AI 每月維護成本：每月要付多少？
AI maintenance cost: what you pay per month
作者估算，自己每月在 Claude、ChatGPT、GLM 訂閲及伺服器上的開支約為 $300；若以 API 按 token 用量付費，完成相同工作估計約需 $30,000。訂閲設有用量限制，API 費用則視乎模型、token 及所用功能；自建系統亦須預算伺服器、第三方服務及維護工作。
I think AI is underpriced right now, although plenty of people would argue with me. There are several ways to use it, and subscriptions are the most common. My recurring spending on three models and a server is around $300 a month
I pay $200 for Claude, $20 for ChatGPT, $20 for GLM and $30 for the server per month. I also track what using those models would cost if I paid directly for tokens through an API. My estimate is around $30,000 for the same volume of work. Tokens are pieces of text a model receives and generates. Subscriptions and token billing work differently, so that is my calculation, not a promise that everyone would save the same amount
Subscriptions, APIs and your own hardware
A subscription has a fixed price, but usage stays within the plan's limits. For example, Claude plans have usage limits, and ChatGPT Plus has limits too. Conversation length, model choice and tools can affect how much of the allowance you use
With an API, the bill depends on usage. According to OpenAI's billing guide, API charges are separate from a ChatGPT subscription and depend on the model, tokens and features used. Before choosing, estimate the workload: a subscription price and an API bill do not necessarily buy the same amount of work
Another option is to run a model on your own hardware. A local system using a powerful modern model needs computing resources, memory and storage. The exact requirements depend on the model; an ordinary laptop may not be enough for some larger ones
What goes into the monthly budget
My main expenses are subscriptions and a remote server. I buy additional tools when I need them. That spending lets me work on roughly 2-3 large projects a month while developing my own system
For a particular system, budget separately for the model, server and third-party services, as well as maintenance work: checking failures, applying updates and adapting it to new tasks. The mix depends on what the system does and which services it needs. I explain what AI system support and development covers on my site
The limits used to be much higher, but the grass used to be greener too
來源：DEV Community · dev.to