GPT Image 2 for Free: How to Try It and What It Costs

Partly, depending on where you use it. Here's what each route gives you as of October 2026.

Spicy Studio: every new account gets 25 free credits, no credit card needed. GPT Image 2 costs 5 to 60 credits per image, so the free credits cover five images at 1K low quality, two at 1K medium or one at 2K medium.

ChatGPT: the free plan includes image creation with its own usage limits. OpenAI rolled Images 2.5 out to all ChatGPT tiers on September 8, 2026, so ChatGPT is no longer the obvious place to test GPT Image 2 itself.

OpenAI's API: pay as you go, per token, with your own OpenAI account, API key and code.

Playground: its free plan includes 3 generations a month, shared across five models, GPT Image 2 among them.

The rest of this guide shows how to spend those free credits well, which settings matter, and when GPT Image 2.5 is the better pick.

GPT Image 2 (API name gpt-image-2) is the OpenAI model behind ChatGPT Images 2.0. OpenAI released it on April 21, 2026, in ChatGPT, Codex and the API at once (OpenAI's announcement). What OpenAI highlighted:

Small and dense text. Lettering, icons, UI elements and crowded compositions that earlier models tended to garble.

Non-Latin scripts. OpenAI reported significant gains in Japanese, Korean, Chinese, Hindi and Bengali text.

Flexible formats. Aspect ratios from 3:1 to 1:3, and output up to 2K in the API, with larger sizes in beta.

Fresher world knowledge. A knowledge cutoff of December 2025, which helps with explainers and infographics.

Thinking in ChatGPT. With a thinking model selected, Images 2.0 could search the web, check its own work and return up to eight related images from one prompt. At launch that was limited to Plus, Pro and Business subscribers.

OpenAI also named the weak spots: tasks that need a full physical model of the world (origami steps, Rubik's Cubes), details on hidden or reversed surfaces, very dense repeating textures, and labels or arrows on diagrams, which still need checking.

1. Open the model. This link opens GPT Image 2 directly. You can also start in the Image Lab and choose it from the model list, or browse all models first. Sign in or create a free account; the 25 credits arrive on sign-up.

2. Write the prompt. Describe the subject, setting, style and light, and put every word that must appear in quotes.

3. Add references if you have them. Up to 16 images: a product shot, a logo, a character sheet, a mood board. On GPT Image 2 they don't change the price.

4. Pick the settings. Choose 1K (the default) or 2K, then low, medium (the default) or high quality. For the shape there's square 1:1, landscape 3:2, 4:3, 5:4, 16:9 and 21:9, or portrait 2:3, 3:4, 4:5 and 9:16. You see the credit cost before you generate.

5. Generate, then change one thing at a time so you can tell which edit helped.

Credits per image:

1K: 5 (low) · 10 (medium) · 34 (high)

2K: 10 (low) · 25 (medium) · 60 (high)

References are free on this model, whether you upload one or sixteen.

A plan for the 25 free credits: three 1K low drafts to settle the composition (15 credits), then one 1K medium version of the best (10). That uses the full allowance and is enough to judge whether GPT Image 2 suits your work.

After that, the smallest top-up is Spark at €2 for 90 credits, which never expire, and monthly plans start with Spicy Basic at €10 for 600 credits. Details are on the pricing page. The same balance pays for every image and video model and for editing tools like the Background Remover.

Three starting points that lean on what OpenAI built the model for:

Bilingual signage: “A chalkboard sign outside a ramen bar, hand-lettered ‘KUMO NOODLE BAR’ across the top, below it ‘Open until midnight’ in English and the same line in Japanese, warm street light, light rain.”

A product composite from references: “Place the sneaker from the first image and the shoebox from the second on a concrete plinth, soft studio light, the lid reading ‘RUNNER’ exactly as in the reference.”

A UI mockup: “A phone screen showing a habit-tracker app called ‘Steady’, a weekly streak bar across the top, three task cards below, clean sans-serif type, light theme.”

A few habits make results more predictable:

- Quote the exact copy and describe the type: weight, case, alignment, placement. - Lead with the main subject, then setting, style, light and camera. - Upload a product or logo instead of describing it. OpenAI's docs say gpt-image-2 processes every image input at high fidelity, so it stays close to what you give it. - Proofread every word before you publish.

The prompt engineering guide covers the full method.

OpenAI followed GPT Image 2 with GPT Image 2.5 on September 8, 2026, and both are in the Image Lab. GPT Image 2.5 comes in two versions, Flare for fast everyday work and Sunburst for precise edits, with 1K, 2K and 4K output, five quality levels (it adds xhigh and max) and 15 aspect ratios.

Choose GPT Image 2.5 when you need 4K, the extra detail of xhigh or max, a very wide or tall frame such as 3:1 or 1:3, or Sunburst's precision on a final edit. It's also slightly cheaper at matching settings: 9 credits against 10 at 1K medium, 56 against 60 at 2K high.

Stay on GPT Image 2 when you work with many references at 1K or 2K. GPT Image 2.5 adds 1 credit per reference and GPT Image 2 doesn't, so twelve references at 2K medium cost 25 credits on GPT Image 2 and 35 on GPT Image 2.5. It's also the easy choice if your prompts are already tuned for it.

The GPT Image 2.5 guide compares Flare and Sunburst in detail.

Spicy Studio wrote this comparison, and all three models sit in its Image Lab on the same credits, so the question is which one to route each job to, not which subscription to buy.

GPT Image 2: use it when the image carries a lot of text, non-Latin lettering or a composite built from many references. 1K or 2K, 5 to 60 credits, up to 16 references.

Nano Banana Pro: made for quick iteration, at 12 credits for 2K or 16 for 4K, with up to 14 references. The Nano Banana Pro guide goes deeper. The newest version, Nano Banana 2.1, is the Image Lab's default model.

Spicy Imagine Pro: runs ByteDance's Seedream 5.0 and is designed for edits that keep faces, lighting and color tones from your references. 8 credits at 2K, 15 at 4K, up to 10 references, which you call “Figure 1”, “Figure 2” and so on in the prompt. It's the way to use Seedream for everyday work on Spicy Studio today.

A practical routing: test layouts cheaply (GPT Image 2 at 1K low is 5 credits, Spicy Imagine Pro at 2K is 8), then make the final, text-heavy version on GPT Image 2 or 2.5 at high quality. For the wider lineup, see AI image generators compared.

No transparency setting. GPT Image 2's options in the Image Lab don't include a transparent background. Generate on a plain, contrasting backdrop and run the Background Remover for a transparent PNG (4 credits).

No 4K. GPT Image 2 stops at 2K here. Use GPT Image 2.5 for native 4K, or Image Upscale to take a 2K image to 4K for 10 credits or 8K for 12.

No loose-interpretation dial. OpenAI doesn't allow input fidelity to be changed for gpt-image-2. That helps when a product or logo must stay exact; if you only want loose inspiration, describe the look in words rather than uploading it.

Text still needs a proofread. OpenAI says labels and diagrams may need review, especially precise arrows and part labels.

High quality is slower. It costs more and takes longer than low or medium, which is why drafts belong at low.

Cleanup happens in the same studio. The Eraser removes a stray object for 8 credits. The Watermark Remover (4 credits) is only for images you own or have the rights to edit.

Content rules apply. GPT Image models filter prompts and outputs under OpenAI's content policy.