A 1200-word system prompt, 200-word query, 150-word reply on GPT-5.5. Estimate the token cost, then name what doubles it.
Free · 2 AI evals / day
Start with the rough numbers. English on o200k_base (GPT-4o, GPT-5, GPT-5.5) tokenizes at about 1.3 tokens per word. So for this ticket:
- Input: (1200 + 200) words * 1.3 = about 1,820 tokens
- Output: 150 words * 1.3 = about 195 tokens
Multiply by published GPT-5.5 input and output prices and you get the per-ticket cost.
Five things can double that cost in production.
**1. Tool schemas.** If the agent has 10 tools with JSON schemas, the schemas are rendered into the prompt on every call. Easily 1,000 to 3,000 extra input tokens per request.
**2. Conversation history.** Multi-turn support tickets re-send prior turns on every call. By turn 5, input tokens are several times the original.
**3. Non-English input.** Hindi, Arabic, or Mandarin user messages tokenize at 2 to 3 tokens per word on o200k_base, well above the English 1.3.
**4. JSON system context.** If the system prompt embeds account data, order history, or preferences as JSON, the keys and punctuation push fertility up. The same content as prose is cheaper.
**5. Structured output.** If you ask for JSON, citations, or chain-of-thought, the reply grows beyond 195 tokens.
Always validate the estimate by running a real production-shape prompt through tiktoken. The 1.3 multiplier is a starting point, not a billing primitive.
**Per-ticket estimate.** English on o200k_base averages roughly 1.3 tokens per word.
- Input: (1200 + 200) * 1.3 = **~1,820 tokens**
- Output: 150 * 1.3 = **~195 tokens**
Multiply by GPT-5.5's published per-million input and output prices. Validate against tiktoken on a real prompt before trusting the number.
**Five doublers in production.**
1. **Tool/function-call schemas.** OpenAI's chat completions API renders declared tools as structured text in every prompt. A non-trivial 10-tool set runs 1,000 to 3,000 tokens. Adding tools to the ticket doubles input cost on its own.
2. **Conversation history.** Multi-turn support flows resend prior turns. Turn 5 input is 5x turn 1 input. Truncation strategies (sliding window, summarization) bound this; without one, costs scale quadratically with turn count.
3. **Non-English user content.** Hindi at roughly 2.5 tokens/word, Arabic at roughly 2 tokens/word, Mandarin at roughly 1.5 tokens/word on o200k_base. A 200-word Hindi message is 500 tokens, not 260.
4. **JSON-formatted system data.** Repetitive keys, brackets, and quotes drive token count above the equivalent prose representation. A flat JSON dump of user context can be 2x the cost of the same content as a paragraph.
5. **Structured or CoT output.** Asking for JSON output, citation blocks, or chain-of-thought reasoning pushes the reply well past 150 words. JSON output specifically can be 2 to 4x prose equivalent, and CoT can 5x the visible answer length.
**Validation discipline.** Pipe a representative production prompt through `tiktoken.get_encoding('o200k_base')` once and use that count as the baseline. Re-validate quarterly because tokenizer behavior can drift with model versions. For cost models that drive budget decisions, sample 200 to 500 real prompts and record actual usage from response.usage rather than estimating from input alone.
Quick reference
TL;DR
Roughly 1,820 input plus 195 output tokens at 1.3 tokens/word on o200k_base. Doublers: tool schemas, conversation history, non-English input, JSON system data, CoT output. Validate with tiktoken.
Think of tokens like syllables a language model has to read or speak. English breaks into roughly 1.3 syllables per word in this model's dictionary. A 1,400-word prompt is about 1,820 syllables in, a 150-word reply is about 195 out. Multiply by the per-syllable price and you get your bill. But the real world adds five things that quietly double the count. Tools the model can call cost extra reading every time. Multi-turn conversations replay everything you said before on every reply. Hindi and Arabic users speak in many more syllables per word than English ones do. Structured data shoved into the prompt as JSON is wordier than the same content as prose. And if you ask the model to think out loud or reply in JSON, the answer balloons. Estimate with the heuristic, then check with the real counter before you trust the number.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Cost modeling for the Gemini family, where billing is per character not per token. Token-based estimates are the wrong unit; character-count drives the bill on Gemini, and the same workload prices differently.
Single-turn, no-tools, English-only batch jobs with no multilingual users. The five doublers do not apply, so the simple 1.3 tokens/word estimate is close to the real cost and the validation effort is overhead.
Early prototyping where exact cost numbers do not yet drive any decisions. Heuristics like 1.3 tokens/word or len/4 are fine for back of the envelope work; the tiktoken validation discipline matters once the estimate gates production decisions.
Quick scan · 2–4 min
Everything important, quickly.
Walk the per-ticket arithmetic: 1.3 tokens/word for English on o200k_base, 1,820 input plus 195 output. Then list and explain the five doublers: tool schemas, conversation history, non-English fertility, JSON-formatted system data, structured or CoT output. Cover the validation discipline of running real prompts through tiktoken and reconciling against response.usage. Close with prompt caching as the production lever for long static system prompts.
Code example
python
import tiktoken
enc = tiktoken.get_encoding('o200k_base') # GPT-4o/GPT-5/GPT-5.5
system_prompt = open('support_system.md').read() # ~1200 words
user_msg = 'My order #12345 has not shipped, can you check?' # ~200 words
assistant_reply = generate_reply(...) # measured at runtime
n_input = len(enc.encode(system_prompt + user_msg))
n_output = len(enc.encode(assistant_reply))
print(f'input={n_input}, output={n_output}')
# Validate against actual API response
response = client.chat.completions.create(...)
print(response.usage.prompt_tokens, response.usage.completion_tokens)
# Drift between tiktoken count and response.usage indicates
# you missed the tool-schema or chat-template overhead.
Real-world examples
Real products, models, and research that use this idea.
1OpenAI's tiktoken library exposes get_encoding('o200k_base') for GPT-4o, GPT-5, and GPT-5.5 token counting; the per-language fertility numbers come from running it on representative samples.
2Anthropic's Claude messages.count_tokens and OpenAI's response.usage both expose actual input and output token counts post-request, which production teams reconcile against estimates nightly.
3OpenAI's prompt-caching pricing in 2026 charges discounted rates on cached prefixes, so long static system prompts only pay full price on the first occurrence, which production cost models need to account for.
4Customer-support agent platforms like Intercom Fin and Zendesk AI agents publish per-conversation cost ranges that explicitly call out tool count and multi-turn flow as the main cost drivers.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does OpenAI's prompt caching change the cost picture for a long static system prompt?
Sign in or go Pro to see the approach for this follow-up.
QYour traffic shifts to 40 percent Hindi users overnight. What happens to the bill and how do you respond?
Sign in or go Pro to see the approach for this follow-up.
QTools versus prompt: where do you put a complex business rule, in the system prompt or as a tool the agent can call?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Estimating from user-visible prompt length alone, missing the tool schema, the chat-template wrappers, and the conversation history that all count toward billed input tokens.