GPT Image 2 & Image Thinking — Image & Multimodal
GPT Image 2 & Image Thinking
Next-Gen Visual Generation (2026 Standard)
GPT Image 2 (gpt-image-2) replaced both retired DALL-E 3 and legacy gpt-image-1 (shut down Oct 2026; interim gpt-image-1.5 / gpt-image-1-mini sunset Dec 1, 2026) as OpenAI's unified visual generation and conversational editing model. It introduces token-based pricing, flexible aspect ratios, and pixel-accurate typography.
Key Capabilities
| Feature | DALL-E 3 (Retired) | GPT Image 2 (gpt-image-2) |
|---|---|---|
| Text in images | Often garbled | Pixel-perfect rendering |
| Aspect ratios | Fixed (1:1, 16:9) | Fully flexible |
| Editing | Inpainting only | Full conversational editing |
| Pricing | Per-image | Token-based (pay for complexity) |
const image = await openai.images.generate({
model: "gpt-image-2",
prompt: "A futuristic Tokyo skyline at sunset, cyberpunk style, 8K detail",
size: "1536x1024",
quality: "high"
});
GPT Image Thinking
A specialized capability that combines agentic layout reasoning with visual generation. It analyzes complex multi-element prompts, plans typography and composition, and refines visual outputs before returning the final image.
Vision Input (Multimodal)
OpenAI's active multimodal models (GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, and GPT-6 Luna, plus deprecated GPT-4o until Jan 14, 2027) natively accept image inputs - upload photos, screenshots, charts, or documents for visual reasoning.
const response = await openai.responses.create({
model: "gpt-6.1-terra",
input: [{
role: "user",
content: [
{ type: "input_text", text: "What's in this screenshot?" },
{ type: "input_image", image_url: "https://example.com/screenshot.png" }
]
}]
});
gpt-image-2 inside the Responses API (tools: [{ type: "image_generation" }]) for multi-turn design iteration in natural language.Synthesize visual assets using OpenAI's flagship image generation endpoint. Request a 1024x1024 high-fidelity image via /v1/images/generations with gpt-image-2.