惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

F
Fortinet All Blogs
Recent Announcements
Recent Announcements
H
Help Net Security
Y
Y Combinator Blog
酷 壳 – CoolShell
酷 壳 – CoolShell
有赞技术团队
有赞技术团队
小众软件
小众软件
Last Week in AI
Last Week in AI
U
Unit 42
Google DeepMind News
Google DeepMind News
博客园 - 司徒正美
H
Hackread – Cybersecurity News, Data Breaches, AI and More
J
Java Code Geeks
Microsoft Security Blog
Microsoft Security Blog
G
Google Developers Blog
N
Netflix TechBlog - Medium
Blog — PlanetScale
Blog — PlanetScale
云风的 BLOG
云风的 BLOG
V
V2EX
博客园 - 聂微东
人人都是产品经理
人人都是产品经理
博客园 - 三生石上(FineUI控件)
阮一峰的网络日志
阮一峰的网络日志
爱范儿
爱范儿

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
I Cut My AI Test Automation Cost by 300x by Ditching Visi...
xulingfeng · 2026-05-20 · via DEV Community

I Cut My AI Test Automation Cost by 300x by Ditching Vision Models

From $0.011 per step to $0.00004 — here's how I learned vision models are overkill for most web testing, and what I built instead.


It started with a $400 monthly API bill (and yes, that's USD — I'm in China, but you'll feel the same pain in any currency).

I was running an AI-powered test automation platform built on Midscene.js with Qwen-VL vision models. Every test step meant sending a full-page screenshot to a multimodal LLM — and paying about $0.011 per step.

A 50-step test case cost about $0.55. Run it daily? $16.50/month. Add a few more test scenarios, and suddenly I was spending more on API calls than on coffee.

And the worst part? Most of those screenshots contained information I already had for free.

The Platform That Taught Me a Lesson

First, a quick backstory.

I built ai-test-platform, a full-stack test automation management system:

  • Frontend: Vue 3 + ElementUI Plus
  • Backend: Express + Node.js + MySQL
  • Test engine: Midscene.js 1.5.2 + Playwright + Qwen-VL
  • Dockerized, with a management UI for test cases, reports, and models

It worked. Beautiful reports, clean UI, easy test management. I even pushed it to Docker Hub (xulingfeng/ai-test-platform:latest).

But every time I ran a test, I could almost hear the coins dropping. $0.011 here, $0.011 there. A 29-step doctor-onboarding flow cost $0.32.

For a solo QA engineer running tests multiple times a day, that adds up fast.

The Moment It Clicked

I was watching a test run one afternoon. The AI was analyzing a screenshot of a web page — and I realized something:

The AI could see 45 interactive elements in the screenshot. But Playwright had already extracted all 45 of them as clean structured text.

I was paying to process pixels when the data was already neatly organized in the DOM tree.

Here's what a page looks like to a vision model:

[screenshot image with pixel data, rendering details, colors, shadows...]

And here's what it looks like in the DOM:

[0] <input placeholder="Search..." name="q">
[1] <button>Sign in</button>
[2] <a>Add new doctor</a>
...

Enter fullscreen mode Exit fullscreen mode

The AI doesn't need to "see" the page. It needs to understand the structure and decide what to click. And structured text does that perfectly.

The 300x Optimization: deep-test

I built deep-test — a pure-text AI testing framework.

The architecture is embarrassingly simple:

Task: "Login system, search product, add to cart"
         ↓
① Extract interactive elements (DOM tree / uiautomator)
   (No screenshots. No vision models.)
         ↓
② DeepSeek V4 analyzes structure + decides next action
   (~2000 tokens/step × $0.14/M = $0.0001/step)
         ↓
③ Execute action (Playwright click / ADB tap)
         ↓
④ Back to ① until task completes

Enter fullscreen mode Exit fullscreen mode

The cost comparison is ridiculous:

Approach Per step 50-step test
Midscene.js + Qwen-VL-Plus ~$0.011 ~$0.55
browser-use + Claude ~$0.10 ~$5.00
deep-test + DeepSeek V4 ~$0.00004 ~$0.002

200-300x cheaper. The 50-step test that cost $0.55 now costs less than a cent.

The Real-World Numbers

I ran a complete hospital management workflow — login, navigate menus, add a new doctor with 12 fields, verify the result. 29 steps total.

Result: 81.8 seconds, ~$0.001 total cost.

For context, that's less than the price of a single step on the vision-based approach.

But Wait — What About Android Apps?

Here's where it gets even more interesting.

Android apps can't give you a clean DOM tree like a web page. So I added a hybrid approach:

  1. Use uiautomator2 to extract the native UI tree (it's text, just like DOM)
  2. Use ADB screencap + OCR only when the UI tree doesn't have enough info
  3. Same DeepSeek V4 decision engine — just different input sources

This means one AI agent handles both Web and Android with the same architecture.

And I even solved the notorious hybrid app WebView input problem — where in-app web views ignore standard automation commands. The fix: uiautomator2.send_keys() instead of set_text(). Took days to figure out, one line to implement.

What I Learned

Vision models are overkill for most web testing.

They're great for:

  • Visual regression testing (did the layout break?)
  • CAPTCHA solving
  • Canvas/SVG-heavy applications

But for standard CRUD operations — filling forms, clicking buttons, navigating menus — the DOM already has all the information you need.

The real optimization isn't about better prompting or smarter AI. It's about choosing the right data format for the job.

The Tools

Both projects are not yet public — they contain real test data from production healthcare applications. I plan to clean and open-source them once the company-specific content is stripped out. If you'd like early access or want to discuss the approach, feel free to reach out.

The tech stack:

  • LLM: DeepSeek V4 Flash ($0.14/M input, $0.28/M output)
  • Web automation: Playwright
  • Android automation: uiautomator2 + ADB
  • OCR: EasyOCR (local, no API cost)

I'm a test manager with 15 years of experience. I've been building AI testing tools on the side because I believe good testing shouldn't cost a fortune. If this resonates, I share more practical testing prompts and techniques in my toolkit: xulingfeng.gumroad.com/l/vkhhq