Recommended OpenClaw Model Stack

The recommended default is a stack, not a single model

The Orange Paper’s most practical advice is to stop asking for one universal best model. Most users need a stack with three roles:

• a strong primary model
• cheaper fallbacks
• a free or local route for low-value background work

The default recommendation

For most English- and Chinese-speaking operators in March 2026, the best editorial default is:

• primary: Claude Sonnet 4.6
• fallback 1: Claude Haiku 4.5
• fallback 2: DeepSeek-V3.2

This gives you strong agent performance without forcing every request through the expensive path.

Why this works

• Sonnet handles difficult reasoning and tool use well
• Haiku is a cheaper buffer when you do not need the strongest model
• DeepSeek is dramatically cheaper for daily traffic

If you add Gemini Flash or a local Ollama model for cron and heartbeat work, the cost profile improves further.

When to choose something else

Prefer a different stack when:

• you are China-first and want GLM, Kimi, MiniMax, or bundled cloud plans
• you are privacy-first and already have strong local hardware
• you need a fully free experimental setup

The important part is still the same: distribute workload by value, not by habit.

The control rules

No model stack is safe by default unless you also:

• set a daily budget limit
• keep a fallback chain
• monitor background jobs
• avoid letting premium models handle low-value scheduled work

Where to go next

• Open the Models Hub
• Read OpenClaw Security & Cost Control
• Return to the Install Hub