Guides / openai api cost
OpenAI API cost across the whole model line
OpenAI API cost runs from $0.20 per million input tokens on nano to $30.00 on the Pro tier, a 150-fold spread, while every model bills output at roughly six times input. Culpa, a local-first LLM cost, margin, and forecast ledger, records which model served each call, so a drift up that ladder shows as a number rather than a surprise.
Why this happens
OpenAI's line has the widest price range of any provider in this book. Nano to Pro is 150x on input and 144x on output, so which model handles a call matters far more than how efficient that call is. A team that optimises prompts while leaving routing alone is working on the smaller lever. The output-to-input ratio stays near six across the whole line, which at least makes the shape predictable once you pick a tier.
What this usually looks like
- Spend is concentrated on a model handling a minority of your calls.
- Nobody has re-tested whether nano or mini could do the routing and classification work.
- A Pro-tier call appears in your top spenders despite a tiny call count.
- Your model mix has drifted upward across quarters with no decision recorded anywhere.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Mistakes that cost the most
| Mistake | Why it hurts | Do instead |
|---|---|---|
| Optimising prompts before optimising routing. | A 150x price ladder means the model choice dominates, so prompt work is the smaller lever. | Rank spend by model first. Move whatever work a cheaper tier handles, then tune what remains. |
| Treating the Pro tier as just another option. | Pro bills twelve times GPT-5.4 and carries no cached rate, so repeated prompts pay full price. | Make Pro an explicit, logged decision per surface, and never let a retry path escalate into it. |
| Assuming a cheaper model changes the price shape. | Every OpenAI model bills output near six times input, so a cheap model with verbose answers still skews. | Cap output length at every tier. The ratio holds even where the absolute rate is tiny. |
Run this check tonight
- List spend by model and compute each one's share. The top line is usually not your top call count.
- Price your current volume at one tier down. That difference is what a routing review is worth.
- Check whether any Pro-tier calls exist, and whether anyone decided they should.
- Confirm your cached-token share. Every OpenAI model except Pro carries a 90% cached discount.
- Compare output to input tokens. On a six-to-one ratio, output is where the bill lives.
The same million tokens, priced at four tiers
Illustrative example
1 million input and 200,000 output tokens, priced across the OpenAI line to show what routing alone is worth.
$0.45 to $66.00 for identical work. Nothing about the prompt changed, only which model answered it. That's the largest single lever on this provider.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| 150x | spread between the cheapest and dearest OpenAI input rates | calculated | $30.00 per million on gpt-5.4-pro divided by $0.20 on gpt-5.4-nano, from the price book, effective 2026-07-02. |
| $0.45 to $66.00 | cost of one million input and 200,000 output tokens across the OpenAI line | calculated | Five tier calculations shown in the teardown, all from price-book rows effective 2026-07-02. |
What a generic answer can’t know
A pricing page shows five rates. It can't tell you which of your features is sitting a tier higher than it needs to, or that a retry path quietly escalates into Pro. That needs per-call model attribution joined to your own surfaces. Culpa records it on your infrastructure, keeps your prompts and responses there, and counts the calls to run your plan.
Questions founders ask next
What is the cheapest OpenAI model?
GPT-5.4 nano at $0.20 per million input and $1.25 output, verified 2026-07-02. It sits 150 times below the Pro tier on input, which is why routing decisions dominate cost on this provider.
Do all OpenAI models have the same price shape?
Close to it. Every model bills output at roughly six times input, so the shape is predictable across tiers. What changes enormously is the absolute rate, and that's where the decisions are.
Which OpenAI models support cached input?
All of them except the Pro tier, at a 90% discount against the standard input rate. The Pro models record no cached rate in this price book, so a repeated system prompt there pays full price on every call.
On your infrastructure
Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.
Free, no card, no account
Run the free Cost Leak Scan
It shows your most expensive conversation before you install anything.
Keep reading
Sources: OpenAI API pricing. Last reviewed 2026-08-01, rates effective 2026-07-02. Plain text version.