Cost & Tokens
You pay per token, split into input (prompt) and output (completion).
LLM pricing is based on tokens, and almost always split into two rates:
- Input tokens — everything you send (system, history, documents).
- Output tokens — everything the model generates. Usually priced higher.
Estimating a call
Cost = (input tokens x input rate) + (output tokens x output rate). The API response reports exact usage so you can track spend.
How to spend less
- Trim history and documents to what is relevant.
- Set a tight
max_tokens. - Use a smaller model for easy tasks.
- Cache repeated prompts (covered later).
Example
{
"usage": {
"prompt_tokens": 850,
"completion_tokens": 120,
"total_tokens": 970
}
}When to use it
- A startup tracks monthly token usage per customer and charges overages above the free tier based on input + output token counts.
- A developer reduces prompt size from 2000 to 400 tokens by moving static context to a cached system prompt, cutting costs by 80%.
- A product team chooses gpt-4o-mini over gpt-4o for a classification task because accuracy is equivalent and the smaller model costs 15x less.
More examples
Calculate request cost in Python
Computes the dollar cost of a request by multiplying token counts by per-token prices.
from openai import OpenAI
# gpt-4o-mini pricing (USD per 1M tokens, as of mid-2025)
INPUT_PRICE = 0.15 / 1_000_000
OUTPUT_PRICE = 0.60 / 1_000_000
client = OpenAI()
r = client.chat.completions.create(
model='gpt-4o-mini',
messages=[{'role': 'user', 'content': 'Summarise the water cycle.'}]
)
u = r.usage
cost = u.prompt_tokens * INPUT_PRICE + u.completion_tokens * OUTPUT_PRICE
print(f'Cost: ${cost:.6f} (in={u.prompt_tokens}, out={u.completion_tokens})')Track cumulative token spend
Accumulates token counts across multiple requests to measure the cost of a whole batch.
from openai import OpenAI
client = OpenAI()
total_in = total_out = 0
questions = [
'What is photosynthesis?',
'What is osmosis?',
'What is mitosis?'
]
for q in questions:
r = client.chat.completions.create(
model='gpt-4o-mini',
messages=[{'role': 'user', 'content': q}],
max_tokens=60
)
total_in += r.usage.prompt_tokens
total_out += r.usage.completion_tokens
print(f'Batch totals — input: {total_in}, output: {total_out}')Reduce cost by shrinking the prompt
Truncates the context to a token budget before sending, reducing prompt cost while keeping key content.
import tiktoken
from openai import OpenAI
enc = tiktoken.encoding_for_model('gpt-4o-mini')
def trim_to_tokens(text, max_tokens=200):
tokens = enc.encode(text)
return enc.decode(tokens[:max_tokens])
long_context = open('/tmp/article.txt').read()
short_context = trim_to_tokens(long_context, max_tokens=200)
client = OpenAI()
r = client.chat.completions.create(
model='gpt-4o-mini',
messages=[
{'role': 'system', 'content': 'Answer based on the context below.'},
{'role': 'user', 'content': short_context + '\n\nWhat is the main topic?'}
]
)
print(r.choices[0].message.content)
Discussion