Overview
GPT-Image-2 is the next-generation image generation model launched by OpenAI on April 21, 2026, with the external brand name ChatGPT Images 2.0 and the API model name gpt-image-2. It adopts a brand-new architecture (non-diffusion), replacing DALL·E 3 as ChatGPT's default image model upon release, and tops the LM Arena leaderboard with an ELO of 1512, about 240 points ahead of Nano Banana 2.
It solves two long-standing problems of image models. The first is text rendering: Chinese typesetting accuracy reaches 99%, so text-heavy scenarios like menus, posters and magazines no longer produce garbled characters, a notorious weakness of every prior image model. The second is comprehension: before generating, it searches the internet, analyzes uploaded files and reasons about the visual structure, executing complex prompts reliably instead of mechanically compositing. It supports up to 4096×4096 resolution and 8 variants per generation, and launched simultaneously on ChatGPT, Codex and the API.
On September 8, 2026 it was upgraded to ChatGPT Images 2.5, available to all paid ChatGPT tiers, ChatGPT Work and Codex on desktop, mobile and web. Compared with 2.0, generation latency is cut by up to 50% with sharper details, more natural lighting and richer textures; precise editing modifies only the specified region with stronger multi-turn consistency; identity preservation for people and objects in reference photos is noticeably better, and complex layouts including transparent backgrounds are supported. The API adds two models: GPT-Image-2.5 Flare for fast, cost-efficient generation and GPT-Image-2.5 Sunburst for premium visual workflows. ChatGPT Images and the GPT-Image API together generate over 3 billion images weekly.
Key Features
- 99% Text Rendering Accuracy: Near-perfect Chinese and English typesetting, supporting text-intensive scenarios such as menus, posters, and magazines.
- 4K Resolution: Up to 4096×4096 ultra-high-definition output, twice as fast as the previous generation.
- Reasoning Capability: Searches the internet, analyzes uploaded files, and reasons about image structure before generation.
- 8 Images at Once: Generates multiple variants in a single run for rapid iteration.
- UI Screenshot Generation: Capable of creating high-fidelity interface designs and web page screenshots.
- Precise Local Editing: Modifies specified areas while keeping the rest unchanged.
Use Cases
- Social media posters, Xiaohongshu covers, WeChat official account headers.
- Key visuals for product launches, event materials.
- E-commerce product images, menus, magazine pages.
- UI design prototypes, web concept drafts.
- Marketing materials with text (Chinese text no longer fails).
Pros
- Topped LM Arena, strongest overall capability.
- 99% accuracy in Chinese text rendering, a first in the industry.
- 4K resolution with doubled speed.
- Introduces reasoning capability, understands complex prompts.
- Simultaneously launched on ChatGPT, Codex, and API.
Pricing
Available with ChatGPT Plus ($20/month). API pricing: $8-$30 per million tokens, equivalent to $0.006-$0.211 per image (depending on resolution and quality). Free-tier users have limited trial usage.
Summary
GPT-Image-2 is the model to try first for Chinese content creation in 2026. It turned text rendering from the industry's biggest weakness into its strongest advantage, and the 2.5 upgrade further improves speed, precise editing and reference-image consistency, making it reliable for high-frequency scenarios such as WeChat official accounts, Xiaohongshu and e-commerce materials. For artistic and stylized work, pairing it with models like Midjourney v8 or FLUX 2 works well; for text-dense output and iterative refinement, GPT-Image-2 does the heavy lifting.
Version History
- ChatGPT Images 2.5 released (2026-09-08): OpenAI released ChatGPT Images 2.5 (official date Sep 8 US Pacific), available to all paid ChatGPT tiers, ChatGPT Work and Codex on desktop, mobile and web. Compared with 2.0: generation latency cut by up to 50%, with sharper details, more natural lighting and richer textures; precise editing that modifies only the specified region while keeping stronger multi-turn edit consistency; better identity preservation when using reference photos, plus complex layouts including transparent backgrounds. The Images API adds two models: GPT-Image-2.5 Flare (default recommendation, fast and cost-efficient) and GPT-Image-2.5 Sunburst (for premium visual workflows). ChatGPT Images and the GPT-Image API together generate over 3 billion images weekly.
- GPT-Image-2 released (2026-04-21): OpenAI released GPT-Image-2 (consumer brand ChatGPT Images 2.0), replacing DALL·E 3 as ChatGPT's default image model with the API open from day one across ChatGPT, Codex and the API. Built on a new non-diffusion architecture, it topped LM Arena with an ELO of 1512, achieved 99% accuracy in Chinese typesetting, and supported 4096×4096 output with up to 8 variants per run. Before generating, it searches the web, analyzes uploaded files and reasons about image structure, shifting from picture generation toward a context-aware design assistant.