Access GPT Image 2 — the image model released alongside ChatGPT Images 2.0, also searched as gpt-image-2 or ChatGPT Images API — through EvoLink's unified image API. Test text-to-image, reference-image editing with masks, and the flat-rate GPT Image 2 Beta route before integrating.
Test GPT Image 2 output quality, estimate the token cost of one request, and generate a sample before you formally integrate.
Create image task
Fixed to gpt-image-2Text-to-image, image-to-image and mask-guided editing. Billed by the tokens the upstream usage object reports.
Upload source or reference images for image-to-image and editing. Each one adds image input tokens.
Pixel budget for the chosen ratio: 1K about 1.05 MP, 2K about 4.19 MP, 4K is 8.29 MP.
Rendering effort. It drives the output token count, so it is the single biggest cost lever.
Transparent keeps an alpha channel in the output. This is a Preview feature — results may be unstable.
Results are kept for 24 hours — download anything you want to keep.
Start from the cost of a single sample, pick a testing budget, then price your own case in the calculator. The raw token rates are at the bottom.
First Test Cost
GPT Image 2 - 1K - 1:1 - Low - 1 imageOne Low-quality image, no reference images
About 1888 tests with $10 in credits.
In Playground, choose 1K resolution, a 1:1 aspect ratio, Low quality and 1 image with the prompt “Cat”. This is a token-based estimate; the final cost follows actual usage. Longer prompts and reference images add to the cost.
GPT Image 2 testing budget guide
Pick an amount based on how many tests you expect.Good for a first validation.
Good for prompt iteration.
Good before production integration.
Estimated cost for one generated 1:1 image. Final cost is settled against the token counts the upstream usage object reports.
This page serves two models and they are billed differently. Each block states its own rule, so read the badge before comparing the numbers.
gpt-image-2Billed by token, so there is no fixed per-image price. These are the raw rates; what a request actually costs depends on its size, quality, reference images and count. Use the calculator below to price your own case.
| Item | Rule | Rate | Billing |
|---|---|---|---|
| Image output tokens | The generated image itself. Token count grows with the resolution tier and the quality tier. | $0.027/1K tokens-10% 1.836 cr/1K tokens$0.030 official price | Output tokens |
| Image input tokens | Applied to each reference image in image_urls, and to a mask when one is sent. | $0.0072/1K tokens-10% 0.4896 cr/1K tokens$0.0080 official price | Input tokens |
| Image cached input tokens | Charged only when the upstream usage object reports cached image tokens. | $0.0018/1K tokens-10% 0.1224 cr/1K tokens$0.0020 official price | Input tokens |
| Text input tokens | Produced by the prompt itself. Typically under 1% of a request. | $0.0045/1K tokens-10% 0.306 cr/1K tokens$0.0050 official price | Input tokens |
| Text cached input tokens | Charged only when the upstream usage object reports cached text tokens. | $0.0012/1K tokens-10% 0.0765 cr/1K tokens$0.0013 official price | Input tokens |
Figures are estimates. Final charges are based on actual token usage.
gpt-image-2-betaOne flat rate per call, whatever the aspect ratio - no token accounting and nothing to estimate. Auto or aspect-ratio sizes at the 1K tier only, one image per call. The OpenAI list price has no equivalent for this route, so no official comparison is shown.
| Item | Rule | Rate | Billing |
|---|---|---|---|
| 1K output | Any aspect ratio or auto, at the 1K tier. One image per call, regardless of prompt length or reference images. | $0.015/image 1.02 cr/image | Per call |
Generate high-fidelity images from text prompts or reference images, with strong poster layouts, data-dense infographics, character sheets, product creative, and mask-guided editing workflows.
GPT Image 2 was released alongside ChatGPT Images 2.0 and is available through the API as model gpt-image-2. Use it for text-rich generation and high-fidelity editing; ChatGPT may add orchestration and tools that are separate from the raw image model.
Best fit: Text-rich generation and high-fidelity image editing · Cost basis: Prompt, size, quality, and reference-image input · Output boundary: Custom sizes up to a 3840 px edge; large outputs are experimental
gpt-image-2Token-billed official route with low/medium/high quality, 1K/2K/4K resolution, explicit pixel sizes, n up to 10, and an inpainting mask.
gpt-image-2-betaEvoLink's alternative flat-rate route: every 1K image costs the same regardless of prompt length or references, one image per call.
Use the EvoLink API for product integration and batch jobs, or call it from Codex, Claude, or Gemini for fast creative and development workflows. Both paths share the same EvoLink API key, balance, model routes, and task history.
Best for: product backends, batch jobs, automated workflows
Call EvoLink’s unified image API from your server and control the model ID, parameters, task queue, callbacks, and result storage.
Best for: creative and development tasks in Codex, Claude, and Gemini
Give the Agent your output goal, assets, and acceptance criteria. It can choose the route, assemble the request, track the task, and return the result without requiring you to hand-code every step.
