OpenAI SDK: Models, parameters & tokens

Which model you call decides quality, speed and cost more than any prompt trick. A small, fast model is often right for extraction and classification, and a flagship reasoning model for multi-step analysis. This part covers how to choose, the parameters that matter (reasoning, max_output_tokens, temperature), and how to count tokens and estimate cost before you ship.

1. The model line-up (September 2026)

2. How to choose: start small, measure, move up

3. The parameters that matter

from fake_openai import fake_client

client = fake_client(script=["Positive"])
r = client.responses.create(
    model="gpt-5.6-luna",
    instructions="Classify the sentiment as Positive, Negative or Neutral. Reply with one word.",
    input="The delivery was late but the support team sorted it out brilliantly.",
    reasoning={"effort": "low"},      # how much the model thinks before answering (reasoning models)
    max_output_tokens=50,             # hard cap on output (includes reasoning tokens on reasoning models)
    store=False,                      # don't keep this response on OpenAI's side (lesson 15)
    metadata={"feature": "review-sentiment"},
)
print(r.output_text)
print(client.fake.requests[0]["body"]["reasoning"], client.fake.requests[0]["body"]["max_output_tokens"])

Output

Positive
{'effort': 'low'} 50

4. Counting tokens and estimating cost

exact counts with tiktoken (pip install tiktoken)

import tiktoken

enc = tiktoken.get_encoding("o200k_base")      # the encoding family used by recent OpenAI models
text = "Delivery to Thiruvananthapuram was delayed by two days."
tokens = enc.encode(text)
print(len(tokens), [enc.decode([t]) for t in tokens][:8])

a cost estimator you can run before launching a feature

# Prices per 1M tokens (USD) from the models page when this was written; keep them in config.
PRICES = {"gpt-5.6-luna": {"input": 0.20, "output": 1.20}}

def estimate_monthly_cost(model, requests_per_day, input_tokens, output_tokens, days=30):
    p = PRICES[model]
    per_request = (input_tokens * p["input"] + output_tokens * p["output"]) / 1_000_000
    return per_request, per_request * requests_per_day * days

def rough_tokens(text):             # ~4 characters per token for English; use tiktoken for accuracy
    return max(1, len(text) // 4)

prompt = "Classify this review as Positive, Negative or Neutral: " + "great product " * 40
per_req, monthly = estimate_monthly_cost("gpt-5.6-luna", requests_per_day=20_000,
                                         input_tokens=rough_tokens(prompt), output_tokens=5)
print(f"~{rough_tokens(prompt)} input tokens → ${per_req:.6f} per request, ${monthly:,.2f} per month")

# the same feature with a 2,000-token output (e.g. a long summary) instead of one word:
_, monthly_long = estimate_monthly_cost("gpt-5.6-luna", 20_000, rough_tokens(prompt), 2_000)
print(f"with 2,000 output tokens: ${monthly_long:,.2f} per month")

Output

~153 input tokens → $0.000037 per request, $21.96 per month
with 2,000 output tokens: $1,458.36 per month

Output length dominates. Asking for a one-word label instead of a paragraph can cut the cost of a feature by orders of magnitude. After launch, track real usage from every response (part 13).

Recap