makes over $20M in 12 months. An average can also hide a weak spot: AA notes Muse ranks last of the 9 top open-weight models on the Index's 4 agentic evaluations.