工具
精华
|
战斗力 鹅
|
回帖 0
注册时间 2004-11-14
|
发表于 2025-1-21 18:15
来自手机
|
显示全部楼层
本帖最后由 alixsander 于 2025-1-21 18:26 编辑
overflowal 发表于 2025-1-21 17:41
怎么就盈利是不可能的,之前V2时api的定价Deepseek都说是赚钱的,这次计算量实在提高了才提价。
实际收费 ...
https://www.reddit.com/r/LocalLLaMA/s/ocUEwqmk0P
I calculated the effective cost of R1 Vs o1 and here's what I found
In order to calculate the effective cost of R1 Vs o1, we need to know 2 things:
1. how much each model costs per million output tokens.
2. how much tokens each model generates on average per Chain-of-Thought.
You might think: Wait, we can't see o1's CoT since OpenAI hides it, right? While OpenAI does hide the internal CoTs when using o1 via ChatGPT and the API, they did reveal full non-summarized CoTs in the initial announcement of o1-preview ([Source](https://openai.com/index/learning-to-reason-with-llms/)). Later, when o1-2024-1217 was released in December, OpenAI stated,
>o1 uses on average 60% fewer reasoning tokens than o1-preview for a given request
([Source](https://openai.com/index/o1-and-new-tools-for-developers/)). Thus, we can calculate the average for o1 by multiplying o1-preview’s token averages by 0.4.
The Chain-of-Thought character count per example OpenAI showed us is as follows, as well as the exact same question on R1 below:
o1 - \[(16577 + 4475 + 20248 + 12276 + 2930 + 3397 + 2265 + 3542)\*0.4\]/8 = 3285.5 characters per CoT.
R1 - (14777 + 14911 + 54837 + 35459 + 7795 + 24143 + 7361 + 4115)/8 = 20424.75 characters per CoT.
20424.75/3285.5 ≈ 6.22
R1 generates 6.22x more reasoning tokens on average than o1 according to the official examples average.
R1 costs $2.19/1M output tokens.
o1 costs $60/1M output tokens.
60/2.19 ≈ 27.4
o1 costs 27.4x more than R1 price-per-token, however, generates 6.22x fewer tokens.
27.4/6.22 ≈ 4.41
# Therefore in practice R1 is only 4.41x cheaper than o1
(note assumptions made):
If o1 generates x less characters it will also be roughly x less tokens. This assumption is fair, however, the precise exact values can vary slightly but should not effect things noticeably.
This is just API discussion if you use R1 via the website or the app its infinitely cheaper since its free Vs $20/mo. |
|