First-try ChatGPT images rarely match what you pictured. With gpt-image-2 (the model behind ChatGPT Images 2.0, shipped April 21 2026), the gap is smaller than it used to be — but a usable hero image still comes from a workflow, not one prompt. This guide is the loop I use to go from a one-line goal to a final, on-brand asset in 10-15 minutes.
TL;DR
- ChatGPT’s current image model is
gpt-image-2(Images 2.0). It renders at native 2K (up to 2048px), reads and writes accurate text including Chinese/Japanese/Korean, and keeps a character or product consistent across up to 8 outputs from one prompt. - Use Thinking mode for hero/brand work (it plans, optionally searches the web for references, and critiques its own draft) and Instant mode for fast throwaway rounds.
- Pick your aspect ratio before you start.
gpt-image-2supports 3:1 down to 1:3 (1:1, 16:9, 9:16, 4:5, and so on) — generating square then cropping to 16:9 wastes outputs. - Iterate by changing one variable per round and editing in conversation (“keep everything, swap the laptop for a phone”) instead of re-prompting from scratch.
- Free tier is for previews only (about 2-3 images/24h, square 1024×1024, standard quality). Plan around Plus or higher if images are part of the job (see the limits table below).
Who this is for
Content creators, indie devs, and marketers who need a specific image on a deadline and cannot rely on a lucky roll. If you can describe the result in words, this workflow fits.
When ChatGPT images are the right tool — and when they are not
Reach for it for a single hero image, social-post visuals, blog illustrations, simple ad creative, product mockups, or anything you can describe in a sentence. As of June 2026, gpt-image-2 also handles in-image text reliably, so headline graphics and labeled diagrams are finally on the table.
Do not use it for pixel-precise UI mockups, exact reproduction of a registered logo, photorealistic faces of real people, or any asset needing a legal model release. Use a designer, Figma, or a stock-with-license source for those.
What changed with ChatGPT Images 2.0
If your mental model of ChatGPT images is still DALL-E 3, it is out of date — DALL-E 2 and DALL-E 3 were retired on May 12 2026. The current model is meaningfully different to work with:
| Capability | DALL-E 3 (retired) | gpt-image-2 (June 2026) |
|---|---|---|
| Max resolution | 1024×1792 | Native 2K, up to 2048×2048 |
| Aspect ratios | 1:1, 16:9, 9:16 | 3:1 to 1:3 (1:1, 16:9, 9:16, 4:5, 3:2, etc.) |
| In-image text | Unreliable, Latin only | ~99% character accuracy, multilingual incl. CJK |
| Editing | Re-prompt from scratch | Conversational inpainting, object removal, background swap |
| Consistency | One image at a time | Up to 8 outputs, same character/product/scene |
| Reasoning | None | Thinking mode plans and self-critiques before drawing |
Two practical takeaways: you can now ask for text in the image (within reason), and you can edit a generated image by describing the change instead of rewriting the prompt.
Instant vs Thinking mode
gpt-image-2 runs in two modes, mirroring the GPT-5.5 Instant/Thinking split:
- Instant returns a fast single-shot render. Use it for exploration rounds where you just want to see how the model reads your description.
- Thinking runs a longer loop: it breaks the prompt into steps, can search the web for references, generates candidates, critiques them, and returns the best one. Use it for the final hero image and anything with text or brand constraints.
Rule of thumb: explore in Instant, finish in Thinking.
Step by step
- Write a one-line goal — subject + style + mood + use-case. Example: “Hero image for a productivity-app blog, calm, isometric, blue-and-orange palette, 16:9.”
- Set the aspect ratio up front. Tell ChatGPT the exact ratio (1200×630 / 1.91:1 for a blog OG card, 1080×1350 / 4:5 for an Instagram portrait). Do not default to square.
- Run an Instant exploration round of 2-3 images to see how the model reads “calm isometric.” Treat this as research, not the deliverable.
- Pick the closest one and edit in conversation. Say what to keep and what to change: “Keep the lighting and composition, change the laptop to a phone, remove the coffee cup.” The model edits that region instead of starting over.
