GPT-5.4 + OpenClaw: Complete Setup Guide for Production (2026)

GPT-5.4 + OpenClaw: Complete Setup Guide for Production (2026)

OpenAI just shipped GPT-5.4. It's their most capable model yet — 1M token context, built-in computer use, and 33% fewer errors than GPT-5.2.

This guide walks through setting up OpenClaw with GPT-5.4 from initial configuration to production deployment.

Why GPT-5.4 for OpenClaw

GPT-5.4 was built for agentic work. Here's what matters for OpenClaw users:

• 1M token context window — Analyze entire codebases, long document collections, or extended agent trajectories in a single request
• Built-in computer use — First mainline OpenAI model that can interact directly with software in a build-run-verify-fix loop
• 33% fewer errors — Individual claims are 33% less likely to be wrong compared to GPT-5.2
• Token efficiency — Accomplishes identical tasks while consuming substantially fewer tokens than earlier models
• Compaction support — Trained to compress long agent trajectories while preserving key context

For most OpenClaw use cases, GPT-5.4 provides an excellent balance of capability and cost. GPT-5.4 Pro is the higher-end option for complex enterprise workflows where maximum accuracy justifies the price.

Prerequisites

Before starting:

1. OpenAI API key — Get one at platform.openai.com
2. OpenClaw installed — Follow the official installation guide
3. Node.js 22+ — Required for OpenClaw

Step 1: Configure Your API Key

Add your OpenAI API key to the OpenClaw configuration.

Option A: Environment variable (recommended)

export OPENAI_API_KEY="sk-..."

Option B: Auth profile via CLI

openclaw configure --section model

The interactive wizard will prompt for your OpenAI credentials. You can also set it directly:

openclaw config set auth.profiles.openai:default '{"provider":"openai","mode":"api_key"}' --strict-json

Option C: Configuration file

In your ~/.openclaw/openclaw.json:

{
  env: {
    OPENAI_API_KEY: "sk-..."
  }
}

Security note: For production, use environment variables or a secrets manager. Don't commit API keys to version control.

Step 2: Select Your Model

Configure GPT-5.4 as your primary model:

{
  agent: {
    workspace: "~/.openclaw/workspace",
    model: {
      primary: "openai/gpt-5.4"
    }
  }
}

Or via the CLI:

openclaw config set agent.model.primary "openai/gpt-5.4"

Model options:

Recommendation: Start with openai/gpt-5.4. Upgrade to Pro only if you need maximum accuracy on high-stakes tasks. The standard model handles the vast majority of agent workloads.

Want fallbacks? Set a fallback model in case OpenAI has an outage:

{
  agent: {
    model: {
      primary: "openai/gpt-5.4",
      fallbacks: ["anthropic/claude-sonnet-4-5"]
    }
  }
}

Step 3: Configure Your Agent

Set your agent's identity and workspace:

{
  identity: {
    name: "Atlas",
    theme: "senior software engineer",
    emoji: "🤖"
  },
  agent: {
    workspace: "~/.openclaw/workspace",
    model: { primary: "openai/gpt-5.4" }
  }
}

OpenClaw uses bootstrap files in your workspace to shape agent behavior:

• AGENTS.md — Operating instructions and memory
• SOUL.md — Persona, boundaries, tone
• TOOLS.md — Tool usage notes and conventions

Edit these files to define your agent's personality and expertise. They're injected into the agent context at the start of every session.

Step 4: Configure Tools

Control which tools your agent can use:

{
  tools: {
    allow: ["exec", "read", "write", "edit"],
    deny: ["browser"]
  }
}

Or via the CLI:

openclaw config set tools.allow '["exec","read","write","edit"]' --strict-json
openclaw config set tools.deny '["browser"]' --strict-json

Security principle: Only enable tools your agent actually needs. Disable everything else.

GPT-5.4's built-in computer use capabilities make it especially effective at tool-heavy workflows — it can interact with software directly, verify results, and fix issues in a loop. But more capability means tighter controls matter even more.

Step 5: Set Up Model Allowlist

Control which models are available for switching via the /model command:

{
  agents: {
    defaults: {
      models: {
        "openai/gpt-5.4": { alias: "gpt" },
        "openai/gpt-5.4-pro": { alias: "gpt-pro" },
        "anthropic/claude-sonnet-4-5": { alias: "sonnet" }
      }
    }
  }
}

This lets you switch between models on the fly without editing config.

Step 6: Test Locally

Start the OpenClaw gateway:

openclaw gateway run

Test with a message:

openclaw message send "What model are you running?"

Expected response should confirm GPT-5.4.

Check gateway health:

openclaw status

Verify tools work by asking the agent to list files or run a command.

Step 7: Production Configuration

For production, lock down the gateway:

{
  gateway: {
    mode: "local",
    port: 18789,
    bind: "loopback",
    auth: {
      mode: "token",
      token: "your-secure-gateway-token"
    },
    controlUi: { enabled: false }
  },
  agent: {
    workspace: "~/.openclaw/workspace",
    model: {
      primary: "openai/gpt-5.4",
      fallbacks: ["anthropic/claude-sonnet-4-5"]
    }
  },
  channels: {
    whatsapp: {
      allowFrom: ["+15555550123"]
    }
  }
}

Key production settings:

• Bind to loopback only — no external access without a reverse proxy
• Token auth on the gateway — prevents unauthorized API access
• Disable Control UI — reduces attack surface
• Set allowFrom on channels — only authorized senders can interact

Common Issues

"Model not found" Error

Model refs use provider/model format (split on first /):

• ✅ openai/gpt-5.4
• ❌ gpt5.4
• ❌ openai:gpt-5.4

"Invalid API key" Error

Check that:

1. Your key starts with sk-
2. The key has not been revoked
3. The environment variable is set: echo $OPENAI_API_KEY
4. Or your auth profile is configured: openclaw config get auth

High Latency

If responses are slow:

• Check if you're sending >272K input tokens (higher-tier pricing and potentially different routing)
• Reduce context size by limiting workspace scope
• Use openai/gpt-5.4 instead of openai/gpt-5.4-pro for faster responses
• Check your network connection to OpenAI's API

Token Limit Errors

If you hit token limits:

• GPT-5.4 supports up to 1M context — increase input limits if analyzing large codebases
• Consider breaking tasks into smaller steps for cost efficiency
• Use the /model command to switch to a cheaper model for simpler tasks

Cost Management

Estimated costs (as of March 2026):

Cost optimization tips:

1. Use GPT-5.4 by default — Pro only when you need maximum accuracy
2. Stay under 272K input tokens — Pricing doubles above that threshold
3. Leverage token efficiency — GPT-5.4 uses fewer tokens than GPT-5.2 for equivalent tasks
4. Use cached input — Repeated prompts cost only $0.25/M (10x cheaper)
5. Set fallbacks to cheaper models — Use Claude Sonnet or Gemini Flash as fallbacks
6. Monitor usage — Check OpenAI dashboard regularly

With Clawctl

If you're using Clawctl, GPT-5.4 configuration is simpler. Sign up at clawctl.com/checkout, and the dashboard setup wizard walks you through it:

1. Select OpenAI as your provider
2. Paste your API key
3. Clawctl validates the connection and deploys

No config files. No environment variables. Just pick your model and go.

Clawctl handles:

• Secure credential storage (AES-256-GCM encrypted)
• Production gateway security
• Token usage monitoring
• Cost alerts

Deploy GPT-5.4 + OpenClaw securely →

Full OpenClaw documentation →