How to Use OpenClaw with DeepSeek for 90% Less — NovAI Setup Guide
OpenClaw is the fastest-growing open-source AI coding assistant. But running it with GPT-4o or Claude gets expensive fast — a heavy day of coding can easily cost $10-25 in API fees.
What if you could get the same coding performance for 90% less?
In this guide, we connect OpenClaw to DeepSeek-v3.2 through NovAI. DeepSeek matches GPT-4o on code benchmarks (90.2% HumanEval) and beats it on math (90.0% MATH-500) — at $0.20/1M tokens instead of $2.50. No Chinese phone number needed.
(~1M input + 500K output tokens)
97% cost reduction
Why NovAI Instead of DeepSeek Directly?
- No Chinese phone number required. DeepSeek requires Chinese phone verification. NovAI: just email signup.
- 8 models, one API key. DeepSeek, Qwen-Max, GLM-4.6V, MiniMax (1M context), and more.
- One free model included. GLM-4.6V-Flash is completely free for testing.
- Hong Kong servers. Sub-80ms time to first token, especially fast in Asia-Pacific.
- Pay with USDT or PayPal (coming soon). No Chinese payment methods needed.
Step-by-Step Setup
1 Get your NovAI API key
Go to aiapi-pro.com, sign up with email, copy your API key. 30 seconds, no credit card.
2 Set the environment variable
# Add to ~/.bashrc or ~/.zshrc export NOVAI_API_KEY="your-novai-api-key-here"
Then: source ~/.bashrc
3 Configure OpenClaw
Open ~/.openclaw/openclaw.json and add NovAI as a custom provider:
{
"models": {
"mode": "merge",
"providers": {
"novai": {
"baseUrl": "https://aiapi-pro.com/v1",
"apiKey": "${NOVAI_API_KEY:-}",
"api": "openai-completions",
"models": [
{
"id": "deepseek-v3.2",
"name": "DeepSeek V3.2 (via NovAI)",
"reasoning": false,
"input": ["text"],
"cost": {"input": 0.0000002, "output": 0.0000004},
"contextWindow": 128000,
"maxTokens": 8192
},
{
"id": "qwen-max",
"name": "Qwen Max (via NovAI)",
"cost": {"input": 0.0000004, "output": 0.0000012},
"contextWindow": 32000,
"maxTokens": 8192
},
{
"id": "minimax-text-01",
"name": "MiniMax 1M Context (via NovAI)",
"cost": {"input": 0.0000002, "output": 0.0000016},
"contextWindow": 1000000,
"maxTokens": 8192
},
{
"id": "glm-4.6v-flash",
"name": "GLM Flash FREE (via NovAI)",
"cost": {"input": 0, "output": 0},
"contextWindow": 128000,
"maxTokens": 4096
}
]
}
}
},
"agents": {
"defaults": {
"model": {"primary": "novai/deepseek-v3.2"}
}
}
}4 Verify it works
# Check OpenClaw sees NovAI models openclaw models list | grep novai # novai/deepseek-v3.2 DeepSeek V3.2 (via NovAI) # novai/qwen-max Qwen Max (via NovAI) # novai/minimax-text-01 MiniMax 1M Context (via NovAI) # novai/glm-4.6v-flash GLM Flash FREE (via NovAI)
Which Model for Which Task?
| Task | Best Model | Why | Cost/1M Input |
|---|---|---|---|
| Coding & debugging | DeepSeek-v3.2 | 90.2% HumanEval | $0.20 |
| Full codebase analysis | MiniMax-Text-01 | 1M token context | $0.20 |
| Translation | Qwen-Max | Best Chinese-English | $0.40 |
| Testing & prototyping | GLM-4.6V-Flash | Completely free | $0.00 |
| Quick tasks | Qwen-Turbo | Fastest, ultra cheap | $0.06 |
Switch models on-the-fly: /model novai/minimax-text-01 for 1M context, /model novai/deepseek-v3.2 for coding.
Pro Tips
Tip 1: Start with the free model
Set glm-4.6v-flash as default first. Free, no rate limits. Test your setup, then switch to DeepSeek.
Tip 2: Use MiniMax for large codebases
MiniMax-Text-01 has 1 million token context — roughly 750K words or a ~3MB codebase in a single prompt. No other provider offers this at $0.20/1M tokens.
Tip 3: Multi-model routing
Use DeepSeek for coding, Qwen-Turbo ($0.06/1M) for commit messages and simple refactors. Another 70% savings on routine tasks.
Real Cost: 30 Days of OpenClaw
Scenario: Full-time developer, heavy usage
~2M input + 1M output tokens/day, 22 working days/month
GPT-4o: (2 x $2.50 + 1 x $10.00) x 22 = $330/month
Claude 3.5: (2 x $3.00 + 1 x $15.00) x 22 = $462/month
DeepSeek via NovAI: (2 x $0.20 + 1 x $0.40) x 22 = $17.60/month
Save $312-$444/month. A new MacBook every quarter.
FAQ
Is DeepSeek good enough for coding vs GPT-4o?
Yes. DeepSeek-v3.2 scores 90.2% on HumanEval (identical to GPT-4o) and 90.0% on MATH-500 (vs 76.6%). Most developers can't tell the difference for coding tasks.
Will NovAI work with the latest OpenClaw?
Yes. NovAI uses standard OpenAI-compatible API format. As long as OpenClaw supports custom providers (it does), NovAI works.
Do I need a Chinese phone number?
No. NovAI handles all upstream authentication. Email signup only, no phone, no VPN, no ID verification.
Can I still use GPT-4o alongside?
Yes. OpenClaw supports multiple providers. Keep OpenAI configured and add NovAI. Use /model to switch.
Start Using OpenClaw for 90% Less
Sign up in 30 seconds. Free model included. No credit card required.
Get Your NovAI API Key →