Image model
GPT Image 2 and 1.5, running in your browser
Two OpenAI image models, four rows in the picker: text to image and image to image for each. A 1K render on GPT Image 2 costs 15 credits, a Standard square on 1.5 costs 10, and the figure sits on the button before you press it. No API key, nothing to install.
What a render costs
| GPT Image 2 — text to image | 15 credits at 1K, 26 at 2K. Any of the five aspect ratios: 1:1, 16:9, 9:16, 4:3, 3:4. |
|---|---|
| GPT Image 2 — image to image | 15 credits at 1K, 26 at 2K. Up to 10 reference images in one render. |
| GPT Image 1.5 — text to image | 10 credits at Standard, 55 at High. Square 1:1 only, and no resolution setting at all. |
| GPT Image 1.5 — image to image | 10 credits at Standard, 55 at High. Up to 16 reference images, the highest cap of any image model here. |
| Images per submission | 1, 2 or 4. Each one is a separate render and a separate charge, so four at High on 1.5 comes to 220 credits. |
| Where credits come from | A plan, of which Solo carries 2,000 credits a month, or a one-off pack starting at 800 credits. A render that fails at the provider is refunded automatically. |
What it is good at, and what it is not
Worth picking it for
- Words inside the picture. Ask for a shop sign, a book cover, a product label or a UI mock with specific wording on it and GPT Image spells it more reliably than anything else in the picker.
- Long constraint lists. It reads a prompt closer to literally than the Nano Banana or Seedream rows, so "low angle, overcast, no people, no signage" tends to survive instead of being averaged into a nice-looking photo.
- Anything not square, on GPT Image 2: all five aspect ratios, 15 credits at 1K or 26 at 2K.
- Reference-heavy editing, on GPT Image 1.5: 16 reference images per render, against 10 on GPT Image 2, both Nano Bananas and every Seedream row.
Pick something else for
- Non-square work on GPT Image 1.5. Its aspect enum is 1:1 and nothing else, so the 16:9 and 9:16 buttons go dead the moment you select it.
- Sharpness on 1.5 without paying for it. There is no 1K/2K control on that version at all, only Standard and High, and High costs 55 credits against Standard's 10.
- Masters above 2K. The resolution row here offers 1K and 2K, so neither version will hand you a 4K file.
- Cheap volume. Nano Banana runs 10 credits flat and Seedream 5.0 Lite 14, both at every aspect ratio, so a hundred throwaway variations cost less elsewhere.
1.5 and 2 are not a ladder
The numbering suggests 2 supersedes 1.5. In the picker they behave more like two tools that happen to share a name. GPT Image 2 takes a real resolution setting, 1K or 2K, and renders any shape you ask for. GPT Image 1.5 has no resolution field in its schema at all: it sizes output through a quality tier, Standard or High, and it only ever returns a square.
That single difference decides most of the choice for you. A 9:16 wallpaper, a 16:9 header, a 4:3 slide: those are GPT Image 2 jobs, 15 credits at 1K. Avatars, album art, app icons, sticker sheets, anything already square: 1.5 at Standard costs 10 credits, which is the cheapest render in the whole image catalogue, tied with the original Nano Banana.
Each version appears twice in the list, once for text to image and once for image to image. Direction does not change the price on either one.
The High tier costs 5.5 times Standard
GPT Image 1.5 is the only model in this app with a live quality selector. Several Seedream rows have quality tiers in their API too, but no per-tier price is configured for them, so the app never offers the choice and always bills the base tier. On 1.5 the choice is real and it is steep: Standard 10 credits, High 55.
5.5 is the literal multiplier in the pricing row, and the credit figure printed on each tier button comes from the same resolver that charges you, so there is nothing to discover at submit time. What makes it a genuinely hard call is what else 55 credits buys. GPT Image 2 at 2K is 26. Seedream 5.0 Pro at 2K is 36. Nano Banana Pro is 45 at any size. High on 1.5 is the single most expensive image render sold here, and it is still square.
