Stop Token Anxiety: Your $20 Subscription is Enough to Run OpenClaw

Do you ever feel this way:

You clearly love OpenClaw. You know it can handle everything at once—writing code, writing content, running automation, managing schedules, and pulling data. But you just don't dare to use it freely.

Because when you open the comments section, group chats, or Twitter, all you see is:

• "Burned $70 in 24 hours."
• "I'm spending hundreds of dollars a week."
• "Even more exaggerated, directly burning over $1,000 a week."

So you start pondering two things every day:

1. How to save Tokens?
2. Are there cheaper domestic alternatives?

You might even have developed a kind of "AI usage habitual poverty"—every time you want the agent to do a bit more work, your hand stops:

"Maybe... forget it?"

But I want to tell you a very counter-intuitive and satisfying fact:

You don't need to spend time on these things at all. Because your three top-tier models'

$20 subscription packages (Claude / GPT / Gemini)

are already enough to cover your OpenClaw in most normal usage scenarios.

Then you'll likely say:

"Huh? That actually works! If that's the case, how do I do it?"

Don't worry. I won't tell you how to install it yet. First, I'll show you what it looks like once it's installed and running.

Look at the "Other Side" first: What does OpenClaw look like after installation?

Don't rush into the steps. I want you to see a picture first: In your OpenClaw, there isn't just one model, but a whole "model arsenal."

Figure 01: My model arsenal

My own configuration looks something like this—

• Master Agent: I named it Friday (this is just my "Master Agent name")
• Default Master Model: I generally use GPT 5.2 as the master
• Model Pool (you can think of it as: switch to whichever you want to use):

1. GPT Series: GPT 5.2, GPT 5.2 Codex, GPT 5.3 Codex
2. Claude Series: Claude Opus 4.6, Opus 4.5, Sonnet 4.5
3. Gemini Series: Gemini 3 Pro, Gemini 3 Flash

More importantly: this isn't just "I configured a bunch of models with nice names."

It's that you can actually open multiple Sessions in OpenClaw and let them each do their own work.

For example, I'll open three extra Sessions:

• Gemini 3 Pro Session: Log in via OAuth, using my Gemini $20 subscription package. You can see a million-level context window in Usage.
• Claude Opus 4.6 Session: Uses the Claude Pro subscription token. 400,000 context window.
• GPT Session: Also OAuth login, 400,000 context window, use it whenever you want.

Figure 02: Using Claude, GPT, and Gemini simultaneously in Telegram

Seeing this, you'll basically start "drooling":

It turns out I can use whichever model I want in the same OpenClaw, without spending an extra penny on API Tokens, and without having to do mental preparation while staring at Token bills every day.

Okay. Do you want to install it like this too? You do, right? Then let me teach you. Next, we'll do a hands-on configuration in three steps:

1. Connect Claude first
2. Then connect GPT
3. Finally connect Gemini

Every step is practical; just follow along and it will work.

Before you start: You only need to prepare these three things

To make sure you don't get stuck, let me clarify the minimum preparation list:

1. You have already installed and can run OpenClaw.
2. You can use npm (we'll need to install several official CLI tools later).

After configuration, we'll use this command to verify everything:

1openclaw models list

If you see the models you selected in the model list, it's working.

Step 1: Connect Claude Pro subscription to OpenClaw

You need to do four things now:

Install Claude Code → Generate Token → Configure in OpenClaw → Verify

1. Install Claude Code CLI

1curl -fsSL https://claude.ai/install.sh | bash

1. Generate Claude Token

1claude setup-token

It will automatically open the browser and jump to the Claude website for authorization. After authorizing, close the browser, return to the terminal, and you will get a Token string. Save it well.

1. Configure Claude in OpenClaw

1openclaw configure

In the wizard:

• Select Anthropic as the provider
• Paste the token
• Select default
• Check the models you want to use (e.g., Opus / Sonnet)

1. Verify

1openclaw models list

You should see Claude models appearing in the list.

Is there a high risk in connecting Claude?

You might ask:

"Will I get banned for using it this way?"

I'll give you the conclusion directly:

• The connection method itself is compliant.
• Anthropic may trigger risk control for certain regions/accounts.

But this has nothing to do with whether you use OpenClaw; even if you use the Claude website normally, some people still get banned. This is an Anthropic platform strategy. So I always suggest: use Claude as an enhancement, and GPT / Gemini as a stable base.

Step 2: Connect GPT Plus subscription to OpenClaw

You do three things now:

Install Codex CLI → Go through OAuth in OpenClaw → Select models → Verify

1. Install Codex CLI

1npm i -g @openai/codex

1. Configure GPT in OpenClaw (Note: Look for OpenAI Codex)

1openclaw configure

One of the easiest pitfalls here is:

You should choose the

OpenAI Codex

path (which opens the browser for OAuth), not the path that says "let me fill in the API Key."

After successful browser verification, return to the terminal to continue.

1. How to choose models

I usually choose:

• GPT 5.2
• GPT 5.2 Codex
• GPT 5.3 Codex

1. Verify

1openclaw models list

If you see OpenAI models in the list, it's OK.

Step 3: Connect Gemini Pro subscription to OpenClaw

You do three things now:

Install Gemini CLI → Select Gemini CLI Auth in OpenClaw → Select models → Verify

1. Install Gemini CLI

1npm install -g @google/gemini-cli

1. Configure Gemini in OpenClaw (Gemini CLI Auth)

1openclaw configure

Select Google as the provider, and for the verification method, choose: Google Gemini CLI Auth

Follow the prompts to complete the browser login.

1. How to choose models

I will check:

• Gemini 3 Pro
• Gemini 3 Flash

1. Verify (I suggest you do both parts; the experience will feel very "successful")

A. Terminal verification:

1openclaw models list

B. Telegram verification:

Go to Telegram and send to your OpenClaw Bot:

• /models

You will see it list the available models.

Then switch to Gemini Pro, ask a random question, and if it replies normally, it's fully working.

Okay, all three are connected: You finally don't have to live your life staring at Tokens

By now, you have connected all three major subscription packages to OpenClaw. Your status should now be:

• Claude / GPT / Gemini, use whichever you want.
• GPT / Gemini are rock solid.
• Claude might have risk control, but it doesn't affect your overall workflow.

You'll find that what you really need to optimize was never "saving Tokens." It's:

How to turn OpenClaw into your productivity exoskeleton and continuously produce value for you.

Let me tell another true story. I have a friend who is a heavy user of ChatGPT's Deep Research. He spends $200 a month on a high-level subscription just for unlimited use of Deep Research. After he installed OpenClaw, he thought he still had to pay extra for API Tokens.

I told him: "Just configure your subscription package directly."

After he finished the configuration, he said to me emotionally: "This $200 is finally being spent where it counts."

So finally, I want to say: Stop worrying about Tokens! Make full use of your subscription package and squeeze out every penny of value.

What we should really care about is how to let OpenClaw create greater value for us, rather than staring at the balance with a racing heart every day.

Spend your time where it's worth it!

As for how to choose models and which one is most suitable for which scenario—if everyone is interested, I can write another dedicated post (or make a video) to talk to you in detail.

I am Kai on AI.

If this post helped you, give me a follow, and we'll continue next time.

Tell me in the comments: Which model do you use most often?