This example shows the shortest runnable flow: create a task, save the returned task ID, then poll or wait for an HTTPS callback. Open the API tab for the complete parameter and response reference.
curl -X POST https://api.evolink.ai/v1/images/generations \
-H "Authorization: Bearer $EVOLINK_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-image-2",
"prompt": "A colorful sunset over the ocean, painterly style",
"size": "auto",
"resolution": "1K",
"quality": "medium",
"n": 1,
"callback_url": "https://your-domain.com/webhooks/image-task-completed"
}'
# Save the returned task ID, then query:
curl https://api.evolink.ai/v1/tasks/{task_id} \
-H "Authorization: Bearer $EVOLINK_API_KEY"Check every value against the parameter cards above: ranges, allowed options, asset counts, and file types must match the selected route.
Check the Authorization bearer token and confirm the available balance and task billing in the console.
Review likeness rights, brand assets, readable text, sensitive content, and input-asset compliance.
A GPT Image 2 task that reaches a final failed state is not charged. Keep the task ID, inspect the final state, and identify whether the cause is validation, review, or execution before retrying.
callback_url must be a public HTTPS URL, not localhost or a private network address. Always keep the task ID as a polling fallback.
| Feature | GPT Image 2 | Nano Banana | Seedream |
|---|---|---|---|
| Starting price | ~$0.048 / image | Lower-cost options | Per-image tiers |
| Billing model | Token based | Per generated image | Per image by output tier |
| Output quality | Low / Medium / High x 1K / 2K / 4K | 1K / 2K / 4K by model | 1K / 1.5K / 2K |
| Reference input | Up to 16 images plus mask | Model dependent | Up to 10 images |
| Best for | Posters, infographics, in-image text | High-quality image generation | Ads, product creative, image editing |
In ratio mode the resolution tier sets the pixel budget (1K is about 1.05 MP, 2K about 4.19 MP, 4K is 8.29 MP). In auto or custom-pixel mode resolution is ignored.
Low, medium, and high change how many output tokens the model spends. Iterate prompts on low, then promote the final render to high.
Credits are charged against the token counts the upstream usage object reports, so a request is quoted before it runs and settled on the real numbers afterwards.
The fields that shape output and cost on POST /v1/images/generations. Generation is asynchronous — poll the returned task ID or use a callback, and save results promptly since image links stay valid for 24 hours. Full schemas and code samples live in the API tab.
| Parameter | Type | What it does |
|---|---|---|
| model | string · required | 'gpt-image-2' for the token-billed official route, or 'gpt-image-2-beta' for the flat-rate 1K route. |
| prompt | string · required | The text instruction. Prompt tokens are metered as text input on the token route. |
| size | string · default auto | 'auto', one of 15 aspect ratios such as 1:1 or 16:9, or explicit WxH pixels in multiples of 16. |
| resolution | string · 1K / 2K / 4K | Pixel budget for ratio mode; ignored when size is auto or explicit pixels. |
| quality | string · low / medium / high | Rendering effort. Drives output-token spend on the token route — iterate on low, promote the final render to high. |
| background | string · opaque / transparent | Alpha channel of the output. Defaults to opaque; transparent is in Preview and results may be unstable. |
| output_format | string · png / jpeg / webp | File format of the returned image. Defaults to png; applies to generation and editing. Transparent output needs png or webp. |
| n | integer · 1–10 | Images per request on gpt-image-2; cost scales linearly. GPT Image 2 Beta returns one image per call. |
| image_urls | array · up to 16 | Reference images for image-to-image and editing. Each one adds image input tokens. |
| mask_url | string · optional | Alpha-channel PNG whose pixel dimensions match the reference image; transparent areas are regenerated. |
| callback_url | string · optional | HTTPS webhook that receives the task result, as an alternative to polling. |
Text rendering is improved, but exact placement, clarity, and structured composition can still miss the prompt.
Custom sizes can reach a 3840 px edge, but outputs with more than 3,686,400 total pixels—the pixel count of 2560x1440—are currently experimental.
background: "transparent" is available, but still in Preview — results may be unstable.
OpenAI notes that complex prompts may take up to two minutes to process.
Token billing is the part of GPT Image 2 most teams find hardest to budget: image output tokens scale with resolution and quality tier, every reference image adds input tokens, and the prompt itself is metered. EvoLink turns that into a quoted number before you commit — the pricing calculator and the Playground both price the exact size, resolution, quality, and reference combination you are about to send.
That changes how you test. Iterate prompts on the low tier, watch what each change does to the estimate, and only promote the final render to high — instead of discovering the cost pattern on the invoice at the end of the month.
The pricing section on this page lists the EvoLink rate next to the official OpenAI rate for every token type — image output, image input, cached input, and text — and shows the current savings on each line, backed by EvoLink's lowest-price guarantee.
Because the comparison is computed from live SKU prices rather than a marketing claim, you can verify it at the moment you integrate, and re-check it whenever your volume grows enough for the difference to matter.
gpt-image-2 and gpt-image-2-beta are served side by side: the official token-billed route with the full 1K/2K/4K, quality-tier, mask, and multi-reference surface, and a flat-rate route where every 1K image costs the same regardless of prompt length or references.
