用户谈订阅预算：DeepSeek V4.1 Flash 1017 适合大规模跑
一位同时付费 Claude、Codex、Muse 的用户表示，月费 100–200 美元可接受，认为 V4.1 Flash 1017 在大规模使用上很划算，目前没有别的模型接近。
Depends on the models and usage. I’d not pay in the range of 500-1000/mo for oss modelsz Around 100-200/mo would work if the value is extremely good, but I already pay claude and codex and muse, so it would have to give me something I can’t get from running luna 24/7. Say deepseek v4.1 flash 1017 would be nice at scale. Not sure if anything else comes close currently. But I’m very happy to experiment. I got a nice swarm and integrating more models is easy for me. Also happy to try models that train on inputs, if that gives significantly more usage.
1. More tokens and speed than I get from Luna, right now it seems DS V4.1 Flash 1017 is better than Luna Max. If that ends up being the case, it'd be worth it 2. I don't understand, you talk about "unlimited" and then say 5-10B tokens/month ... that's nothing, if that's what you call unlimited then we have a very different idea, I used 6.2B tokens in the past 24 hours ... (1.3B opus, 2.7B sol, 1.4B deepseek) ... or are we talking uncached tokens? I hope you won't do what http://Z.Ai does, where they sell you "tokens" and then count cache towards the tokens, so if you run 98% cached you destroy your usage at 5x more than API would cost. On Opencode Go in past 24 hours I have 30M input, 1.21B cache read, 7.1M output for $12.47 API cost, over 11791 requests. 3. I don't care about ZDR, you can have my code, especially since a big chunk of my current work will be on something that's soon to be opensource Just to clarify, the only thing that beats Luna is either "better model" or "significantly faster per task" :)