How to Cut GPT-5 API Costs Without Losing Quality

Most teams overspend on GPT-5 for one reason: output tokens cost eight times more than input tokens, and reasoning models generate a lot of hidden output. Fixing the obvious waste usually drops the bill by 60 to 80 percent.

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.

Task	Route to	Cost impact
Classification, tagging	Small model	-95%
Drafting, analysis	GPT-5 medium	baseline
Hard math, debugging	GPT-5 high	+200%

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.