GPT Image 2 — the text-legible production image model
OpenAI · Apr 22, 2026 · Proprietary (API only) · site
OpenAI's state-of-the-art image model — sharper text rendering, 2K output, and thinking-mode reasoning for layout-heavy work.
What it is
GPT Image 2 is OpenAI's flagship image generation and editing model, released April 22, 2026 under the `gpt-image-2` alias (snapshot `gpt-image-2-2026-04-21`). It targets production workflows where an image has to be readable, on-brand, and correctly laid out — not just pretty. At launch it swept the Image Arena leaderboards, posting a +242 Elo lead on text-to-image over the next model and topping both single- and multi-image edit tracks. Available through `v1/images/generations` and `v1/images/edits`; pairs with a reasoning model for thinking-mode generation.
What's new in GPT Image 2
Text rendering that ships
Multi-word labels, signs, UI copy, and non-Latin scripts render correctly at the sizes you actually need — the biggest jump of any OpenAI image model to date.
Layout-aware structured imagery
Diagrams, infographics, posters, comics, and multi-panel scenes hold their structure instead of turning into abstract shapes.
2K output and flexible ratios
More export ratios and higher-resolution output up to 2048px — suitable for ads, product flows, presentations, and documentation without an upscaler stage.
Thinking-mode generation
Paired with a reasoning model, the image pipeline can research the web, transform source material, produce multiple candidates, and self-check outputs before returning.
Tighter instruction following
Better at preserving requested details, relating objects accurately, and rendering dense compositions — fewer "almost right" passes.
Arena #1 across all tracks
At launch: 1512 on text-to-image (+242 Elo over the runner-up), 1513 on single-image edit, 1464 on multi-image edit — all three Image Arena leaderboards in one sweep.
Generated with GPT Image 2
What people build with GPT Image 2
Marketing creatives
Social graphics, ad banners, and email headers where the copy has to render exactly as written.
Diagrams & infographics
Charts, explainer graphics, and process diagrams with real labels — not "almost readable" text.
UI mockups & product shots
Realistic screenshot-style mockups for decks, docs, and early-stage product concepts.
Posters & print
2K output at flexible ratios — hero frames, event posters, and print-ready pieces from a prompt.
Strengths & limitations
- +Best-in-class text rendering — multi-word labels, signs, UI, and non-Latin scripts render cleanly
- +Strong layout and typography for diagrams, infographics, posters, comics, and multi-panel scenes
- +Flexible aspect ratios with high-resolution output up to 2K
- +Thinking-mode generation: pairs with a reasoning model to research, transform source material, and self-check outputs
- +Topped every Image Arena leaderboard at release with a record +242 Elo lead on text-to-image
- −Closed-weights and API-only — no self-hosting, fine-tuning, or LoRAs
- −Priced for production, not iteration: $30 / 1M output image tokens
- −Tier-gated rate limits (5 IPM on Tier 1, up to 250 IPM on Tier 5) — bulk workloads need an upgraded tier
How GPT Image 2 stacks up
| Feature | GPT Image 2 | GPT Image 1 | FLUX.2 Dev | Seedream 4.0 |
|---|---|---|---|---|
| Open weights | No | No | Yes (NC) | No |
| Max resolution | 2048 | 1024 | 4MP | 4K |
| Text rendering | Excellent | Good | Good | Good |
| Thinking mode | Yes | No | No | No |
| Cost / image | $$$$ | $$$ | $$ | $$ |
About GPT Image 2
Need stock imagery on a brief?
OKSLOP handles routing, provider selection, and delivery — you describe what you need and we hand back the shots.