In practice, teams put drafts, prompt iteration, and bulk 1K jobs on the flat route where cost is perfectly predictable, and reserve the token route for 2K/4K finals, masks, and reference-heavy edits. Switching is a model-ID change, not a vendor migration.
The same API key that calls gpt-image-2 also calls Seedance for video, Nano Banana 2 and Seedream for alternative image routes, and the LLM catalogue — one balance, one task history, one authentication path.
That matters for image work specifically because pipelines rarely end at a still: generate the key frame with GPT Image 2, animate it with Seedance, and fall back to another image model when a job fits it better — without maintaining separate provider accounts, keys, and billing records.
Every generation reports submitted, processing, completed, or failed, with error details in the response and the console. Failed tasks are not treated as successful billable generations, so a bad batch of jobs does not quietly become a bill.
For production systems this is the difference between "request sent" and "image delivered": poll the task, verify the final state, and reconcile spending against per-task records in the console when you need to explain a number.
Compare GPT Image generations under one EvoLink account and API key, then choose by output quality, latency, and cost.
GPT Image 2
CurrentGenerate high-fidelity images from text prompts or reference images, with strong poster layouts, data-dense infographics, character sheets, product creative, and mask-guided editing workflows.
GPT Image 1.5
Previous OpenAI image route with fixed 1024-class sizes, token billing, and settled production behaviour.
View API
Nano Banana 2
Google image model for high-quality generation and flexible creative workflows.
View API
Nano Banana Pro
Premium Nano Banana route for higher-fidelity images and production-grade output.
View API
Seedream 5.0 Pro
BytePlus image route with 1K/1.5K/2K tiers, multi-reference editing, and layer decomposition.
View API
Grok Imagine Image 2.0
xAI image route where the number of input images picks the mode, with 1K/2K output and up to 10 images per request.
View APIThe request contract, size and resolution interplay, async polling, and token billing worked through with real numbers.
What actually changed between the two routes, and which one to keep in production.
The fast GPT Image 2.5 variant OpenAI released on September 8, 2026, live on EvoLink with five quality tiers up to max and the same token pricing as GPT Image 2.
The precision GPT Image 2.5 variant for workflows where editing control matters most — same endpoint, parameters and token rates as Flare.
GPT Image 2 is the API-accessible image model released alongside ChatGPT Images 2.0, with model ID gpt-image-2. ChatGPT may combine image generation with reasoning, web search, or multi-image orchestration, so the raw API model should not be treated as identical to every ChatGPT product workflow.
They are separate EvoLink route variants. gpt-image-2 uses the token-billed official route with low/medium/high quality, 1K/2K/4K resolution, explicit WxH pixels, n up to 10, and an inpainting mask. gpt-image-2-beta is EvoLink's alternative fixed-price route for 1K output and one image per call; it is not an official OpenAI model ID.
It is billed by token. Image output tokens dominate the bill and scale with resolution and quality tier; image input tokens are added per reference image; text input tokens come from the prompt. The Playground estimates the selected combination before you submit. If you want one fixed per-image price instead, GPT Image 2 Beta bills a flat rate per call, and the pricing calculator estimates your per-image cost.
The pricing section compares the current EvoLink rate with the OpenAI list price and shows any savings available at the time you check. Use the calculator for your exact size, quality, and reference-image combination instead of relying on a universal per-image claim.
With the Image API, choose gpt-image-2 directly for generation or editing. With the Responses API, choose a mainline model that supports the image_generation tool; the tool handles image-model selection for conversational or multi-step workflows. EvoLink provides its own unified route for direct model access.
Yes. Use the Playground to test output quality, estimate credits, and inspect the request JSON before wiring it into your product.
size picks the shape - auto, one of 15 aspect ratios, or explicit WxH pixels. resolution picks the pixel budget for ratio mode only, and is ignored when size is auto or explicit pixels.
Yes. Every image in image_urls adds image input tokens. A mask is reserved as one extra input image at quote time and settled against the real upstream usage.
Width and height must be multiples of 16, total pixels must fall between 655,360 and 8,294,400, no edge may exceed 3840 px, and the aspect ratio must stay between 1:3 and 3:1. Outputs with more than 3,686,400 total pixels—the pixel count of 2560x1440—are currently experimental.
On gpt-image-2 yes - n accepts 1 to 10 and the cost scales linearly. GPT Image 2 Beta returns a single image per call.
mask_url takes a PNG with an alpha channel whose pixel dimensions match the reference image exactly. Transparent pixels mark the area to regenerate, opaque pixels are preserved. Sent without a reference image, the mask is dropped.
Check the task status and error details in the response or console. Failed tasks should not be treated as successful billable generations.
Complex prompts may take up to two minutes. Text placement, recurring-character consistency, and precise structured composition can still need iteration.