Pricing
GPT Image 2.6 is billed by token usage. Image dimensions, aspect ratio, quality, reference images, and reported text-output tokens determine the total cost. The default quality is high.
Estimated output cost per image
The table below shows the estimated image-output cost for one still image. It does not include reference images, prompt text, or reported text-output tokens.
Values are shown before request-level rounding. Actual charges depend on reported token usage.
| Output size | low | medium | high | xhigh | max |
|---|
| 1024×768 | $0.01689 | $0.02979 | $0.03612 | $0.06420 | $0.14445 |
| 1024×1024 — square | $0.02469 | $0.04347 | $0.05268 | $0.09366 | $0.21072 |
| 1536×1536 — square | $0.03516 | $0.06216 | $0.07533 | $0.13392 | $0.30132 |
| 2048×2048 — square | $0.05001 | $0.08832 | $0.10704 | $0.19029 | $0.42816 |
| 2880×2880 — square | $0.08304 | $0.14682 | $0.17790 | $0.31626 | $0.71157 |
| 1024×1536 | $0.01992 | $0.03396 | $0.04116 | $0.07377 | $0.16464 |
| 1920×1072 | $0.01851 | $0.03138 | $0.03945 | $0.07014 | $0.15780 |
| 2560×1440 | $0.02583 | $0.04731 | $0.05529 | $0.09828 | $0.22110 |
| 3840×2160 | $0.04674 | $0.08565 | $0.10008 | $0.17790 | $0.40026 |
Use exact dimensions rather than labels such as “2K.” A 2048×2048 square and a 2560×1440 landscape have different output costs.
Pricing depends on both pixel area and the model’s token grid. Square images use a larger grid at the same quality, so pixel count alone does not predict cost. For example, 1024×1024 costs more than 1024×1536 despite containing fewer pixels.
Token rates and total request cost
| Component | Price per 1 million tokens |
|---|
| Text input | $5.00 |
| Reference-image input | $8.00 |
| Reported text output, including reasoning when reported as text | $10.00 |
| Image output | $30.00 |
Total cost = text input + reference-image input + text output + image output.
The total is rounded up to the next $0.0001 once per request.
Each reference contributes image-input tokens. Its cost depends on token usage rather than a fixed fee per uploaded image. Up to 16 reference images are supported. Generating multiple images also increases the image-output component.
Example: a 2048×2048 medium-quality edit
This hypothetical example assumes one output, 1,000 reference-image input tokens, 100 text-input tokens, and no text-output tokens:
| Component | Calculation | Cost |
|---|
| Image output | 2,944 tokens × $30 / 1M | $0.08832 |
| Reference-image input | 1,000 tokens × $8 / 1M | $0.00800 |
| Text input | 100 tokens × $5 / 1M | $0.00050 |
| Subtotal | | $0.09682 |
| Total after rounding | | $0.0969 |
These input-token counts are illustrative, not a fixed allowance or a guaranteed cost for a particular reference resolution. If the request also reports 1,000 text-output tokens, the rounded total becomes $0.1069.
How pricing differs from GPT Image 2.5
The image-output rate remains $30 per million tokens, but the estimated number of output tokens differs:
| Quality | GPT Image 2.6 compared with GPT Image 2.5 at the same dimensions |
|---|
| low | Approximately 4.2× the image-output tokens |
| medium | Approximately 3.3× the image-output tokens |
| high, xhigh, max | Same image-output token estimate |
These ratios apply to image output, not the entire request. Reference images, prompt text, reported text output, and rounding affect the final price.
For example, a 2048×2048 medium-quality image has an estimated output cost of $0.02676 on GPT Image 2.5 and $0.08832 on GPT Image 2.6, before other components.
Settings that affect estimates
- Quality: Defaults to
high. Set quality explicitly for more predictable estimates. auto lets the model choose and is not a fixed-price tier.
- Reasoning effort: Controls planning independently of rendering quality and defaults to
high. Reported text-output tokens are billed at the rate above.
- Resolution: Still-image dimensions are aligned to multiples of 16, with a maximum edge of 3,840 pixels, a maximum area of 8,294,400 pixels, and an aspect ratio of at most 3:1. A 2880×2880 square fits these limits; 4096×4096 does not.
- GIF output: Frame count affects usage. The still-image table does not represent the total cost of an animation.
openai/gpt-image-2.6/flare/edit
