AI 编程工具到底花多少钱:629 行数,每行带原句和日期
English · 中文 · Español · 日本語 · 한국어 · Tiếng Việt · Français · Deutsch · Русский · Bahasa Indonesia
这是一份公开数表。64 篇实测笔记里出现过的每一个数 —— 价格、百分比、倍数、token 数、耗时 —— 都抽成一行,每行都附上它原来所在的那句话和发布日期。
跑 agent 的模型,今天多少钱
68 个在 Design Arena 的 agents 各档里排过名的模型,每百万 token 的挂牌价——不是账单:缓存命中、batch、不同接入方各有各的价。取自 OpenRouter 的公开目录,最后核到 2026-10-11。最便宜的三个:
| $ in / 1M | $ out / 1M | Model | Best agents rank |
|---|---|---|---|
| $0.06 | $7.00 | GLM 5.2 | #10 agenticgamedev |
| $0.07 | $4.80 | GLM 5.3 | #8 python-pptxslides |
| $0.09 | $0.36 | Solar Pro 4 | #35 webapps |
同一个数,相反的答案。 Google 和 xAI 都在 200,000 个输入 token 上跳贵价档 —— 而正好 200,000 这一发,Google 按便宜的收,xAI 按贵的收。别处的价目表印一个门槛数就没了下文。哪家算在哪一边、引的是厂商自己那页上的原话、附核对日期:same number, opposite answer(英文)。
上面那张是挂牌价,没有一张账单跟它对得上。 差额在缓存命中、重试、和你要付两遍钱重发的上下文里 —— 这些目录表一样都看不见。有一篇专门去追这段差额:钱到底花在哪儿了。它结尾问的正是这张表答不上来的那个问题:上个月你付了多少,其中有多少是你不情愿重发一遍的上下文? 那一页有回复框,这一页没有。
先看数
下面几行是英文原文照抄,没有翻译:一个数离开它原来那句话就没法核对 —— $1.43 可能是每百万 token、每月、也可能是每个席位。
| Figure | The sentence it came from | Write-up |
|---|---|---|
$7,600 |
An AI crypto trading bot reportedly made $7,600 in a week.” That line gets clicks. | How a Multi-Agent AI System Made $7,600 in 7 Days for Under $100 in API Costs |
$5 |
The tool runs on a $5/month VPS using Docker. | How Alibaba’s Open Code Review Slashed AI Code Review Costs by 90% — And What It Means for Independent Developers |
$0.028 |
With DeepSeek V4 Flash priced at just $0.028 per million tokens, developers can assign lightweight tasks like boilerplate generation or syntax checking to this model, reserving higher-cost models for nuanced logic or debugging. | Bypass Codex Rate Limits: The Local Proxy Path to 70% Cost Savings |
40% |
40% of review tasks needed manual backfill, costing about 15 extra minutes each time. | Your AI Didn’t Misread Your Code by Accident. You Handed It the Wrong Context. |
$1.43 |
The price spread is wide even among the Western flagships, which becomes obvious on ReactBench, where one run with GPT 5.6 Sol costs about $1.43 while one run with Fable 5 costs $9.05, which means that a single Fable 5 run comes to a bit more than six times as much as the GPT 5.6 Sol run does. | Stop Hitting the 5-Hour Limit: Routing Your IDE’s AI Requests to Local Models |
20% |
I’d argue that building these decoupling layers is the only way to survive the shift toward platforms like ChatGPT Work, where non-programming users are expected to jump from 20% to 60% of the total user base within a year. | Four Circuit Breakers Every Unattended AI Pipeline Needs (Learned the Expensive Way) |
$170,000 |
In contrast to one-off projects, micro-automation tools excel at addressing single, high-frequency pain points A former Alibaba P8, after facing three months of unsuccessful job applications, pivoted to building AI software that has five core features—including scheduled automation and skill packages—to generate $170,000 in monthly revenue The data from these 27 successful cases reveals that the most effective tools prioritize “scheduled automation,” which can reduce manual task time from several hours each day to zero minutes. | Stop Building Custom AI Agents: How to Earn $63K/Month with Micro-Automation |
80% |
The machine runs for about 11 minutes, and human review takes 10–20 minutes, leading to a total time reduction of ~80%–90%. | I built a WeChat Official Account writing Agent that cuts work time by 80% — here’s the workflow filter and tool combo I used |
54% |
Through its internal multi-agent system, Sol achieves 54% higher token efficiency on agentic coding tasks compared to peer models, showing how strategic design choices can transform from a development challenge into a cost-saving advantage. | The Hidden Costs of Over-Prompting in AI Coding: Lessons from Claude Code’s Optimization |
$10 |
An AI image generation tool used Google Ads to boost its monthly paying subscribers from approximately 80 to over 500, maintaining a customer acquisition cost of around $10 per paying user. | 63 Billion Tokens in One Month, Spent Writing Google Ads for Amazon Sellers |
$1.43 |
One GPT-5.6 Sol run might cost $1.43, while other models could cost upwards of $9.00 for the same task. | Claude Code Can Model a Flange but Not a Freeform Surface: A Six-Step Handoff Checklist for Non-Code Agent Output |
$63,000, |
Consider Jordan, who noticed his partner spending hours manually sharing items on Poshmark, and by developing a simple 30-line JavaScript automation script to solve this pain point, he created Resellbot, which eventually scaled to a monthly revenue of $63,000, while this progression highlights how identifying such tedious manual tasks can serve as a potent foundation for building highly profitable and scalable software solutions. | Building High-Income Single-Page Tool Sites via SEO |
curl -s https://cdn.jsdelivr.net/gh/xyzs996/llm-api-pricing@main/data/figures.json
curl -s https://cdn.jsdelivr.net/gh/xyzs996/llm-api-pricing@main/data/figures.csv
字段里的 published 是那篇笔记发出去的日子,不是这个价钱当时还作数的日子。价格一直在变,每一行都得按它自己的日期读。
正文
正文是英文,在这儿:https://spectracodeai.com/ —— 带数表、题目页和厂商页。只想要数的话,上面两条 curl 就够了,不用读正文。
说一句
- 收藏这个仓库,数表更新会跟着走。数据是 CC BY 的,收不收藏都随便拿,这儿不拿数据换 star。收藏改变的是下一个找这些数的人能不能找到 —— GitHub 的搜索排序和「相关仓库」推荐都把 star 数算进去。
- 数不对? 哪个价钱变了、你自己跑出来是另一个数 —— 提个issue。这个仓库就是干这个的。 (issue)
- 想要的数这儿没有? 说清楚要哪个指标、哪家、什么单位,一行就够 —— 表单只有一格必填,收到就收成新行。 (form)
CC BY 4.0:随便复制、转载、再加工、拿去卖。唯一的条件是说一声出处,挂个链接回 https://spectracodeai.com/ 就行。