Tokens 如今就是租金:如何在 AI 計價翻倍前從中獲利
Tokens Are Rent Now: How to Earn From AI's Meter Before the Price Doubles
這份指南拆解 AI token 計費方式,並提出三種獲利路徑:製作模板反覆銷售、租用 GPU 提供運算成果,以及教授用戶管理帳單。Gemini 4 Carbon 的每百萬 input tokens 價格將於 Dec 1 由 $2 升至 $4,output tokens 則由 $10 升至 $20;提示詞快取、批次處理及模型路由可降低成本。
Tokens Are Rent Now: How to Earn From AI's Meter Before the Price Doubles
Google doubles Gemini 4 Carbon on Dec 1. Input moves from $2 to $4 per million tokens. Output moves from $10 to $20. The meter is moving, and the meter is the business.
A token is about four characters. Providers charge for every token a model reads and writes. One formula runs every bill: input tokens times the input price, plus output tokens times the output price, divided by one million. That is the whole economy in one line.
This guide shows the math and three ways to earn from it. Build once, sell repeatedly.
The meter
Output costs more than input. Frontier models charge about five times more for output than for input. The ratio holds across the market.
The spread is the story. GPT-6 Astra charges $10 in and $50 out per million tokens. GPT-6 Luna charges $0.10 in and $0.50 out. Same unit, 100 times apart.
A 2,000-token prompt with a 500-token reply on Claude Sonnet 5 costs $0.009. Less than a cent. At 100,000 requests a day, the same request costs $900 a day. About $27,000 a month.
The request that looked free in testing is a salary at scale.
| Model | Input per 1M | Output per 1M |
|---|---|---|
| GPT-6 Astra | $10 | $50 |
| Claude Fable 5.1 | $10 | $50 |
| GPT-6.1 Sol | $2 | $10 |
| Claude Sonnet 5.5 | $2 | $10 |
| Gemini 4 Carbon | $2 ($4 from Dec 1) | $10 ($20 from Dec 1) |
| GPT-6 Luna | $0.10 | $0.50 |
| Claude Haiku 5.5 | $0.10 | $0.50 |
Prices are list prices from the providers, checked November 2026. They change often.
Cut the bill first
Prompt caching means the provider reuses text it has already seen, at a lower price. It cuts repeated input to 10% of list price. On Claude Fable 5.1 it falls to 2.5%.
Batch processing means the provider runs your requests in a slow lane for half price. It halves the whole bill at OpenAI, Anthropic, and Google.
Routing matters more than list price. Send each task to the cheapest model that passes your quality bar. Simple tasks can cost 99% less this way.
And the cheapest list price can still ship the most expensive bill. Cost per useful output can vary several times across stacks that picked the cheapest rate. Routing and prompt shape swamp the headline number.
A gateway is a middle layer that routes and watches your API calls. One enterprise report puts gateway markup at 15 to 35 percent above raw API list. The same report puts retry waste at 10 to 25 percent of agent token volume before any user value appears. Count those in your math, or your margin is fiction.
Path 1: Package it once, sell it forever
Sell the outcome, not the token. Fin charges $0.99 per resolved conversation. HubSpot charges $0.50. Their token cost is a fraction of a cent. The buyer pays for a solved problem.
Build the template around one vertical. Pick one industry and one workflow. Aim it at one clear result. A generic prompt pack competes with free. A finance template for freelancers competes with nothing.
Price it two ways. Per use for the curious, monthly for the committed. The monthly plan is the repeat income.
Your first step: pick one workflow you already do by hand, turn it into a template, and price the result.
Path 2: Rent the machines, sell the gap
You do not need to own a GPU. An RTX 5090 rents for $0.79 an hour on Vast.ai. RunPod lists the same card at $1.39. Buying one costs $6,199 to $11,896.
An RTX 4090 rents from $0.30 an hour. Buying one costs $4,899. That price equals about 16,300 rental hours. Just under two years running nonstop.
H100 time runs about $3 per GPU-hour on market indexes. AWS on-demand lists it near $7. The gap between the cheap market and the dear cloud is the margin.
The play is simple. Rent the card by the hour, run the job, sell the result. A fine-tune. A render. A batch of images. The customer pays for the finished file, not the hour.
Watch the traps. Storage and bandwidth bill separately. Vast.ai hosts vary in reliability. A cheap host that dies mid-job costs more than a dear one that finishes.
Path 3: Teach the bill
Most users cannot read their own AI bill. That gap is a course.
Teach beginners to build one real AI workflow. Then teach them to read the meter: what a token costs and where the bill hides. Then show which lever cuts it. Call it AI plus a practical skill.
The course sells because the pain is real. A team that finds its $27,000-a-month leak pays for the lesson that found it.
Your first step: record one screen session of you cutting a real bill in half. That recording is lesson one.
What can go wrong
Prices move. Gemini doubles on Dec 1. Gemini 3.8 Flash doubles on Jan 1. A margin built on today's table breaks on the next one.
The realized bill beats the list price. The rate card is the floor. Gateway markup sits on top of it. So do retries and tool-call waste. One enterprise report puts frontier work at $3 to $12 per million blended tokens.
None of the three paths pays on day one. Templates need a buyer. Compute needs a customer. And courses need trust, which takes the longest to earn. The repeat income starts after the first sale, not before it.
Verdict
The token is the unit now. Every AI service bills by it. The people who earn from it are the ones who read the meter before they build.
Start with one number. Open any provider pricing page. Compute what one real task costs you today. Then pick one path and build the smallest version of it this week.
Follow along
I test AI tools and build in public:
- GitHub: https://github.com/ovu-lab
- YouTube: https://youtube.com/@oVuCyber
來源:DEV Community · dev.to