GPT Image 2
GPT Image 2 is OpenAI's image generation and editing model, released to the API on 21 April 2026 as gpt-image-2 and to ChatGPT as ChatGPT Images 2.0. It was the model behind the ChatGPT image generator until Images 2.5 replaced it on 8 September 2026, and it stays available on the API. It reasons before it draws, honours exact pixel sizes, makes transparent cut-outs and edits with masks.
Anvisha Pai, Co-founder & CEO, ModaTested OpenAI · released
›What we asked for
A self-portrait of GPT Image 2, OpenAI's image model, as a typographic poster: the words 'GPT IMAGE 2' set huge in a heavy grotesque, filled with a photorealistic scene of a city at dusk seen through the letterforms, smaller lines of type reading 'reasons before it draws' and 'any size, any shape', black background, Swiss poster layout.
openai/gpt-image-2
image_size: {"width":1360,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1360 by 1024 pixels
At a glance
- Any size
- Ask for exact pixel dimensions, anything from a 1:3 strip to a 3:1 banner up to 3840 on the long edge.
- Quality
- Low, medium or high. Low costs under a cent and takes seconds; high costs 30 times more and takes about two minutes. The ladder below shows what each looks like.
- Transparent backgrounds
- A background setting gives you a real cut-out PNG. No Nano Banana tier can do this.
- Mask edits
- Paint a mask over the part to change and it leaves the rest untouched.
- Reference images
- Up to 16 input images per edit request.
Other GPT Image models: GPT Image 2.5 Flare, GPT Image 2.5 Sunburst. Compare all 3.
Anvisha's take
GPT Image 2 is the model that got the hard composition exactly right, three ice cubes and all, and the one whose edit changed only what I asked. It is also the one I would no longer start a project on. At high quality it took about two minutes a picture on our run and, on fal's list, four times what 2.5 Flare charges for the same size and tier. Everything it does well the 2.5 models now do faster.
GPT Image 2 real output: the same 12 prompts every model gets
Every model on this site gets the same 12 prompts, so you can compare like with like. These are the pictures GPT Image 2 returned, one try each, nothing re-rolled or retouched. Generated on . The 12 pictures took 72 to 163 seconds each and cost $2.16 in total.
Spelling, kerning and hierarchy of three lines of set text.
›What we asked for
A concert poster that reads 'MODA SUMMER SESSIONS' in large bold type, with the subtitle 'Friday 12 June, 8pm, Roundhouse' and a smaller line 'Tickets at the door'. Risograph style, two colours, tangerine and navy, grainy ink texture.
openai/gpt-image-2
image_size: {"width":1024,"height":1360}, quality: high, output_format: png, num_images: 1
returned 1024 by 1360 pixels
Catalogue lighting, a clean seamless background and a believable contact shadow.
›What we asked for
Studio product photograph of a matte black ceramic pour-over coffee dripper on a seamless white background, soft top light, subtle contact shadow, centred, catalogue style.
openai/gpt-image-2
image_size: {"width":1024,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1024 by 1024 pixels
A natural face, correct finger count and hands holding an object.
›What we asked for
Close-up portrait of a woman in her 40s with short grey hair, laughing, holding a red ceramic mug with both hands, natural window light, 85mm lens, shallow depth of field.
openai/gpt-image-2
image_size: {"width":1024,"height":1360}, quality: high, output_format: png, num_images: 1
returned 1024 by 1360 pixels
A flat single-colour mark with a legible wordmark and no gradients or photo texture.
›What we asked for
A flat vector logo mark for a bicycle repair shop called 'Spoke & Chain': a single-colour icon combining a chain link and a wheel spoke, with the wordmark below in a geometric sans-serif, black on white.
openai/gpt-image-2
image_size: {"width":1024,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1024 by 1024 pixels
Many distinct objects and people in one frame without merging or duplicating them.
›What we asked for
A crowded Saturday farmers market seen from slightly above: at least twelve stalls with striped awnings, shoppers carrying tote bags, a dog on a leash, crates of oranges, a musician with a guitar, bunting overhead, late morning light.
openai/gpt-image-2
image_size: {"width":1360,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1360 by 1024 pixels
A named historical style applied faithfully rather than a generic painterly look.
