Guides / gpt-5.4 pricing
GPT-5.4 pricing, and what it really costs per month
GPT-5.4 pricing is $2.50 per million input tokens and $15.00 per million output tokens, on rates verified 2026-07-02. Output bills six times input, so answer length drives your bill harder than prompt length. Culpa, a local-first LLM cost, margin, and forecast ledger, prices every call from that rate and names the conversation that spent it.
Why this happens
GPT-5.4 is the production default for most teams, which makes it the least re-examined line on the bill. Its price shape matters more than its headline rate. The six-to-one output spread means a verbose response format costs more than a long system prompt, and the cached rate at $0.25 per million is a 90% discount most teams never claim because the repeated block isn't at the front of the prompt.
What this usually looks like
- Your bill climbs while call count holds, which points at response length rather than traffic.
- Cached tokens sit near zero even though an identical system prompt goes out on every call.
- Nobody has re-tested whether GPT-5.4 mini would handle the routing and classification work at a third of the rate.
- One feature with open-ended responses outspends several features with far higher call volume.
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 |
|---|---|---|
| Comparing GPT-5.4 to another model on the input rate alone. | Output bills six times input here, so the headline input price hides most of your real cost. | Compare on your own token mix. Multiply both rates by the input and output volume you actually send. |
| Leaving the response format open-ended. | On a six-to-one spread an extra hundred output tokens costs what six hundred input tokens would. | Cap max tokens and ask for the shortest form that works, then measure whether quality actually moved. |
| Treating GPT-5.4 as the default for every step. | Classification and routing rarely need it, and mini runs at $0.75 input against $2.50 here. | Route the mechanical steps to mini and keep GPT-5.4 for the work that needs the quality. |
Run this check tonight
- Pull input and output token totals separately for the last 30 days. Combined totals hide the six-to-one spread.
- Multiply each by the rates above. A large gap against your invoice means calls you aren't counting.
- Check what share of input billed at the cached $0.25 rate. Near zero is a 90% discount left untouched.
- Price your single longest completion this month. On a six-to-one spread it's rarely a rounding error.
- List which features use GPT-5.4 for work a mini model could do, and price the difference.
A month on GPT-5.4, with the math shown
Illustrative example
One feature sending 1 million input tokens and 200,000 output tokens in a month. Round numbers, so you can scale them to your own volume.
Output is 55% of the bill on a fifth of the tokens. Caching the repeated input takes 25% off the total, and neither number is visible from a rate card alone.
Every number, with its confidence and source
| Figure | What it means | Confidence | Source |
|---|---|---|---|
| $2.50 / $15.00 per million | GPT-5.4 input and output rates | calculated | Culpa price book row for openai/gpt-5.4, $0.0025 and $0.015 per 1k, effective 2026-07-02, verified against OpenAI API pricing. Shown here per million. |
| 6x | output rate as a multiple of the input rate on GPT-5.4 | calculated | $15.00 divided by $2.50, from the same price-book row. |
| 90% | discount the cached input rate carries against standard input | calculated | 1 - ($0.25 / $2.50), using the cached and input rates from the same row. |
What a generic answer can’t know
A rate card tells you $15.00 per million output tokens. It can't tell you that one conversation burned a fifth of your GPT-5.4 spend, or that the feature driving it serves customers on your cheapest plan. That needs request-level data, which means prompts and responses. Culpa keeps those on your infrastructure and counts the calls to run your plan.
Questions founders ask next
How much does GPT-5.4 cost per million tokens?
$2.50 for input and $15.00 for output, on rates verified 2026-07-02 against OpenAI's pricing page. Cached input runs at $0.25 per million, a 90% discount that only applies if the repeated block sits at the front of your prompt.
Why is my GPT-5.4 bill higher than this calculation?
A published rate is a rate, not an invoice. Retries, tool definitions and conversation history resent on every turn all add tokens the arithmetic above doesn't include. The gap between the two is your effective cost.
Is GPT-5.4 mini worth switching to?
For classification, routing and extraction, usually. Mini runs at $0.75 input and $4.50 output against $2.50 and $15.00 here, so the saving is roughly threefold on the same token volume. Test quality on your own data before moving reasoning work.
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.