System prompt budget checker
Flag long system prompts and estimate their cost per call.
FreeRuns in your browserNo signupUpdated 2026-10-29
Rates last checked October 2026. Input prices only; the user message and reply are extra. Prices change, so check the provider's pricing page.
System prompt size
108 tokens · Lean
Under 1,000 is lean; over 4,000 is worth trimming or caching
Cost per month
$0.65
$0.00001 per call, sent 60,000 times a month
| Check | Result |
|---|---|
| Rules or bullet points | 5 |
| Absolute words (always, never, must, do not) | 2 |
| Duplicate lines | None |
| Share of a 1M-token context window | 0.01% |
| Model | Tokens | Per call | Per month |
|---|---|---|---|
| GPT-6 Astra | 108 | $0.00108 | $64.8 |
| GPT-6.1 Sol | 108 | $0.00022 | $12.96 |
| GPT-6.1 Terra | 108 | $0.00011 | $6.48 |
| GPT-6 Luna | 108 | $0.00001 | $0.65 |
| Claude Fable 5.1 | 151 | $0.00151 | $90.6 |
| Claude Opus 5.5 | 151 | $0.0006 | $36.24 |
| Claude Sonnet 5.5 | 151 | $0.0003 | $18.12 |
| Claude Haiku 5.5 | 151 | $0.00002 | $0.91 |
| Gemini 4 Argon | 108 | $0.00022 | $12.96 |
| Gemini 3.8 Flash | 108 | $0.00008 | $4.86 |
| Gemini 3.5 Flash-Lite | 108 | $0.00003 | $1.94 |
OpenAI and Gemini rows use the o200k_base tokenizer, a close estimate for GPT-6 and Gemini. Claude counts are an estimate at 3.5 characters per token. Prompt caching on OpenAI and Anthropic cuts the repeated system prompt cost sharply once it passes the provider's minimum cacheable length.
About this tool
How it works and how to read the result.
Need this done for you?
AI agents & chatbots
AI assistants that answer, act and hand over when needed.
See the serviceWhat it does
System prompt length is a cost you pay on every call, so this checker makes it visible. Paste the prompt, choose a model and enter calls per day. The page counts the tokens, labels the prompt lean, long or very long, and shows what the prompt alone costs per call and per month at the model's input price. It also runs simple checks: how many rules the prompt contains, how many absolute words it uses, and whether any lines are duplicated. The count runs in your browser.
How to use it
- Paste the full system prompt as your code sends it, including any policy text.
- Pick the model it runs on and enter a realistic calls-per-day figure.
- Read the size badge and the monthly cost.
- Look at the checks table for duplicate lines and a high rule count.
- Trim, then paste again to compare. Copy the summary for your notes.
How to read the result
Take a 1,650-token system prompt for a clinic assistant on GPT-6 Luna, sent 2,000 times a day. At $0.10 per million input tokens, the prompt costs $0.000165 per call and about $9.90 a month. That is modest. If the same prompt measured 1,650 tokens on Claude Sonnet 5.5 at $2 per million, it would cost about $198 a month before the user's message or the reply are counted.
The badge turns to Long above 1,000 tokens and Very long above 4,000. Long is not wrong; it is a prompt worth caching. Very long usually means policy documents or examples are pasted in full, and those are better retrieved when relevant.
Twelve or more absolute words, or several duplicate lines, point to a prompt that has been patched repeatedly. Rewriting it as a short list of grouped rules often improves behaviour as well as cost.
We review prompts like this when building AI agents. Start a cleaner version with the prompt template builder.
Questions
Frequently asked.
How long should a system prompt be?
As short as it can be while still fixing the behaviour you care about. Under 1,000 tokens is lean for most assistants. Between 1,000 and 4,000 is normal for bots with detailed policies. Above that, look for content that belongs in retrieved documents rather than in every call.
Why does the page count absolute words?
Words like always, never and must signal rules. A handful are fine. Dozens suggest a prompt that grew by patching one incident at a time, and models follow a long list of rules less reliably than a short one. Group related rules and drop any that testing shows are no longer needed.
Can prompt caching remove this cost?
Mostly. OpenAI and Anthropic both discount input tokens that repeat from a recent request, and a system prompt is the textbook case. The discount only applies above a minimum prompt size and for a limited time between calls, so check each provider's rules before you rely on it.
More in AI & automation
Related tools.
Your next move
Like what you see? Tell us about it.
A process that takes too long, a website that undersells you, or an idea ready to take shape. Tell us where you are and we will suggest a useful next step.
India · Working with clients worldwide