What Is Token Padding? How to Detect It in AI APIs

Have you ever wondered why the same prompt sent to different services shows very different token counts on your bill? You may not be imagining it. Some AI API relays use "token padding" to report false numbers in the usage field, so you pay more. This article explains how the problem works, how to recognize it, and how to detect it automatically with tools.

What Is Token Padding?

Token padding (also called token inflation) is when an AI API relay, after calling the upstream model, modifies the usage object it returns, inflating the actual consumed token counts before passing them back to you:

// Actual response from the upstream model
{
  "usage": {
    "prompt_tokens": 120,
    "completion_tokens": 80,
    "total_tokens": 200
  }
}

// Response after the relay tampered with it and returned it to you
{
  "usage": {
    "prompt_tokens": 156,  // inflated by 30%
    "completion_tokens": 104,  // inflated by 30%
    "total_tokens": 260
  }
}

Because most developers do not check usage line by line, minor padding (5–15%) is rarely noticed. But at a monthly volume of several million tokens, the extra cost adds up considerably.

Token Padding vs Model Substitution: Which Is Harder to Spot?

Model substitution is commonly called "jiangzhi" (降智, capability downgrading) in Chinese-speaking communities. For how to tell, see What Does "Jiangzhi" (Capability Downgrade) Mean? Three Detection Methods.

How Large Is Typical Padding?

Based on public research and community reports, the size of token padding varies widely:

• Minor (5–10%): the most common; rarely noticed, but significant over time
• Moderate (10–20%): requires careful bill checking to detect
• Severe (30% or more): the difference is obvious even in a single request

Note: Illustrative calculation

If your monthly usage is 5 million prompt_tokens (GPT-4o billed at $2.50 per 1M tokens), 15% padding means you overpay by 750K tokens × $2.50 = about $1.875 per month. At scale (50 million tokens per month), that is a hidden loss of about $18.75.

How to Check for Token Padding Manually

Method 1: Send a prompt of known length

Use tiktoken or another tokenizer to count the exact tokens in your prompt, then compare with the usage.prompt_tokens the API returns:

# Python example (requires tiktoken)
import tiktoken, openai

enc = tiktoken.encoding_for_model("gpt-4o")
prompt = "What is 2+2?"
expected_tokens = len(enc.encode(prompt))

resp = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": prompt}]
)

actual = resp.usage.prompt_tokens
diff_pct = (actual - expected_tokens) / expected_tokens * 100
print(f"expected: {expected_tokens}, reported: {actual}, gap: {diff_pct:.1f}%")

Method 2: Detect automatically with BazaarLink Probe

Manual counting is slow and requires knowing each model's tokenizer differences. A faster option is BazaarLink Probe, which automatically runs a set of standardized tests, including token count comparison, and outputs a 0–100 score with a detailed report.

• Enter your API endpoint and key
• Probe automatically sends prompts with precisely calculated token counts
• Compares the returned usage against expected values
• Gap exceeds the threshold → flagged as "Token padding risk"
• Also checks for model substitution, system prompt injection, and other risks

Choosing an AI API Provider: Billing Transparency Compared

The most fundamental way to avoid token padding is to choose a provider that connects directly to the official vendor, or a platform with clear, transparent billing. Here is a comparison of several major options:

Technical Background on Token Padding

LLM supply chain attacks have drawn ongoing attention from academia and security researchers in recent years. The paper Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain (ACM CCS 2026) systematically classifies relay attack techniques, including:

• AC-1 Payload Injection: model substitution
• AC-1.b Conditional Injection: conditional system prompt injection
• Token Inflation: token padding (tampering with the usage field)
• AC-2 Secret Exfiltration: key theft risk

BazaarLink Probe's design draws on this research framework and covers automated detection for each of the attack categories above.

Conclusion

Token padding is a hidden way for AI API relays to profit. It is hard to see with the naked eye, but automated tools can detect it systematically. Whether you are an individual developer or an enterprise user, billing transparency and verifiability are key considerations when choosing an AI API provider.

Go to BazaarLink Probe now to test the endpoint you are currently using and check for token padding or other supply chain risks.

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API