Using the GLM-5.3 LLM for Free: TokenRouter Setup & Gotchas Guide

TokenRouter recently rolled out a limited-time free model endpoint: z-ai/glm-5.3-free. You can get it up and running in minutes—just sign up, generate an API key, and configure your client. It comes with an 8 RPM (requests per minute) rate limit and a 1M token context window, but keep in mind that this is not permanently free.

What Is This?

TokenRouter is a third-party aggregation gateway. The endpoint tagged with -free is not Zhipu's official paid tier; it's a promotional channel provided by the gateway running on shared compute resources.

The base GLM-5.3 model is Zhipu's flagship model released in August 2026, featuring a 744B MoE architecture and a 1M context window. The official price sits around $1.40 / 1M input tokens and $4.40 / 1M output tokens. While Zhipu offers official free tiers (like GLM-4.7-Flash and 4.5-Flash), GLM-5.3 is not officially permanently free. Therefore, pricing on this gateway route could change at any time, though it is currently $0.

Current status:

• Price: $0 (subject to change at any time)
• Rate limit: 8 requests/min (8 RPM)
• Context window: 1M tokens
• SLA: None
• Reliability: Shared compute; queuing during peak hours, with relatively high Time-To-First-Token (TTFT)

It's great for learning, prototyping, and lightweight Q&A. However, avoid using it for long-running autonomous agent workflows or core production pipelines—the concurrency is too limited. Also, double-check the model pricing page before using it to ensure it is still listed at $0.

Registration and API Key Setup

1. Go to tokenrouter.com and sign up with your email.
2. In the left navigation bar, click API Keys → Create New Key.
3. Enter any name you like.
4. In the Allowed Models field, make sure to manually enter and lock it exclusively to z-ai/glm-5.3-free (the most critical step).
5. Enable Unlimited Quota.
6. Once created, immediately copy the key starting with sk-. You won't be able to view it again after closing the dialog.

Required Parameters

Common pitfalls:

• Omitting the z-ai/ prefix → Returns a 404 or "Model not found" error.
• Mistyping it as glm-5.3-flash → Might route through paid endpoints.
• Appending /chat/completions to the Base URL → The client will append it again, triggering a 404.
• Not locking Allowed Models on the API key and selecting a paid model in your client dropdown → Silently burns your balance.

Code Example

from openai import OpenAI
 
client = OpenAI(
    api_key="sk-xxx",
    base_url="https://api.tokenrouter.com/v1"
)
 
response = client.chat.completions.create(
    model="z-ai/glm-5.3-free",
    messages=[{"role": "user", "content": "Explain the MoE architecture in one sentence."}],
    max_tokens=512
)
 
print(response.choices[0].message.content)

You only need to override api_key and base_url; everything else works exactly like standard OpenAI SDK calls. The Anthropic /v1/messages format is supported as well.

Handling rate limits with a simple retry mechanism:

import time
import random
from openai import OpenAI, RateLimitError
 
client = OpenAI(api_key="sk-xxx", base_url="https://api.tokenrouter.com/v1")
 
def chat_with_retry(prompt, max_retries=3):
    for i in range(max_retries):
        try:
            return client.chat.completions.create(
                model="z-ai/glm-5.3-free",
                messages=[{"role": "user", "content": prompt}]
            ).choices[0].message.content
        except RateLimitError:
            wait = (2 ** i) + random.uniform(0, 1)
            print(f"Rate limited. Retrying in {wait:.1f}s...")
            time.sleep(wait)
    return None

Client Configuration

Claude Code (CLI)

export ANTHROPIC_BASE_URL="https://api.tokenrouter.com/v1"
export ANTHROPIC_API_KEY="sk-xxx"

After launching, set the model name to z-ai/glm-5.3-free.

Avoiding Hidden Charges with Free Keys

This is the easiest trap to fall into.

If you don't lock Allowed Models on your API key and your client defaults to the paid z-ai/glm-5.3, you might think you are using the free tier while racking up charges on every single request.

To stay safe, follow these best practices:

1. Explicitly lock Allowed Models to z-ai/glm-5.3-free in the dashboard.
2. Manually enter the model name in your client rather than picking from a dropdown.
3. Periodically check the model pricing page to verify the price remains $0.

Is It Worth Using Long-Term?

It works fine as a free sandbox, but not as production infrastructure.

The raw model capability is impressive, and the 1M context window is great for long documents. However, due to shared compute, lack of an SLA, and the 8 RPM rate limit, requests can lag during peak traffic, and continuous agent workflows will quickly hit the ceiling. Furthermore, traffic passes through a third-party gateway, so sensitive data should not be sent here.

A more reliable free alternative is Zhipu's official GLM-4.5-Flash / 4.7-Flash, which is permanently free and backed by official uptime guarantees.

You can also implement simple fallback routing in code: route requests to TokenRouter's free tier first, and automatically fall back to local Ollama instances or official Flash models when encountering 429s to keep tasks running smoothly.

Frequently Asked Questions (FAQ)

How does this differ from the official glm-5.3-flash?
This endpoint offers TokenRouter's limited-time, full-parameter model—delivering higher capabilities, but without an official SLA and with the possibility of price changes down the road. The official Flash tier is a lightweight, permanently free model with slightly lower capabilities but higher stability.

Getting a 401 even though the model name is correct?
Ensure your API key is copied completely (starts with sk-) and that the Base URL doesn't contain extra slashes or subpaths. Deleted keys or exhausted quotas can also result in 401 errors.

Is 8 RPM sufficient?
It is plenty for standard chat and lightweight testing. However, continuous agent tasks (e.g., coding → running → debugging loops) can saturate the limit within seconds. We recommend adding retry logic or throttling concurrency.

Are there other free models available?
The list changes over time. Check the dashboard for models tagged with a free suffix. Other aggregators offer similar routes, but reliability and compliance vary significantly.

Is it still active?
As of September 5, 2026, it remains at $0, though no end date has been announced. Always check the current unit price before using it.

If you hit 429s, 401s, or "model not found" errors during setup, feel free to share your error logs for troubleshooting.