DEV Community: 于侃

Why I Stopped Overpaying for AI APIs (And Built a Gateway to Chinese Models Instead)

于侃 — Sat, 14 Mar 2026 17:06:12 +0000

Last month, a friend showed me his AI API bill. It was shocking.

He was paying hundreds of dollars every month just to access GPT-4, Claude, and other Western AI models for his startup. As a solo developer, that was eating up most of his budget.

"There's got to be a better way," he said.

There wasn't. So we built one.

The Problem Nobody Talks About

Western AI APIs are expensive. And if you're building in Asia or serving Asian markets, you're paying premium prices for models that weren't even optimized for your use case.

Meanwhile, Chinese AI companies like DeepSeek, Zhipu AI, and Moonshot AI were building incredibly capable models at a fraction of the cost. The problem? No easy way to access them.

Documentation in Chinese
Payment methods that don't work internationally
No standardized API format
Complex authentication flows

What We Built

NovAI is an open-source API gateway that solves this:

One API format - Use the familiar OpenAI-compatible interface
Access to Chinese models - DeepSeek, GLM-4, Moonshot, and more
Significant cost savings - Often 60-80% cheaper than Western alternatives
Simple authentication - Just an API key, no complex setup

Here's how simple it is:

import openai

client = openai.OpenAI(
    api_key="your-novai-key",
    base_url="https://aiapi-pro.com/v1"
)

response = client.chat.completions.create(
    model="deepseek-chat",
    messages=[{"role": "user", "content": "Hello!"}]
)

That's it. No new SDK to learn. Just change the base URL and start saving.

Real Results

My friend? His monthly AI costs dropped by over 80%. He's now running his entire AI infrastructure on Chinese models through our gateway, and his users can't tell the difference.

The models are that good.

Why Open Source?

We believe AI infrastructure should be:

Accessible - Not locked behind expensive paywalls
Transparent - You should know what you're running
Flexible - Deploy it yourself if you want

That's why we open-sourced NovAI. Check it out: https://github.com/novai-gateway/novai

Try It Yourself

We wrote an open letter sharing our full story and vision:

Read our Open Letter

Or just try the API right now at aiapi-pro.com

Have you tried Chinese AI models? What's been your experience with API costs? Let's discuss in the comments.

NovAI Agent - Open Source AI Coding Assistant for Automated Code Review & Refactoring

于侃 — Sat, 14 Mar 2026 04:55:07 +0000

Introduction

NovAI Agent is an open-source AI coding assistant built on the NovAI API, helping developers automate code review, refactoring, and test generation.

Core Features

🤖 Interactive AI Chat - Smart answers to programming questions
🔍 Automated Code Review - Detect security vulnerabilities, performance issues, and code style problems
🔄 Intelligent Code Refactoring - Automatically optimize code based on goals
🧪 Automated Test Generation - Generate unit tests with edge case coverage
💰 Extremely Low Cost - Using NovAI API at 1/10th the price of OpenAI
🌐 Global Access - Stable access from anywhere

Quick Start

# Install
pip install novai-agent

# Configure API key
novai-agent config --api-key your-api-key

# Code review
novai-agent review app.py

# Code refactoring
novai-agent refactor legacy.py --goal "reduce complexity"

# Generate tests
novai-agent test utils.py --framework pytest

Cost Comparison

Service	Input Price	Output Price
OpenAI	$0.15/1M tokens	$0.60/1M tokens
NovAI	~$0.55/1M tokens	~$1.65/1M tokens

Actual cost is approximately 1/10th of OpenAI

Comparison with Competitors

Product	Type	Price	Open Source
GitHub Copilot	IDE Plugin	$10/month	❌
Cursor	AI Editor	$20/month	❌
Tongyi Lingma	IDE Plugin	Free	❌
NovAI Agent	CLI Tool	Pay-per-use	✅

Use Cases

Automated code review
Legacy code refactoring
Test case generation
CI/CD integration
Batch code analysis

Tech Stack

Python 3.8+
OpenAI-compatible API
tiktoken tokenization
CLI interface

Links

Star ⭐ and contributions welcome!

AI API Latency Test: US Servers vs Hong Kong from Asia

于侃 — Fri, 13 Mar 2026 12:18:52 +0000

I ran latency tests on 5 major AI API providers from Asia. The results surprised me.

Why Latency Matters

When building AI applications, every millisecond counts. For a chat interface with 10 back-and-forth messages:

300ms latency = 3 seconds of total wait time
80ms latency = 0.8 seconds total

That's the difference between a snappy app and a frustrating experience.

The Test Setup

I tested from 3 locations in Asia:

Singapore (AWS)
Tokyo (GCP)
Hong Kong (Alibaba Cloud)

Tested providers:

OpenAI (US West)
Anthropic (US East)
OpenRouter (US)
NovAI (Hong Kong)
DeepSeek (China)

Results: First Token Latency (ms)

Provider	Singapore	Tokyo	Hong Kong	Average
NovAI	75ms	82ms	68ms	75ms
DeepSeek	145ms	160ms	120ms	142ms
OpenAI	220ms	235ms	195ms	217ms
Anthropic	245ms	260ms	220ms	242ms
OpenRouter	210ms	225ms	185ms	207ms

Key Findings

1. Geography beats everything
Hong Kong-based servers are 3x faster than US-based ones from Asia.

2. Network quality matters
CN2 GIA routing (NovAI) vs standard internet makes a 20-30ms difference.

3. Provider optimizations
Some providers use edge caching and connection pooling to reduce latency.

Real-World Impact

I migrated my OpenClaw app from OpenRouter to NovAI:

Before: 2.3s average response time
After: 0.9s average response time
User satisfaction scores improved 40%

Methodology

Tests were run over 7 days, 100 requests per provider per location. Measured time to first token (TTFT) using identical prompts.

