Controle de custos: contagem de tokens e orçamentos

# pip install tiktoken
import tiktoken

enc = tiktoken.encoding_for_model('gpt-4o-mini')
tokens = enc.encode('Hello, world!')
print(len(tokens))   # 4
print(tokens)        # [9906, 11, 1917, 0]

def count_message_tokens(messages, model='gpt-4o-mini'):
    enc = tiktoken.encoding_for_model(model)
    total = 0
    for m in messages:
        total += 4  # role + structural
        total += len(enc.encode(m['content']))
    return total + 2  # priming

count = client.messages.count_tokens(
    model='claude-sonnet-4-5',
    system='You are helpful.',
    messages=[{'role': 'user', 'content': 'Hello'}],
)
print(count.input_tokens)

INPUT_PRICE = 0.150 / 1_000_000   # gpt-4o-mini: $0.15 / 1M input tokens
OUTPUT_PRICE = 0.600 / 1_000_000

input_tokens = count_message_tokens(messages)
est_input_cost = input_tokens * INPUT_PRICE
print(f'Estimated input cost: ${est_input_cost:.6f}')

total_cost = 0

response = client.chat.completions.create(...)
cost = (
    response.usage.prompt_tokens * INPUT_PRICE +
    response.usage.completion_tokens * OUTPUT_PRICE
)
total_cost += cost

MAX_COST_PER_RUN = 0.50  # USD

if total_cost > MAX_COST_PER_RUN:
    raise BudgetExceeded(f'Cost {total_cost} exceeds limit')

key = f'cost:{user_id}:{today_date}'
current = float(redis.get(key) or 0)
if current + cost > 5.0:
    raise QuotaExceeded()
redis.incrbyfloat(key, cost)
redis.expireat(key, midnight_tomorrow_ts)

response = client.chat.completions.create(
    model='gpt-4o-mini',
    messages=messages,
    max_tokens=2048,  # never more than 2k output
)

MAX_STEPS = 20
for step in range(MAX_STEPS):
    response = run_step(...)
    if response.is_final():
        break
else:
    raise StepLimitExceeded()