Find where the money goes
Open the usage page in your OpenAI dashboard and sort by model. Most bills come from one or two models and one or two features. Output tokens cost several times as much as input on every GPT model, so long answers add up faster than long prompts.
The example on this page is 100,000 replies a month, about 2,500 tokens in and 400 out each. On GPT-5.5 at list price that is $2,450 a month.
Use the smallest model that does the job
The same month of support replies, priced on each GPT model. Try a sample of real requests on the smaller ones before you switch; the price only matters if the answers are good enough.
| Model | Input | Output | Per month |
|---|---|---|---|
| GPT-5.5 | $1,250 | $1,200 | $2,450 |
| GPT-6.1 Sol | $500 | $400 | $900 |
| GPT-6.1 Terra | $250 | $200 | $450 |
| GPT-5.4 mini | $188 | $180 | $368 |
| GPT-5 mini | $62.50 | $80.00 | $143 |
| GPT-5 nano | $12.50 | $16.00 | $28.50 |
You do not have to pick one model for everything. Short, simple requests can go to a small model and hard ones to a large model. That is what an AI API router does.
Keep cached input working
OpenAI caches repeated prompt starts on its own, and cached input costs a tenth of the input price on GPT-5.5. The cache only hits when the start of the prompt is the same, so put your fixed instructions first and anything that changes, like the date or the user's name, after them.
| Model | List price | 50 in 100 input tokens cached | Batch price |
|---|---|---|---|
| GPT-5.5 | $2,450 | $1,888 | $1,225 |
| GPT-6.1 Sol | $900 | $663 | $450 |
| GPT-6.1 Terra | $450 | $331 | $225 |
| GPT-5.4 mini | $368 | $283 | $184 |
Batch what can wait
The Batch API runs jobs within 24 hours at half the standard price. Nightly reports, tagging old records, and evals rarely need an answer in seconds.
Watch long prompts and regions
Several GPT models charge more once a prompt passes 272K tokens, and regional processing adds 10 in 100 to the price for models released from March 5, 2026. Trim old chat turns and large files you do not need, and only pick a region when you must.
| Model | Input | Cached input | Output | Notes |
|---|---|---|---|---|
| GPT-5.5 | $5 | $0.50 | $30 | Over 272K prompt tokens: $10 in, $45 out. |
| GPT-6.1 Sol | $2 | $0.10 | $10 | Over 272K prompt tokens: $4 in, $15 out. |
| GPT-6.1 Terra | $1 | $0.05 | $5 | Over 272K prompt tokens: $2 in, $7.50 out. |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | Over 272K prompt tokens: $0.20 in, $0.75 out. |
| GPT-5.4 mini | $0.75 | $0.075 | $4.50 | None |
| GPT-5 mini | $0.25 | $0.025 | $2 | None |
| GPT-5 nano | $0.05 | $0.005 | $0.40 | None |
| GPT-4o mini | $0.15 | $0.075 | $0.60 | Marked deprecated, still served. |
Cap each key
Set a monthly budget in your OpenAI project settings, and use a separate key per app or person so one runaway script cannot spend for everyone. More ideas that work across providers are in how to reduce your LLM API costs.