Worth it when the deliverable is square and the detail is the product: a print you will actually print, a portrait where hair and skin fall apart at Standard, small type that has to stay readable inside the frame. Not worth it while you are still exploring. Asking for four images at High in one submission costs 220 credits, because quantity fires four separate renders and charges each one. Search at Standard, then re-run the keeper at High.
What the form actually asks you
Open the image workspace and pick one of the four GPT Image rows. Which controls appear next depends on which one you picked. GPT Image 2 gives you an aspect row with all five shapes live, a 1K/2K pair and a count of 1, 2 or 4. GPT Image 1.5 shows the same aspect row with only 1:1 selectable, swaps resolution for Standard and High, and prints the credit cost on each of those two buttons.
Then the prompt, then Generate. The number on the Generate button is the total for the settings in front of you. It leaves your balance when the job is submitted rather than when it lands, and if the render fails at the provider the credits come back on their own, no support ticket. Writing and rewriting prompts is free.
Finished images drop into your library as they arrive. None of it needs a key or a local install: the model runs on hosted infrastructure and the browser is the whole client. Credits come from a plan or a one-off pack, priced on the pricing page.
Editing a photo you already have
The image-to-image rows start from your file instead of from a description. Upload, say what should change, generate. GPT Image 1.5 accepts up to 16 reference images in a single render and GPT Image 2 up to 10, and that cap matters more than it sounds: references are not only the thing being edited. They are also how you hand over a palette, a product from three angles, or the one face that has to stay the same across a set.
Direction does not change the price. Image to image costs the same as text to image on both versions, so there is no penalty for working from a reference. For the technique itself, and for how the other models here handle the same job, see the how image-to-image works here.
GPT Image against Nano Banana and Seedream
Three families cover image generation here and they are not interchangeable. GPT Image is the literal one: it follows a list of constraints and puts readable text in a frame. Nano Banana is the conversational editor, better than most at keeping a subject recognisable across a run of edits. Seedream is the volume option, 14 credits flat on 5.0 Lite, with the widest stylistic range of the three.
The price ranking flips depending on which column you read. GPT Image 1.5 at Standard is the cheapest render available here. GPT Image 1.5 at High is the most expensive. GPT Image 2 at 1K, 15 credits, sits under Nano Banana 2, Nano Banana Pro and Seedream 5.0 Pro, which is friendlier than the model's reputation suggests.
Five prompts to start from
Written against this picker, not lifted from a prompt pack. The first three are square and run on either version. The fourth needs GPT Image 2 for its shape; the fifth belongs on an image-to-image row.
- Text in the frame
A hand-painted enamel shop sign photographed straight on, dark green ground, cream serif lettering reading exactly: NORTH POINT COFFEE. A smaller line beneath in the same cream reading exactly: EST. 1974. Chipped paint at two corners, one screw missing top left. Flat overcast daylight, no shadow. No people, and no other text anywhere in the frame.
Spell the words out and use "reading exactly". Then close the door with "no other text anywhere in the frame", because the common failure here is invented extra lettering rather than a misspelling.
- Product on a surface
A matte black ceramic pour-over dripper on a raw concrete slab. Three-quarter view from slightly above. Single soft window light from the left, deep shadow falling to the right. Shallow depth of field, the rear rim just out of focus. Square crop, object centred with even margin on all four sides. No hands, no steam, no branding.
Camera position, light direction and depth of field are named as separate instructions. GPT Image treats them as four things to satisfy instead of blending them into a generic product shot.
- Exclusion list
An empty municipal swimming pool at dusk. Drained, tiles cracked, a single folding chair on the far side. Cold blue-grey light, long shadows from a low sun. Photographic, 35mm, slight grain. Do not include: people, water, text, signage, birds, litter.
A colon-separated exclusion list at the end beats a scattering of "without" clauses. The literal reading that makes this model good at typography also makes it good at a checklist.
