Average tokens per request and monthly volume. PTUs are sized for the peak.
1,000,000 requests a month · 1,500 in / 300 out tokens
| Option | Per month | Break-even |
|---|---|---|
| Standard (Global) Pay per token. No capacity commitment; subject to rate limits (TPM quota). | $5,400 | — |
| Batch 50% cheaper, but results can take up to 24 hours. Only for work that can wait. | $2,700 | — |
| Provisioned, hourly 65 PTU × 730 hours. Billed while the deployment exists, used or not. | $47,450 | never |
| Provisioned, 1-month reservation 65 PTU reserved for a month. Reservations discount the hourly meter; they don't reserve capacity. | $16,900 | 3,129,630 req |
| Provisioned, 1-year reservation 65 PTU reserved for a year, shown per month. | $14,365 | 2,660,186 req |
Break-even: monthly requests at which the PTU option costs the same as Standard, for the PTU count below. Standard and Batch prices are per token; Batch assumes no cached input.
Microsoft's formula for GPT-4.1: 3,000 input tokens per minute per PTU, 1 output token = 4 input tokens.
Global deployments start at 15 PTUs and grow in steps of 5.
A 1-month PTU reservation pays off from about 3,129,630 requests a month at this size (3.1× your volume).
| Tokens | Price / 1M | Per month |
|---|---|---|
| Input | $2.00 | $3,000 |
| Cached input | $0.500 | $0 |
| Output | $8.00 | $2,400 |
Prices: Azure Retail Prices API, refreshed 19 Oct 2026, USD. PTU: Azure Retail Prices API, refreshed 19 Oct 2026. Prompts over 128K tokens aren't supported on provisioned deployments.
Global deployment, cheapest Standard first. PTUs sized for each model's throughput.
| Model | Standard / month | PTUs | PTU 1-month | PTU 1-year / month | |
|---|---|---|---|---|---|
| GPT-5 nano | $195.00 | — | — | — | Use |
| GPT-4.1 nano | $270.00 | 15 | $3,900 | $3,315 | Use |
| GPT-4o mini | $405.00 | 15 | $3,900 | $3,315 | Use |
| GPT-5.6 luna | $660.00 | 15 | $3,900 | $3,315 | Use |
| GPT-5 mini | $975.00 | 15 | $3,900 | $3,315 | Use |
| GPT-4.1 mini | $1,080 | 15 | $3,900 | $3,315 | Use |
| GPT-5.4 mini | $2,475 | 30 | $7,800 | $6,630 | Use |
| o4-mini | $2,970 | 35 | $9,100 | $7,735 | Use |
| o3-mini | $2,970 | 75 | $19,500 | $16,575 | Use |
| GPT-5.1 | $4,875 | 60 | $15,600 | $13,260 | Use |
| GPT-5.1 codex | $4,875 | 60 | $15,600 | $13,260 | Use |
| GPT-5 | $4,875 | 60 | $15,600 | $13,260 | Use |
| GPT-4.1 | $5,400 | 65 | $16,900 | $14,365 | Use |
| o3 | $5,400 | 65 | $16,900 | $14,365 | Use |
| GPT-5.6 terra | $6,600 | 80 | $20,800 | $17,680 | Use |
| GPT-4o | $6,750 | 75 | $19,500 | $16,575 | Use |
| GPT-5.3 codex | $6,825 | 80 | $20,800 | $17,680 | Use |
| GPT-5.2 | $6,825 | 80 | $20,800 | $17,680 | Use |
| GPT-5.2 codex | $6,825 | 80 | $20,800 | $17,680 | Use |
| GPT-5.4 | $8,250 | 95 | $24,700 | $20,995 | Use |
| GPT-5.6 sol | $12,000 | 190 | $49,400 | $41,990 | Use |
| GPT-5.5 | $16,500 | 190 | $49,400 | $41,990 | Use |
| o1 | $40,500 | 805 | $209,300 | $177,905 | Use |
"pending": no price for this deployment type in our data yet. GPT-6 and image models size PTUs with weighted tokens and aren't included.
Microsoft sizes provisioned throughput from the peak minute: input tokens per minute (minus cached tokens) plus output tokens per minute multiplied by the model's output-to-input ratio, divided by the model's input TPM per PTU. Round up to the scale increment (5 PTUs for Global and Data Zone on most models) and to at least the minimum (15). This calculator does exactly that for the model you pick.
When the deployment is busy most of the time. PTUs are billed for every hour they exist, so they only pay off when the workload keeps them well used. Sized for a 3× peak, a steady workload averages about a third of its PTU capacity, which is why Standard usually wins below several thousand dollars a month.
Global routes requests to any Azure region and is the cheapest. Data Zone keeps processing inside the US or EU data zone at about 10% more. Regional keeps it in one region; Regional PTU deployments also need a larger minimum (25–50 PTUs).
Cached input tokens are billed at the cached price on Standard and don't use PTU capacity at all. Reasoning tokens from GPT-5 and o-series models are billed as output tokens and use capacity like output tokens.
No. A reservation is a discount on the hourly PTU meter for a month or a year. Create the deployment first to confirm the region has capacity, then buy the reservation for the same deployment type.
Token and PTU prices are read every week from the Azure Retail Prices API (USD list prices; Data Zone and Regional for eastus2). Models the API hasn't priced yet show list prices checked on 2026-10-15. PTU sizing parameters are from Microsoft Learn. GPT-6 models use a weighted token formula for PTUs and aren't covered.