Full details: https://aiapi-pro.com/blog/ai-api-latency-test

What latency are you seeing from your location?

DeepSeek API Timeout? 5 Alternatives with Lower Latency from Asia

于侃 — Fri, 13 Mar 2026 12:18:21 +0000

If you're building AI applications in Asia, you've probably experienced DeepSeek's API timeout issues. Here's what I found after testing 5 alternatives.

The Problem

DeepSeek's official API has been struggling with:

30+ second response times
Frequent 504 timeouts
300-500ms network latency from Asia
Aggressive rate limiting (10 RPM)

The Alternatives I Tested

I spent a week testing providers specifically for low-latency access from Asia:

1. NovAI (Hong Kong) - ~80ms latency

Best for: Chinese models (DeepSeek, Qwen, GLM)
Pricing: $0.20/1M tokens (cheaper than DeepSeek direct)
Pros: Hong Kong servers, OpenAI-compatible API

2. OpenRouter (US) - ~220ms latency

Best for: Wide model selection
Pricing: Varies by model
Cons: US-based adds latency for Asia users

3. SiliconFlow (China) - ~150ms latency

Best for: Domestic Chinese access
Pricing: Competitive
Cons: Requires China business registration

4. AWS Bedrock (Singapore) - ~120ms latency

Best for: Enterprise users
Pricing: Higher but includes support
Cons: Complex setup, limited model selection

5. Google Vertex (Singapore) - ~95ms latency

Best for: Google Cloud users
Pricing: Premium
Cons: Limited Chinese model support

Key Findings

Server location matters more than expected.

For a chat app with 10 back-and-forth messages:

DeepSeek direct: 3.5 seconds total wait time
Hong Kong provider: 0.8 seconds total wait time

That's a 4x improvement in user experience.

My Recommendation

For production apps serving users in Asia:

Use a Hong Kong-based provider for Chinese models
Consider Singapore endpoints for Claude/GPT
Always test latency from your target region

I wrote a detailed comparison with code examples here:
https://aiapi-pro.com/blog/deepseek-api-timeout-alternatives

What API providers are you using for AI apps in Asia?

I Made a Free GitHub Copilot Alternative Using Chinese AI Models

于侃 — Wed, 11 Mar 2026 07:58:27 +0000

GitHub Copilot costs $19/month. OpenAI's API needs a credit card. And if you're outside the US, getting set up with either can be a real headache.

So I built NovAI Coder — a free, open-source Windows app that gives you AI coding assistance through 7 Chinese AI models that rival GPT-4o in quality.

Why Chinese AI Models?

DeepSeek V3.2 scores 90.2% on HumanEval — the exact same as GPT-4o. But it costs $0.14 per million input tokens vs GPT-4o's $2.50. That's 18x cheaper.

The catch? Accessing these models directly requires a Chinese phone number and navigating Chinese-language dashboards. NovAI removes that barrier — sign up with email, get an API key, start coding.

What's Inside

NovAI Coder bundles OpenClaw (an open-source coding agent) with pre-configured access to:

Model	Cost	Best For
GLM-4.6V-Flash	FREE	Testing, prototyping
Qwen-Turbo	FREE	Quick tasks
DeepSeek V3.2	$0.14/1M	Coding, reasoning
Qwen-Plus	$0.20/1M	Multilingual
MiniMax-Text-01	$0.20/1M	1M context, entire repos
GLM-4.6V	$0.40/1M	Vision + text
Qwen-Max	$0.40/1M	Creative writing

Getting Started

Download from GitHub Releases
Run the installer (one click)
Register at aiapi-pro.com (email only, $0.50 free credits)
Paste API key → start coding

The whole process takes under 2 minutes.

Open Source

MIT license. Full source on GitHub.

If you're tired of Copilot's pricing or can't get an international credit card, give it a try. Feedback and contributions welcome!

How I Cut My OpenClaw API Costs by 97% (From $330/mo to $18)

于侃 — Tue, 10 Mar 2026 13:19:06 +0000

I love OpenClaw. It's genuinely changed how I code. But after my first month, I looked at my API bill and nearly choked — $330 on GPT-4o alone.

So I started experimenting. After testing multiple alternatives, I found a setup that gives me the same coding quality for $18/month: DeepSeek-v3.2 through a gateway called NovAI.

Here's exactly how to set it up.

Why DeepSeek?
Before you dismiss this as "just use a cheaper model" — look at the benchmarks:

Benchmark DeepSeek-v3.2 GPT-4o Claude 3.5
HumanEval (code) 90.2% 90.2% 92.0%
MATH-500 90.0% 76.6% 78.3%
Input / 1M tokens $0.20 $2.50 $3.00
Output / 1M tokens $0.40 $10.00 $15.00
DeepSeek matches GPT-4o on code and beats it on math. The price difference is 12x on input, 25x on output.

Why NovAI Instead of DeepSeek Directly?
DeepSeek's official API requires a Chinese phone number to sign up. That's a dealbreaker for most of us.

NovAI
https://aiapi-pro.com
is a gateway that solves this:

Email signup only — no phone, no VPN, no ID verification
8 models, one API key — DeepSeek, Qwen, GLM, MiniMax, Moonshot
One FREE model — GLM-4.6V-Flash, no usage limits, perfect for testing
OpenAI-compatible API — works as an OpenClaw custom provider out of the box
Hong Kong servers — sub-80ms TTFT, especially fast in Asia-Pacific