Score how token-efficient your prompt is
Every token in your prompt is billed on every single call. A bloated system prompt with filler phrases, repeated instructions, and decorative padding quietly inflates your bill across millions of requests. This tool gives your prompt a 0–100 efficiency score and shows exactly where the waste is.
How the score works
The scorer estimates token count for your chosen tokenizer, then applies penalties for the patterns that waste tokens without adding information:
- Filler and hedging — words like “please”, “kindly”, “just”, “really”, “very”, “in order to”, “as you can see”.
- Repetition — repeated 3-word phrases that say the same thing twice.
- Padding — excessive whitespace, runs of punctuation, and decorative separators that cost tokens but carry no meaning.
A lean, instruction-dense prompt scores near 100; a verbose one drops fast.
Where the waste usually hides
Most prompts bleed tokens in a predictable set of places. Understanding them is the fastest route to a better score.
System prompt headers and dividers. Patterns like ====== INSTRUCTIONS ====== or ### TASK ### are purely decorative. The model does not need headers to read a prompt; they exist for human readability in code. Replace them with a blank line or nothing.
Polite framing. Opening with “I need you to please carefully” instead of just stating the instruction directly wastes several tokens before any information is conveyed. LLMs respond identically to direct instructions regardless of politeness.
Restated rules. A constraint stated once is obeyed. Restating “always respond in English, respond only in English, do not respond in any other language” triples the token cost of one rule with no added reliability.
Long chain-of-thought instructions. “Think step by step, reason through each part carefully, then provide your final answer after full consideration” can often become think step by step with identical effect.
Worked example
Consider this 52-word instruction:
“Please make sure that you always respond in a professional and formal manner. It goes without saying that you should be polite and considerate. In order to provide the best possible assistance, please carefully think through your response before outputting it.”
Scored: roughly 38/100 due to filler (“it goes without saying”), hedging (“please”, “make sure”), and redundant framing (“best possible”). Trimmed version:
“Respond formally and professionally. Think before answering.”
That is 7 words instead of 52, with no change in model behaviour on most tasks.
Tips to raise your score
- Replace “in order to” with “to”, “due to the fact that” with “because”.
- State each rule once. Models do not need repeated emphasis to obey an instruction.
- Cut polite filler — “please”, “thank you”, “if you could” — it adds tokens and changes nothing in the output.
- Use compact formatting. Long ASCII dividers and decorative headers are pure token cost. Keep structure minimal.
- If you have a system prompt sent on every request, a score of 70+ means your token cost per million calls is meaningfully lower than a score of 40.