›What we asked for
A lighthouse on a rocky cliff during a storm, in the style of a Japanese ukiyo-e woodblock print, with visible wood grain texture, flat colour areas and a bokashi gradient sky.
openai/gpt-image-2
image_size: {"width":1024,"height":1360}, quality: high, output_format: png, num_images: 1
returned 1024 by 1360 pixels
Six labelled bars at the stated heights, in order, with readable axes.
›What we asked for
A clean bar chart titled 'Monthly active users' showing six months, January through June, with the values 120, 150, 180, 170, 220 and 260, labelled axes, white background, one accent colour, no 3D effects.
openai/gpt-image-2
image_size: {"width":1360,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1360 by 1024 pixels
A 21:9 composition that leaves the left third clear for headline text.
›What we asked for
Website hero banner: a wide panorama of a sunlit modern co-working space with plants, the left third of the frame kept as clear wall space for text, warm neutral palette.
openai/gpt-image-2
image_size: {"width":2016,"height":864}, quality: high, output_format: png, num_images: 1
returned 2016 by 864 pixels
Four numbered steps in the given order with their exact wording, plus an icon each.
›What we asked for
An infographic card titled 'How to brew pour-over' with four numbered steps in this order: 1. Rinse the filter, 2. Bloom for 30 seconds, 3. Pour in slow circles, 4. Wait 3 minutes. A simple icon beside each step, muted palette, clean sans-serif type.
openai/gpt-image-2
image_size: {"width":1024,"height":1360}, quality: high, output_format: png, num_images: 1
returned 1024 by 1360 pixels
Appetising, menu-grade food with correct textures and no melted anatomy.
›What we asked for
Overhead photograph of a margherita pizza on a wooden board, fresh basil leaves, one slice pulled out with the cheese stretching, restaurant menu photography, natural light.
openai/gpt-image-2
image_size: {"width":1024,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1024 by 1024 pixels
Exact counts and left-right placement of five named objects.
›What we asked for
A glass of water on a wooden table with exactly three ice cubes in it, a yellow pencil resting across the rim of the glass, a blue book to the left of the glass and a red apple to the right, photographed from table height.
openai/gpt-image-2
image_size: {"width":1360,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1360 by 1024 pixels
Change one object's colour and add a part while leaving background, light and framing untouched.
›What we asked for
Change the coffee dripper to a deep forest green and add a small brass handle on its right side. Keep the background, lighting and composition exactly as they are. (applied to the product on white result above)
openai/gpt-image-2/edit
image_size: {"width":1024,"height":1024}, quality: high, output_format: png, num_images: 1
returned 1024 by 1024 pixels
How the cost under each picture was worked out: fal's published per-image figure for gpt-image-2 at high quality, nearest listed size 1024x1536; fal bills this endpoint by tokens and does not return the charge.
GPT Image 2 settings that matter
Quality
The text was correct at every level, so quality here means texture and finish, not whether the words come out. Low is a fortieth of the price of high and came back in about 20 seconds; medium in 45 to 50; high in two to three minutes. Medium is the sensible default and low is fine for layout drafts. If speed matters, the 2.5 models do the same jobs in a fraction of the time.
Text in an image
›What we asked for
A concert poster that reads 'MODA SUMMER SESSIONS' in large bold type, with the subtitle 'Friday 12 June, 8pm, Roundhouse' and a smaller line 'Tickets at the door'. Risograph style, two colours, tangerine and navy, grainy ink texture.
openai/gpt-image-2
image_size: {"width":1024,"height":1360}, quality: low, output_format: png, num_images: 1
returned 1024 by 1360 pixels
Food photography
›What we asked for
Overhead photograph of a margherita pizza on a wooden board, fresh basil leaves, one slice pulled out with the cheese stretching, restaurant menu photography, natural light.
openai/gpt-image-2
image_size: {"width":1024,"height":1024}, quality: low, output_format: png, num_images: 1
returned 1024 by 1024 pixels
GPT Image 2 pricing
| Where | Unit | Price | Source |
|---|---|---|---|
| OpenAI API | Per million image output tokens Text input is $5, image input $8 per million. OpenAI publishes no per-image table; cost depends on size, quality and prompt length. | $30.00 | OpenAI API pricing, read |
| OpenAI API, Batch | Per million image output tokens Batch halves every rate. | $15.00 | OpenAI API pricing, read |
| fal | 1024 by 1024, high quality fal's published typical figures for the same model. Low is $0.006 and medium $0.053 at this size. fal bills by tokens, so these are typical, not fixed. | $0.211 | fal model page, read |
| fal | 3840 by 2160, high quality The largest listed size. Low is $0.012 and medium $0.101 at 4K. | $0.401 | fal model page, read |
Is GPT Image 2 free?