- Layout you can dictate
A flat vector diagram on an off-white ground, 16:9. Three labelled boxes in a horizontal row, connected by two arrows pointing right. Box labels from left to right, exactly: DRAFT, REVIEW, PUBLISH. Thin charcoal strokes of even weight, one accent colour (burnt orange) used only on the arrows. Generous margins. No drop shadows, no gradients, no extra icons.
Pick GPT Image 2 for this one; 1.5 returns squares only. Giving the exact label strings and a stroke description is what keeps it out of default stock-illustration territory.
- An edit, not a new picture
Keep the subject, the pose and the framing exactly as they are. Change only the background to a plain studio sweep in warm grey, and relight to match: soft key from camera left, gentle fill on the right cheek. Do not alter the face, the hair or the clothing. Keep the original crop.
On an image-to-image row, say what must not change before you say what must. "Keep the original crop" is worth adding every single time.
How it compares, in credits
Every figure is the live cost from the same pricing table the Generate button reads. Dearest is the highest tier this app will actually sell you, which is not always the highest tier the model has.
| Cheapest render | Dearest render | Shapes | Reference images | |
|---|---|---|---|---|
| GPT Image 2 | 15 at 1K | 26 at 2K | All five | 10 |
| GPT Image 1.5 | 10 at Standard | 55 at High | 1:1 only | 16 |
| Nano Banana 2 | 20 at 1K | 30 at 2K | All five | 10 |
| Nano Banana Pro | 45, no size tiers | 45, no size tiers | All five | 10 |
| Seedream 5.0 Pro | 18 at 1K | 36 at 2K | All five | 10 |
| Seedream 5.0 Lite | 14, no size tiers | 14, no size tiers | All five | 10 |
GPT Image FAQ
Is GPT Image 2 free?
Not here, and not anywhere genuinely running it. The account is free and needs no card, and prompts cost nothing to write or rewrite, but a finished render spends credits. The smallest pack is 800 credits: 80 renders at GPT Image 1.5 Standard, 53 at GPT Image 2's 1K rate, 14 at 1.5 High.
How much does GPT Image 2 cost per image?
15 credits at 1K, 26 at 2K, the same from a prompt or from a reference photo. GPT Image 1.5 is 10 at Standard and 55 at High. Quantity multiplies: asking for 4 fires four renders and charges four times. A render that fails is refunded automatically.
What is the difference between GPT Image 1.5 and GPT Image 2?
GPT Image 2 takes a 1K or 2K resolution and renders all five aspect ratios. GPT Image 1.5 has no resolution setting at all, returns squares only, and sizes output through a Standard or High tier. Reference caps go the other way: 16 on 1.5, 10 on 2.
Can GPT Image make a 16:9 or a vertical image?
GPT Image 2 can, at 1:1, 16:9, 9:16, 4:3 or 3:4. GPT Image 1.5 cannot: its API accepts a square and nothing else, so those buttons are disabled while 1.5 is selected rather than accepted and quietly ignored.
Is the High quality tier worth 55 credits?
For a square deliverable where detail is the point, often. While you are still exploring, no. High is 5.5 times Standard and the most expensive single image render sold here, above Nano Banana Pro at 45 and Seedream 5.0 Pro at 36. Iterate at Standard, re-run the keeper at High.
Is there a GPT Image 2 API?
No, and that is deliberate. SynthPulse is a web app rather than an API reseller, so there is no key to collect and no endpoint to call. You get the model behind a form, with prompt history, a visible credit balance and a library that keeps every render. If programmatic access is the actual requirement, this is not the product for it.
Can GPT Image edit a photo I already have?
Yes, on the image-to-image rows. Upload the picture, describe the change, and the model works from your file instead of starting over. GPT Image 1.5 takes 16 reference images per render and GPT Image 2 takes 10, so the edit target, a palette and a second angle can all go in at once. Same price as text to image.
What kind of prompts work best with GPT Image?
A complete description of the finished picture, then an exclusion list. Name the subject, the camera position, the light and the medium, and spell out any wording with the phrase "reading exactly". GPT Image follows a list more literally than the other models here, so "stunning" is wasted on it and "no people, no text, no signage" actually holds.