Setting Up OpenAI (GPT-5.2) with OpenClaw — Complete Guide

TL;DR: Get an API key from OpenAI, add it to your OpenClaw config with the openai-responses provider type, pick a model, restart. Takes 10 minutes. Expect $10-40/month in API costs for moderate use.

Estimated time: 10-15 minutes
Difficulty: Beginner

Why OpenAI with OpenClaw?

OpenAI makes some of the most capable models available. GPT-5.2 is their latest flagship — excellent at reasoning, coding, and general conversation. Combined with OpenClaw, you get GPT-5.2 as a persistent assistant across all your messaging channels.

You could just use ChatGPT directly, of course. But OpenClaw gives you things ChatGPT doesn't: memory across sessions, integration with your tools and services, multi-channel access, and control over exactly how your assistant behaves. More on this below.

Step 1: Get Your API Key

1. Go to platform.openai.com
2. Sign up or log in (this is separate from your ChatGPT account — you need a platform account)
3. Navigate to Settings → API keys (or go directly to platform.openai.com/api-keys)
4. Click "Create new secret key"
5. Name it something memorable like "OpenClaw" and click Create
6. Copy the key immediately — OpenAI only shows it once. It starts with sk-

Add Credits

OpenAI's API is prepaid. You need to add credits before it works:

1. Go to Settings → Billing (platform.openai.com/settings/organization/billing)
2. Add a payment method
3. Add credits — $10 is a good starting amount for testing and initial use
4. Optionally, set up auto-recharge so you don't run dry mid-conversation

Set a Usage Limit

While you're in billing, set a monthly limit so you never get a surprise bill:

1. Go to Settings → Limits
2. Set a monthly budget (e.g., $30)
3. Set a notification threshold (e.g., $20)

This is genuinely important. Without a limit, a misconfigured cron job or runaway conversation can burn through credits fast.

Step 2: Configure OpenClaw

OpenClaw supports OpenAI through the Responses API (openai-responses), which is OpenAI's latest API format. This is the recommended setup.

config.json

{
  "providers": {
    "openai": {
      "type": "openai-responses",
      "apiKey": "sk-proj-xxxxxxxxxxxxxxxxxxxxxxxxxxxx"
    }
  },
  "defaultModel": "openai/gpt-4.1"
}

config.yaml

providers:
  openai:
    type: openai-responses
    apiKey: "sk-proj-xxxxxxxxxxxxxxxxxxxxxxxxxxxx"

defaultModel: "openai/gpt-4.1"

Using Environment Variables (Recommended)

Don't put API keys directly in config files — especially if your config is in a git repo.

export OPENAI_API_KEY="sk-proj-xxxxxxxxxxxxxxxxxxxxxxxxxxxx"

Then in your config:

{
  "providers": {
    "openai": {
      "type": "openai-responses",
      "apiKey": "${OPENAI_API_KEY}"
    }
  },
  "defaultModel": "openai/gpt-4.1"
}

Step 3: Choose Your Model

OpenAI has several models. Here's what matters for OpenClaw:

Available Models (Early 2026)

Recommendation for most people: Start with gpt-4.1. It's capable, fast, and reasonably priced. Switch to gpt-4.1-mini if you want to save money, or gpt-5.2 if you want the newest and best.

Setting Your Model

{
  "defaultModel": "openai/gpt-4.1"
}

Step 4: Reasoning Mode Setup

OpenAI's o-series models (o3, o4-mini) use "reasoning" — they think step-by-step before answering. This produces better results for complex tasks but costs more and takes longer.

Configuring Thinking Levels

You can control how much thinking the model does:

{
  "providers": {
    "openai": {
      "type": "openai-responses",
      "apiKey": "${OPENAI_API_KEY}",
      "thinking": {
        "enabled": true,
        "budgetTokens": 10000
      }
    }
  },
  "defaultModel": "openai/o4-mini"
}

Thinking Budget Options

Cost note: Thinking tokens are billed as output tokens. A 10,000-token thinking budget on o3 could add $0.08 per response on top of the visible output. It adds up.

When to Use Reasoning Models

• ✅ Math and logic problems
• ✅ Complex code debugging
• ✅ Multi-step analysis
• ✅ Tasks where accuracy matters more than speed
• ❌ Simple conversation (use gpt-4.1 instead)
• ❌ Quick questions (reasoning adds unnecessary latency)
• ❌ High-frequency tasks like cron jobs (gets expensive fast)

Step 5: Restart and Test

clawdbot gateway restart

Check the logs:

clawdbot gateway logs

Send a message to your bot. You should get a response from GPT within a few seconds.