ChatGPT's free plan includes image generation, and OpenAI says ChatGPT Images 2.5 is rolling out to every ChatGPT tier including free. The API has no free tier. OpenAI publishes no fixed daily number for free accounts; the cap moves with demand and the banner inside ChatGPT is the only authoritative figure.
- •ChatGPT Free: image generation is included. OpenAI says Images 2.5 is rolling out to ChatGPT, ChatGPT Work and Codex users across all tiers. Limits are dynamic and unpublished.
- •ChatGPT Plus and Pro: higher limits, again unpublished as a fixed number.
- •OpenAI API: no free tier. Every image is billed per token.
- •fal, Replicate and other hosts: pay per image, no free allowance beyond any sign-up credit.
Where to use GPT Image 2
- ChatGPT · Now runs Images 2.5; GPT Image 2 is no longer selectable there.
- OpenAI API · Model id gpt-image-2 on the generations, edits and batch endpoints.
- fal · openai/gpt-image-2 and its edit endpoint, with published per-image figures.
- Replicate · Hosted with a narrower set of controls: three aspect ratios, no transparent background.
- Moda · Available as the GPT Image 2.0 model inside the Moda design agent, with the quality control and custom dimensions.
Generate with GPT Image 2 in Moda
Pick GPT Image 2 in the Moda design agent, then edit the result as vectors, text and layers rather than a flat image.
Open Moda →What we noticed running GPT Image 2
- •It draws real logos when it thinks they fit. The self-portrait at the top of this page arrived with an OpenAI mark in the corner that the prompt never asked for. Check anything brand-adjacent for trademarks before it ships.
- •Counts and placement were exact. The glass holds three ice cubes, the pencil rests across the rim, the blue book is on the left and the apple on the right. This is the prompt most models fail, and it was the one it got fully right. See the hard composition picture.
- •It renders what you asked for and nothing more. The infographic carries the four step headings with their exact wording and no invented body copy, and the chart shows six bars at the stated values in a designed typeface rather than a plotting-library default. See the infographic with steps picture.
- •Set text is reliable and it will decorate around it. All three poster lines came back correctly spelled, with an unrequested but on-brief illustration of a round venue and a crowd underneath. See the text in an image picture.
- •Style prompts arrive signed by someone real. The ukiyo-e print carries a fabricated title cartouche, a red seal and a signature reading 広重画, 'painted by Hiroshige'. Treat any script in a stylised output as a liability and remove it. See the specific art style picture.
- •Negative-space instructions are read literally. Asked to keep the left third clear for text, it rendered a blank wall filling that third, which is usable but is a wall, not a composition with room in it. See the wide aspect ratio picture.
- •Product shots are catalogue-clean but flat. The white background is pure and the object is centred, with almost no contact shadow despite the prompt asking for one, and it added a handle the prompt did not mention. See the product on white picture.
- •Edits are precise. The recolour kept the dripper's silhouette, fluting and base identical and swapped the handle for a brass one in the same position; only the requested parts changed. See the edit an existing image picture.
- •It is slow at high quality. Across the twelve generations the fastest returned in 72 seconds and the slowest in 163, with a median around 115. The same prompt set took Nano Banana Pro 15 to 87 seconds. Use low or medium quality for drafts, or the 2.5 Flare model, if latency matters.
- •Custom sizes are honoured to the pixel. Requests for 1024 by 1360, 1360 by 1024 and 2016 by 864 came back at exactly those dimensions, which no fixed-aspect-ratio model can promise.
- •The price depends on the size you ask for, not just the quality. At high quality fal lists 1024 by 1024 at $0.211 but 1024 by 1536 at $0.165, so a taller image can cost less than a square one; the token count follows the model's internal tiling, not the pixel count you would expect.
- •Small incidental text is mostly right and occasionally not. Stall signs in the market read correctly ('Fresh cut flowers', 'Support local farmers', 'Olive oil') while one or two lines of fine print are close but not exact. See the busy scene picture.
