The 70% Waste Problem
To review prompt noise, label passages that add requirements, context or data and passages that do not serve the task. Count tokens with the relevant tokenizer and retain the labeled examples. This article does not establish an average noise share across users.
At GPT-4 pricing ($10 per million input tokens), those 33 noise tokens cost $0.00033 per prompt. Run 1,000 prompts per month and you waste $0.33. That sounds small. Now picture a company with 500 employees, each sending 50 prompts a day. That is 25,000 prompts per day × 33 noise tokens × $10/M = $8.25 per day in pure noise tokens. That adds up to $247.50 per month and $2,970 per year. And that is only the input side.
Input noise also creates output noise. A messy prompt gets a longer, less useful reply. Every noise input token typically produces 3 to 5 noise output tokens. Output tokens cost $30 per million. So the real cost of noise is 3 to 5 times the input cost.
What Noise Costs in Dollars
Scroll sideways to see all table columns.
| Illustrative workload | Example prompts/month | Input cost to review | Output cost to review | Total cost to review |
|---|---|---|---|---|
| Individual | 500 | Not measured | Not measured | Not measured |
| Small team (10) | 5,000 | Not measured | Not measured | Not measured |
| Mid company (200) | 100,000 | Not measured | Not measured | Not measured |
| Enterprise (2,000) | 1,000,000 | Not measured | Not measured | Not measured |
| Large enterprise (10,000) | 5,000,000 | Not measured | Not measured | Not measured |
That is $18,975 per month in wasted tokens for a large enterprise. It comes to $227,700 per year. All of that money buys zero value. And this is at $10/$30 per million. Prices will likely go up as models get better.
Anatomy of a Wasteful Prompt
Here is a real prompt from my data:
"Hey there! I was hoping you could help me out with something. So basically what I need is for you to take a look at our quarterly report and kind of summarize the key findings for me. If you could make it pretty concise that would be great. Thanks so much!"
Total tokens: ~52
Signal tokens: "summarize" + "quarterly report" + "key findings" + "concise" = ~6 tokens
Noise tokens: 46
SNR: 6/52 = 0.115
Here is a better version:
TASK: Summarize this quarterly report. CONSTRAINTS: Maximum 200 words. Only key findings. No commentary. FORMAT: Bullet points. One finding per bullet. DATA: [report text]
Total tokens: ~30 (excluding report text)
Signal tokens: ~27
SNR: 27/30 = 0.90
The revised example states requirements directly instead of relying on a greeting or “pretty concise.” Compare token counts and output quality for the two versions; concision alone does not establish a stronger signal or better answer.
The Reduction Math
Structured prompts cut token usage in 3 ways:
- Input review: remove irrelevant filler and repetition, then check that required context remains.
- Output review: use constraints and format rules to request the needed detail, then measure the returned length and usefulness.
- Retry review: record which responses require correction and test whether clearer instructions address the cause.
Review API costs against the same workload, model settings, and output checks before attributing a change to prompt structure.
Enterprise Token Economics
The AI cost problem is not about model prices. It is about wasted prompts. In my analysis, companies spending $50,000 per month on AI API costs are getting $5,000 worth of signal and $45,000 worth of noise. A cheaper model will not fix that. A cleaner signal will.
I built sinc-LLM to clean the signal. Paste any prompt and get a structured 6-band signal back. The SNR improvement is instant and easy to measure. Lower costs follow right away.
Every noise token you pay for is a tax on unclear writing. The model does not charge extra for noise. It just gives you a worse answer. You end up paying twice: once for the noise tokens, and again when you redo the work because the output was wrong.
// Production AI Engineering
Build AI systems that hold up in production.
sinc-LLM designs, audits, and stabilises production AI infrastructure: from vendor evaluation and cost accountability to incident controls and MCP architecture.
See what we do →
sinc-LLM