Short answer: set reasoning_effort="low" for anything mechanical, cap max_output_tokens, enable prompt caching for repeated system prompts, and send only relevant chunks instead of whole documents.

Why is GPT-5 so expensive to run?

GPT-5 is priced at $1.25 per million input tokens and $10 per million output tokens. That 8:1 ratio means a prompt that reads 100K tokens and writes 5K tokens costs $0.125 in input and $0.05 in output — but the reasoning tokens you never see are billed as output too.

At high reasoning effort, a single hard question can generate 20K to 40K reasoning tokens. That is $0.20 to $0.40 for one answer. Multiply by a few thousand calls a day and it adds up fast.

How do you cut GPT-5 output costs?

1. Cap the output. Set max_output_tokens=800 for chat replies. Unbounded generation is where bills explode.
2. Match effort to task. Use low for extraction and routing, medium for drafting, high only when you have measured that it fixes real errors.
3. Turn off visible reasoning. If you are logging chain-of-thought you do not read, stop paying for it.
4. Ask for structured output. JSON mode returns fewer tokens than prose and parses reliably.

resp = client.chat.completions.create(
    model="gpt-5",
    reasoning_effort="low",
    max_output_tokens=800,
    response_format={"type": "json_object"},
    messages=[...]
)

Does prompt caching actually help GPT-5?

Yes, and it is the single biggest win for agents. If your system prompt is 8K tokens and you run 10,000 calls a day, you are paying $100 a day just on instructions. Prompt caching cuts the cached input portion by up to 90 percent on repeated prefixes.

The pattern is to put stable content first — system prompt, tool definitions, style guide — and variable content last. Caches match on prefixes, so order matters more than people expect.

When should you route away from GPT-5?

A surprising share of calls do not need a frontier model. Sentiment tagging, language detection, and simple reformatting can run on a small model at one-hundredth the price with equal accuracy on those narrow tasks.

Build a router that scores input complexity first. Even a crude keyword-based router catches the easy 40 percent.

What is the one metric to track?

Track cost per resolved task, not cost per token. A model that answers in one call at $0.30 beats a cheaper model that needs three retries at $0.08 each. Teams that only optimize token price often end up spending more because retry loops multiply output tokens.

This cost-cutting playbook suits teams running more than a few thousand calls a day. If you make 20 requests a week, the effort is not worth it — just use the ChatGPT UI.

Quick Summary

• GPT-5 output tokens cost 8x input tokens, and reasoning tokens are billed as output.
• Cap max_output_tokens and use low reasoning effort for mechanical tasks.
• Prompt caching can cut cached input costs by up to 90 percent on repeated prefixes.
• Route simple classification and extraction to a small model for roughly 95 percent savings.
• Measure cost per resolved task, because retries hide the real bill.

Short answer: use GPT-5 in the ChatGPT app or via the API, set reasoning effort to medium for daily work, and keep individual prompts under roughly 50K tokens unless you actually need the full window.

What is GPT-5 and what changed?

GPT-5 is a multimodal large language model from OpenAI that accepts text, images, and audio input and returns text. It ships with a 400,000-token context window and roughly 128K tokens of maximum output, which is a real jump over GPT-4o’s 128K context.

The feature that matters most day to day is reasoning effort. You choose between low, medium, and high. Low answers fast and cheap. High spends 30 seconds to two minutes thinking before it writes anything.

Pricing on the API sits at $1.25 per million input tokens and $10 per million output tokens for the base tier. ChatGPT Plus is $20 a month and includes GPT-5 usage caps that reset every three hours.

How do you get started with GPT-5?

1. Sign in at chatgpt.com or openAI’s API dashboard, or download the desktop app for macOS and Windows.
2. In the model picker, choose GPT-5. The mobile app hides it behind the model dropdown at the top of the chat.
3. For API work, create a key under API keys and copy it once — OpenAI never shows it again.
4. Install the SDK: pip install openai and set export OPENAI_API_KEY=sk-....

Your first call looks like this:

from openai import OpenAI
client = OpenAI()
resp = client.chat.completions.create(
    model="gpt-5",
    reasoning_effort="medium",
    messages=[{"role": "user", "content": "Summarize this contract"}]
)

Which reasoning tier should you actually pick?

Medium is the right default for 80% of tasks. Low is for classification, extraction, and formatting where the answer is obvious. High is worth the wait only for multi-step math, debugging tricky code, or long-form planning.

High effort can triple your output token bill. If you are prototyping, start at low and only move up when the output is genuinely wrong, not when it is merely terse.

Is the 400K context window worth using?

Sometimes — but it is a trap more often than people admit. Cost scales linearly with input tokens, so pasting a 300K-token codebase costs around $0.38 per call at input pricing, and quality drops in the middle of very long contexts.

Better approach: use retrieval. Split documents into chunks, embed them, and send only the five to ten most relevant chunks. You get faster responses and a smaller bill for the same result.

What does GPT-5 still get wrong?

It still fabricates citations if you ask for sources without web search enabled. It also over-explains when reasoning effort is high — you get the full chain of thought as visible output, which is verbose for a chat UI.

GPT-5 is a poor fit if you need guaranteed deterministic output. Any workflow requiring bit-for-bit reproducible answers should not route through an LLM at all.

Who should use it: developers writing code, analysts summarizing documents, and writers who need a first draft. Who should skip it: anyone whose task is a single lookup — a search engine is faster and free.

Quick Summary

• GPT-5 has a 400K-token context window and three reasoning tiers: low, medium, high.
• Use medium effort as your default; reserve high for hard math and debugging.
• API pricing is $1.25 per million input tokens and $10 per million output tokens.
• Retrieval beats pasting huge contexts — it is cheaper and more accurate.
• Verify any factual claim GPT-5 makes without web search, because it can still hallucinate sources.