Cost Expectations: Real Numbers

Here's what actual OpenClaw usage costs with OpenAI models:

Daily Cost Estimates

Monthly Estimates

Cost Killers to Watch

1. Long conversations without resets: Context grows, input costs grow with it
2. Reasoning models for everything: Use gpt-4.1 for chat, reasoning models only when needed
3. Cron jobs on expensive models: Route heartbeats and background tasks to gpt-4.1-nano
4. No compaction: Enable context compaction to keep costs linear instead of exponential

See our API costs guide for detailed strategies.

ChatGPT / Codex Subscription vs API Key

This is the question everyone asks: "I already pay $20/month for ChatGPT Plus. Why would I use the API?"

What ChatGPT Plus ($20/month) Gets You

• ChatGPT web/mobile app access
• GPT-4o and GPT-5.2 in the app
• DALL·E image generation
• Rate limits (generous but they exist)
• ChatGPT-specific features (memory, custom GPTs, canvas)

What ChatGPT Plus Does NOT Get You

• API access (that's a separate product)
• Custom system prompts
• Integration with your tools and services
• Multi-channel messaging (Telegram, WhatsApp, Discord)
• Persistent memory that you control
• Fine-grained model selection per task
• Background automation (cron, heartbeats)

What Codex ($200/month) Gets You

• Cloud-based coding agent
• Works on GitHub repos
• Parallel task execution
• Powered by codex-1 (o3-derived model)

What OpenClaw + API Key Gets You

• Everything above in the "Does NOT" list
• Pay-per-use instead of flat fee (cheaper for light users, more expensive for very heavy users)
• Choice of any model (not just what ChatGPT offers)
• Total control over your assistant's behavior
• Can run alongside ChatGPT — they're not mutually exclusive

The Math

For light-to-moderate users, the API is usually cheaper. For heavy users, ChatGPT's flat rate wins on cost — but you lose the integration and control that OpenClaw provides.

Many people use both: ChatGPT for casual browsing and quick questions, OpenClaw for their persistent AI assistant workflow.

Multiple OpenAI Models in One Config

You can configure several models and switch between them:

{
  "providers": {
    "openai": {
      "type": "openai-responses",
      "apiKey": "${OPENAI_API_KEY}",
      "models": [
        { "id": "gpt-5.2", "name": "GPT-5.2 (Latest)" },
        { "id": "gpt-4.1", "name": "GPT-4.1 (Balanced)" },
        { "id": "gpt-4.1-mini", "name": "GPT-4.1 Mini (Budget)" },
        { "id": "o4-mini", "name": "o4-mini (Reasoning)" }
      ]
    }
  },
  "defaultModel": "openai/gpt-4.1",
  "cron": {
    "defaultModel": "openai/gpt-4.1-nano"
  }
}

This uses gpt-4.1 for normal chat, gpt-4.1-nano for background tasks, and gives you access to other models when needed.

Common Issues

"Incorrect API key provided"

• Make sure the key starts with sk- (project keys start with sk-proj-)
• Check for extra whitespace or newlines
• Verify the key hasn't been revoked in the OpenAI dashboard
• If using env vars, make sure the variable is exported and available to the OpenClaw process

"You exceeded your current quota"

You've run out of credits. Add more at platform.openai.com/settings/organization/billing.

"Rate limit reached"

You're sending too many requests too fast. OpenAI has per-minute rate limits that vary by tier. Solutions:

• Wait a minute and try again
• Upgrade your usage tier (based on total spend history)
• Use a different model (less popular models have higher limits)

Responses are slow with reasoning models

That's expected. o3 and o4-mini "think" before responding, which adds 5-30 seconds of latency. If you don't need reasoning, switch to gpt-4.1.

Provider type confusion: openai vs openai-responses

• openai-responses — Uses the Responses API (recommended, supports latest features)
• openai — Uses the older Chat Completions API (still works, but fewer features)

Use openai-responses unless you have a specific reason not to.

The Easy Way

Getting an OpenAI API key and adding it to a config file isn't hard — but getting the model selection, thinking levels, and cost optimization right takes some experimentation.

Don't want to manage server infrastructure? lobsterfarm provides managed OpenClaw hosting — deployment, updates, and support handled for you.

Get started with lobsterfarm →