Choose an AI Model

OpenClaw supports multiple LLM providers. You can use different models for different scenarios and even set up automatic switching. This guide helps you make the best choice.

---

Model Comparison Overview

Model	Provider	Cost	Speed	Quality	Privacy	Best For
Claude Opus 4.6	Anthropic	$$$	Medium	5/5	Cloud	Complex reasoning, long-form writing
Claude Sonnet 4.5	Anthropic	$$	Fast	4/5	Cloud	Daily conversation, general tasks
GPT-5.2 Codex	OpenAI	$$$	Medium	5/5	Cloud	Software development, technical tasks
GPT-5.2 Mini	OpenAI	$	Very fast	3/5	Cloud	Simple chat, quick responses
Gemini 3 Pro	Google	$$	Fast	4/5	Cloud	Multimodal, search integration
DeepSeek V3	DeepSeek	$	Fast	4/5	Cloud*	Chinese tasks, high cost-effectiveness
Ollama (Local)	You	Free	Hardware-dependent	2-4/5	5/5	Fully offline, maximum privacy
GLM-5	Zhipu AI	$	Fast	4/5	Cloud	Chinese tasks
KIMI K2.5	Moonshot	$$	Medium	4/5	Cloud	Long context, Chinese tasks

*DeepSeek is a Chinese company; data processing is subject to Chinese law. Evaluate carefully if you have privacy concerns.

---

Detailed Introductions

Claude (Anthropic)

Claude is one of the most popular models in the OpenClaw community, especially strong in nuanced reasoning and safety scenarios.

# ~/.openclaw/providers/anthropic.yaml
provider:
  name: anthropic
  api_key: "${ANTHROPIC_API_KEY}"
  models:
    primary:
      id: claude-opus-4-6
      max_tokens: 8192
      temperature: 0.7
    fast:
      id: claude-sonnet-4-5
      max_tokens: 4096
      temperature: 0.5
  routing:
    enabled: true
    complexity_threshold: 0.6

Strengths: Exceptional reasoning quality, natural responses, comprehensive safety guardrails, ultra-long context (1M tokens) Weaknesses: Opus pricing is high, higher latency in some regions

GPT (OpenAI)

The GPT-5.2 series still leads in code generation and technical tasks.

# ~/.openclaw/providers/openai.yaml
provider:
  name: openai
  api_key: "${OPENAI_API_KEY}"
  models:
    primary:
      id: gpt-5.2-codex
      max_tokens: 8192
      temperature: 0.7
    fast:
      id: gpt-5.2-mini
      max_tokens: 2048
      temperature: 0.5

DeepSeek

DeepSeek V3 delivers near-top-tier quality at a fraction of the cost, especially for Chinese-language scenarios.

provider:
  name: deepseek
  api_key: "${DEEPSEEK_API_KEY}"
  endpoint: "https://api.deepseek.com"
  models:
    primary:
      id: deepseek-v3
      max_tokens: 8192
      temperature: 0.7

DeepSeek Privacy Notice

DeepSeek is a Chinese company and its API service data processing is subject to Chinese regulations. If you handle sensitive personal data or trade secrets, evaluate carefully. You can run DeepSeek open-source models locally via Ollama instead.

Ollama (Local Models)

Run open-source LLMs on your own machine. Fully offline, zero cost, maximum privacy.

curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.3:70b
ollama pull deepseek-v3:distill-32b

provider:
  name: ollama
  endpoint: "http://127.0.0.1:11434"
  models:
    primary:
      id: llama3.3:70b
      max_tokens: 4096
      temperature: 0.7

Hardware recommendations:

Model Size	Minimum RAM	Recommended GPU
7B params	8 GB RAM	Not required (CPU works)
13B params	16 GB RAM	6 GB VRAM
27-32B params	32 GB RAM	12 GB VRAM
70B params	64 GB RAM	24 GB VRAM

Apple Silicon Users

M1 Pro/Max and newer Macs run Ollama well thanks to unified memory architecture. M4 Max 64GB can smoothly run 70B models with near-cloud-API response speeds.

---

Local vs Cloud Trade-offs

Factor	Local (Ollama)	Cloud API
Privacy	Data never leaves your machine	Sent to third-party servers
Cost	Only electricity	Pay per usage
Speed	Hardware-dependent	Usually faster
Quality	Open-source models slightly lag	Top-tier closed models
Availability	Always available	Depends on network and API status

Hybrid Mode Recommended

Best practice is hybrid mode: use local models for daily chat (zero cost, high privacy), auto-switch to cloud API when high-quality reasoning is needed.

routing:
  strategy: hybrid
  default: ollama
  escalation:
    provider: anthropic
    model: claude-opus-4-6
    trigger:
      - complexity_score: 0.7
      - explicit_request: true

---

Cost Estimates

Based on 100 conversations per day (average 500 input + 1000 output tokens each):

Model	Daily Cost (USD)	Monthly Cost (USD)
Claude Opus 4.6	~$2.25	~$67.50
Claude Sonnet 4.5	~$0.45	~$13.50
GPT-5.2 Codex	~$2.10	~$63.00
GPT-5.2 Mini	~$0.15	~$4.50
Gemini 3 Pro	~$0.53	~$15.75
DeepSeek V3	~$0.08	~$2.40
Ollama (local)	$0	$0

---

Next Steps

After choosing your AI model, it is time to give it a unique personality:

• SOUL.md Personality Config — Design your AI agent's character and behavior
• Connect Channels — If you haven't connected a platform yet
• MasterClass Courses — Deep-dive into advanced features