- Iterate 3-5 rounds, one variable per round (lighting, then palette, then a single object). Note what changed each round so you can roll back.
- Finish in Thinking mode at production size. Ask for the final image at the exact ratio and 2K resolution, then request any in-image text explicitly.
- Run a release check: text legibility, brand-safe colors, no anatomical glitches, no unintended logos, correct dimensions.
A real OG-image run
Goal: a 1.91:1 blog OG card. Prompt 3 Instant variants → pick one → edit twice in conversation (first the background gradient, then the headline placement) → switch to Thinking mode and regenerate at 1200×630 with the headline text baked in → one pass in any image editor to nudge contrast. Total: about a dozen minutes, four to five generations.
ChatGPT image limits by plan (June 2026)
Image generation is metered separately from chat. These are the practical caps as of June 2026; OpenAI adjusts them, so confirm in-app before you commit a workflow.
| Plan | Price (USD/mo) | Image generation | Resolution / quality |
|---|---|---|---|
| Free | $0 | ~2-3 images per 24h (rolling) | 1024×1024 only, standard |
| Go | $8 | Higher than Free, below Plus | Standard + larger sizes |
| Plus | $20 | ~50 images per 3-hour window | Full 2K, all ratios |
| Pro | $100 / $200 | Effectively unlimited (abuse guardrails) | Full 2K, priority |
If images are part of your actual deliverables, Plus is the floor — the Free tier’s 2-3 square previews per day are for trying the model, not shipping with it.
Common mistakes
- Describing everything in the first prompt. The model averages a long wishlist into a muddy result. Lead with subject + style, then refine.
- Not setting the ratio until export. Generating 1:1 and then needing 16:9 wastes outputs against your quota. Set it in step 2.
- Re-prompting instead of editing. With
gpt-image-2you can edit in conversation; rewriting the whole prompt throws away the parts that already worked. - Judging a prompt on one bad roll. Re-run the same prompt two or three times before you decide the wording is wrong.
- Assuming Free-tier output is final. Free is capped at square 1024×1024 standard quality — fine for a preview, not for a 2K OG card.
Advanced tips
- Name a concrete style (“editorial illustration,” “Bauhaus poster,” “isometric 3D render”) rather than vague adjectives. The model anchors to named styles far better.
- Save a house-style prefix. Keep a reusable prompt prefix (palette, render style, lighting) and prepend it to every image for brand consistency.
- Use multi-image consistency for a set: ask for the same character or product across up to 8 outputs in a single prompt instead of fighting drift across separate generations.
- Change one line, regenerate. When a prompt is close but the lighting is off, edit only the lighting clause — do not rewrite the prompt.
Copy-ready prompt
A [subject] in [style] style, [composition] composition, [lighting],
[color-palette], for use as [purpose]. Aspect ratio [ratio], 2K resolution.
[Include headline text "[exact words]" / no text overlay]. No watermark.Swap the bracketed placeholders, drop the text clause if you do not want in-image text, and start in Instant mode.
FAQ
- Which model does ChatGPT use for images now?:
gpt-image-2, branded ChatGPT Images 2.0, launched April 21 2026. It replaced DALL-E 2 and DALL-E 3, both retired May 12 2026. - Can I make images with the free plan?: Yes, but only about 2-3 per 24-hour rolling window, at square 1024×1024 standard quality. For full 2K and all aspect ratios, you need Plus ($20/mo) or higher as of June 2026.
- Can
gpt-image-2put readable text in an image?: Yes — it renders text with roughly 99% character-level accuracy, including Chinese, Japanese, Korean, Arabic, and Hebrew. Quote the exact words in the prompt and use Thinking mode for the final pass. - Can I use these images commercially?: OpenAI permits commercial use on paid plans, but verify the current terms and avoid trademarks, real faces, and copyrighted characters. Liability for those sits with you.
- Why does my logo still look glitchy?: The model cannot reliably recreate a specific registered logo. Generate the background only, then place your real logo in an editor afterward.
- How do I edit one part of an image without starting over?: Describe the change in the same chat (“keep everything, replace the mug with a notebook”).
gpt-image-2inpaints that region instead of regenerating the whole frame.