GPT Image 2 limits and API parameters
| Limit | Value |
|---|---|
| Output size | Any width and height in multiples of 16, longest edge 3840 pixels, 655,360 to 8,294,400 pixels in total, aspect between 1:3 and 3:1. |
| Input images for edits | Up to 16 per request. fal edit API reference, read |
| Rate limits | OpenAI API tier 1: 100,000 tokens and 5 images per minute, rising to 8 million tokens and 250 images per minute at tier 5. OpenAI model page, read |
| Endpoints | Image generation, image edit and batch on the OpenAI API. |
›Every setting, for developers
OpenAI API
OpenAI image generation guide, read
| Parameter | Type | Values | Default |
|---|---|---|---|
| prompt | string | Any text Required. | |
| size | string | 1024x1024, 1536x1024, 1024x1536, or any WIDTHxHEIGHT in multiples of 16 Aspect between 1:3 and 3:1, longest edge 3840, total pixels 655,360 to 8,294,400. Above 2560x1440 is marked experimental. | auto |
| quality | enum | low, medium, high, auto The 2.5 models add xhigh and max. | auto |
| background | enum | transparent, opaque, auto Transparent needs png or webp output. | auto |
| output_format | enum | png, jpeg, webp | png |
| output_compression | integer | 0 to 100 jpeg and webp only. | |
| moderation | enum | auto, low | auto |
| n | integer | Number of images per request | 1 |
| partial_images | integer | 0 to 3 Streaming previews. | |
| image and mask | files | Up to 16 input images, optional mask Edits endpoint. |
fal (openai/gpt-image-2)
fal API reference, read
| Parameter | Type | Values | Default |
|---|---|---|---|
| prompt | string | Any text Required. | |
| image_size | enum or object | square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9, auto, or {width, height} | landscape_4_3 |
| quality | enum | auto, low, medium, high | high |
| background | enum | auto, transparent, opaque | auto |
| output_format | enum | jpeg, png, webp | png |
| num_images | integer | 1 or more | 1 |
| image_urls | list of URLs | Up to 16 Edit endpoint only, where it is required. | |
| mask_url | URL | A mask image Edit endpoint. |
What GPT Image 2 will not draw
OpenAI's usage policies apply and every prompt and output passes a moderation stage. The API's moderation setting has two levels, auto and low, and a blocked request says whether the prompt or the picture tripped it and which category: harassment, self-harm, sexual content or violence. In ChatGPT there is no override. Real people, minors in unsafe contexts and violent or sexual content are the practical boundaries for design work. Every image carries C2PA metadata and an invisible watermark.
Alternatives to GPT Image 2
Frequently asked questions
What is GPT Image 2?
GPT Image 2 is OpenAI's image generation and editing model, released as gpt-image-2 on the API and as ChatGPT Images 2.0 in ChatGPT on 21 April 2026. It was the model behind the ChatGPT image generator until Images 2.5 replaced it on 8 September 2026.
How much does GPT Image 2 cost?
On the OpenAI API it is token-billed at $30 per million image output tokens, read on 9 September 2026. fal publishes typical per-image figures for the same model: $0.006 low, $0.053 medium and $0.211 high at 1024 by 1024, rising to $0.401 for 4K at high.
Is the ChatGPT image generator free?
Yes, with limits. ChatGPT's free plan includes image generation and OpenAI says Images 2.5 reaches every tier including free. The daily cap is not published and moves with demand. The API itself has no free tier.
Low, medium or high quality?
Medium for most work: it is a tenth of the price of high and, on our two ladder prompts, most of the way there. High for final print or type-heavy work. Low for layout drafts, where it costs under a cent.
What changed in GPT Image 2.5?
OpenAI released two API models on 8 September 2026: GPT Image 2.5 Flare, the fast default, and GPT Image 2.5 Sunburst, the precision tier. Both add xhigh and max quality settings and OpenAI cites up to 50 percent lower latency than Images 2.0.
Can GPT Image 2 generate transparent backgrounds?
Yes. Set background to transparent and output format to png or webp. Nano Banana models have no equivalent.
Is GPT Image 2 still available in ChatGPT?
No. ChatGPT moved to Images 2.5 on 8 September 2026 and there is no model picker for images. GPT Image 2 stays available on the OpenAI API as gpt-image-2, and on fal and Replicate.
How this page is made
Prices, limits and settings are read from the linked vendor and host pages on the dates shown. Every picture was generated by us through fal with the request shown under it, and the original files are kept unedited. A generation that failed is shown as a failure. Model and vendor names are used to identify the products; no vendor imagery appears on this page.
