- Blog
- GPT Image 2.6 Pricing: Official API Rates and Image Budget Examples
GPT Image 2.6 Pricing: Official API Rates and Image Budget Examples
on 10 days ago

GPT Image 2.6 pricing starts with token rates, so a single price per picture cannot describe every request. OpenAI lists $5 per million text input tokens, $8 per million image input tokens, and $30 per million image output tokens for both Sunburst and Flare. Your total depends on the tokens billed for the work you request.
This guide turns those rates into practical budgets for creators and small teams. Rates were checked on November 10, 2026. The calculations below are hypothetical examples, not measured benchmarks or guaranteed prices for a particular resolution. They also describe direct OpenAI API charges, not Shotyard credit prices.
GPT Image 2.6 pricing: the official rate table
The official Sunburst model page and Flare model page publish the same rates:
| Billing component | Standard price per million tokens | Listed cached input price |
|---|---|---|
| Text input | $5.00 | $1.25 |
| Image input | $8.00 | $2.00 |
| Image output | $30.00 | Not applicable |
Text input covers the prompt; image input covers source or reference images; image output covers generated image tokens. The prices are quoted per million tokens; the examples below scale them down to individual requests. Any account funding requirements are a separate question.
Treat the cached figures as published rate categories, not automatic savings. The model pages list them, but those entries alone do not establish when your particular request receives cached billing. Use standard input rates for planning until the applicable documentation and your billing records confirm otherwise. Reusing the same picture is not enough evidence to budget at the discounted rate.
Calculate the cost of one image request

For an ordinary raster-image request using uncached inputs, calculate each component separately:
Request cost in USD = text input tokens × 0.000005 + image input tokens × 0.000008 + image output tokens × 0.000030.
Suppose a request uses 250 text input tokens, no reference image, and 1,200 image output tokens. The prompt costs $0.00125 and the output costs $0.036, producing a total of $0.03725. This example chooses token counts to explain the arithmetic; it does not claim that a specific image size produces 1,200 tokens.
Now imagine an edit using the same prompt and output counts, plus 1,500 image input tokens. Those image inputs add $0.012, bringing the request to $0.04925. This is why a text-only generation and an image edit can have different totals even when their finished pictures look similar.
| Hypothetical request | Text input tokens | Image input tokens | Image output tokens | Calculated cost |
|---|---|---|---|---|
| New image | 250 | 0 | 1,200 | $0.03725 |
| Edit with reference input | 250 | 1,500 | 1,200 | $0.04925 |
| Larger output-token budget | 250 | 0 | 2,400 | $0.07325 |
Keep the decimals until you calculate the total. Rounding a $0.03725 request to $0.04 adds $2.75 to a forecast of 1,000 identical requests.
Why size, quality, and format change the estimate
The official image generation cost guide lets you estimate output costs by model, quality, and size. It lists five quality settings for GPT Image 2.6: low, medium, high, xhigh, and max. The guide also warns that models can use different token counts at the same quality setting.
For a repeatable estimate, choose explicit settings. An automatic setting leaves a variable in the calculation. Record the model, dimensions, quality, and number of requested outputs beside the cost so another person can reproduce the comparison.
Do not apply the raster formula blindly to every format. OpenAI's guide says SVG generation includes the image model's reference-image cost and the authoring model's token cost. It also says markup-guided edits have no extra markup fee and use normal edit token rates. A feature without a separate surcharge can still involve billable inputs and outputs.
Test the format you intend to deliver. Budget vector work and animations separately from ordinary still images.
Sunburst versus Flare: compare the finished job
OpenAI describes Sunburst as a model for detailed visual work and Flare as a model for fast iteration. Both use the same published token prices. That does not establish equal request totals, equal generation times, or equal costs for a finished campaign.
A useful comparison starts with one brief: the same subject, composition, reference material, and delivery dimensions. Decide what counts as an acceptable result before generating anything. For a product image, that might mean the correct silhouette, readable packaging, and space for a headline.
Then record the billed spend and the number of acceptable images from each run. Cost per accepted image equals total generation spend divided by accepted images. A model that needs fewer revisions may be economical for that brief even when its individual requests cost more. This is a measurement to make in your workflow, not a claim that either variant always wins.
Build a monthly budget that includes revisions

Start with the number of assets you must deliver, then estimate attempts per accepted asset. If a hypothetical month requires 200 approved images and three generated attempts per approved image, you are budgeting for 600 generation attempts.
Using the first example above, 600 requests at $0.03725 would cost $22.35. If all 600 instead matched the reference-edit example, the total would be $29.55. Neither figure is an advertised package or a prediction of your account's usage; both depend entirely on the stated assumptions.
Keep extra editing rounds visible. If each approved asset needs another paid edit, that is another 200 requests in this scenario. Storage, delivery, application subscriptions, taxes, and staff time also sit outside the token arithmetic shown here.
After a small pilot, replace the assumed attempts and token counts with your own records. Track simple backgrounds separately from demanding product compositions so a cheap category does not hide an expensive one.
Compare direct API costs with Shotyard credits
A provider's token table and an application's credit system answer different questions. The API table explains underlying metered usage. An application may package access and workflow features into credits or a subscription, so dividing its headline price by an API rate does not establish what a customer will pay.
Visit the Shotyard GPT Image 2.6 page to review the available workflow, then check Shotyard pricing for current plan details. Confirm the displayed generation cost for your chosen settings before starting a large job.
When comparing routes, use the same deliverable and count revisions. For an occasional campaign, convenience may matter more than a small difference in estimated inference spend. For recurring automated production, compare a representative month of usage records before choosing.
FAQ
How much does GPT Image 2.6 cost per image?
There is no single total that fits every request. Multiply each billed token category by its rate and add the results. The $0.03725 example here uses assumed counts; it is not a standard price for every generated image.
Is Flare cheaper than Sunburst?
Their published token rates are identical. Compare actual token consumption and the attempts needed to obtain acceptable results before deciding which costs less for your work.
Can I assume a Batch API discount?
No. The official model pages currently mark Batch as unsupported for both variants. Plan against the supported image generation and image editing endpoints, and recheck the model pages before relying on a future discount.
Does repeating a reference image guarantee cached pricing?
No guarantee follows from the rate table alone. Confirm eligibility for the exact endpoint and inspect billing evidence before reducing your budget. This guide's worked examples use standard uncached rates throughout.
Conclusion
GPT Image 2.6 pricing becomes easier to plan when you separate the official token rates from assumptions about each request. Calculate inputs and outputs, allow for revisions, and measure spend per accepted asset. Use a small representative pilot to replace the example numbers before committing your full production budget.
