General · EdgepediaSeptember 17, 20267 min read

GPT Image

GPT Image is a family of image generation and editing models developed by OpenAI, beginning with the native image generation capability built into GPT-4o and announced on March 25, 2025, and continuing through the API models gpt-image-1, gpt-image-1.5, gpt-image-2, and the gpt-image-2.5 pair (Sunburst and Flare). Unlike OpenAI's earlier DALL·E models, which operate as diffusion models, GPT Image generation is autoregressive (generates output piece by piece, each part predicting from previous parts) and embedded natively in the GPT-4o architecture, allowing it to use the model's text knowledge and chat context when producing or editing images.¹

Key fact	Detail
First release	4o image generation announced March 25, 2025, live in ChatGPT and Sora for $200/month Pro subscribers, with wider rollout planned²
Architecture	Autoregressive model natively embedded within GPT-4o, unlike the diffusion-based DALL·E¹
Instruction following	Handles up to 10-20 different objects per prompt, where other systems struggle with about 5-8³
API pricing (gpt-image-1)	$5 per 1M text input tokens, $10 per 1M image input tokens, $40 per 1M image output tokens; roughly $0.02, $0.07, and $0.19 per generated image at low, medium, and high quality⁴
GPT Image 1.5	Launched December 16, 2025; up to 4x faster than its predecessor, with image inputs and outputs 20% cheaper⁵
GPT Image 2.5	Two API models: Flare for fast general generation; Sunburst for editing-heavy workflows⁶
DALL·E retirement	OpenAI has retired the official DALL·E GPT from ChatGPT; image generation continues through ChatGPT Images⁷^•⁸

Origin and architecture

Before 2025, image generation in ChatGPT was handled by DALL·E 3, a separate diffusion model. OpenAI's stated position was that image generation should be a primary capability of its language models, and the March 2025 release moved generation inside GPT-4o itself.³ The company's system card addendum states that, unlike DALL·E, 4o image generation is an autoregressive model natively embedded within ChatGPT, and that this embedding lets it draw on everything the model knows, including chat context and uploaded images.¹

Independent analysis broadly supports the autoregressive characterization while adding detail. The GPT-ImgEval benchmark paper proposes, based on classification of the model's generated data, that GPT-4o's image generation consists of an autoregressive component combined with a diffusion-based head for image decoding, rather than VAR-like architectures.⁹ OpenAI trained the model on the joint distribution of online images and text, with aggressive post-training, and told The Wall Street Journal it used publicly available data plus proprietary data from partnerships such as Shutterstock.³^•²

Capabilities

Text and instruction following. OpenAI positioned the model around workhorse imagery: logos, diagrams, menus, and infographics where text must render accurately. The launch post states the model can handle up to 10-20 different objects in one prompt, with tighter binding of objects to their traits and relations than systems that struggle with about 5-8.³ Because generation is native to the chat model, images can be refined through multi-turn conversation, with character appearance kept coherent across iterations.³

Image-to-image editing. The model takes one or multiple images as input and produces related or modified images, including edits of images containing people and inpainting of foreground and background details.¹^•² An empirical study found GPT-4o strong at large-scale edits such as background changes while maintaining coherence, and strong at subject preservation, but noted it can fail to retain details outside the intended edit region and struggles with unusual geometries such as a "pentagonal pyramid".¹⁰ The GIE-Bench editing benchmark, which compared GPT-Image-1 against other state-of-the-art editing models across more than 1,000 examples in 20 categories, found it led in instruction-following accuracy but often over-modified irrelevant image regions.¹¹

Product and API history

4o image generation (March 2025). Announced March 25, 2025, the feature went live in ChatGPT and Sora for Pro subscribers first, with rollout to Plus, free, and API users planned. TechCrunch reported that GPT-4o with image output "thinks" longer than DALL·E 3, which it effectively replaced, to produce more accurate and detailed images.²

gpt-image-1 (API, 2025). The API model exposes the same capabilities with the same safety guardrails as 4o image generation in ChatGPT, including C2PA metadata in generated images and a moderation parameter set to auto or low. Pricing is per token, and OpenAI cited early integrators including Canva, GoDaddy, HubSpot, Instacart, and invideo.⁴

GPT Image 1.5 / ChatGPT Images (December 2025). On December 16, 2025, OpenAI launched a new Images experience in ChatGPT with a dedicated Images tab, preset filters, and the API model GPT-Image-1.5, up to four times faster than its predecessor and better at editing and instruction following; image inputs and outputs were 20% cheaper than GPT Image 1.⁵ OpenAI's release notes describe stronger preservation of lighting, composition, and people's appearance across edits, denser and smaller text rendering, and acknowledged remaining limitations in styles, multiple faces, and multilingual output.⁵

GPT Image 2 and 2.5 (2026). The changelog lists GPT Image 2 with flexible image sizes, high-fidelity image inputs, token-based pricing, and Batch API support with a 50% discount; Batch API support also covers gpt-image-1.5, chatgpt-image-latest, gpt-image-1, and gpt-image-1-mini.¹² ChatGPT Images 2.5, launched in September 2026, added a Sketch feature for drawing prompts directly in ChatGPT, templates for formats such as posters and merchandise, and prompt sharing; in the API it ships as two models, GPT-Image-2.5 Flare for fast general generation and GPT-Image-2.5 Sunburst for editing-heavy workflows.⁶ Axios's early testing found the update faster than its predecessor and better at preserving the likeness of people and pets, with editing tools that let users type and place text, sketch objects to add, and indicate changes visually rather than only in words.¹³

DALL·E retirement. OpenAI has retired the official DALL·E GPT from ChatGPT, directing users to ChatGPT Images; Tom's Guide reported the retirement date as August 30, 2026, with users advised to download images they wanted to keep.⁷^•⁸

Safety

The system card addendum identifies three capability-driven risks distinct from earlier models: image-to-image transformation, photorealism that can resemble photographs, and detailed instruction following including text and diagrams. Without controls, the model could create or alter photographs detrimental to depicted people or provide schematics for weapons, so the launch relied on chat-model refusals and the existing safety infrastructure built for DALL·E and Sora.¹ The later ChatGPT Images 2.0 system card notes that heightened realism could, absent safeguards, allow more convincing deepfakes of real people, places, or events, and describes a layered stack: text classifiers that block violative requests before generation, image classifiers that screen input images, and a final check of the generated image before display.¹⁴ Generated images carry C2PA metadata, and ChatGPT Images 2.5 continues to use C2PA metadata and invisible watermarking to identify images made with the tools.⁴^•⁶

Evaluation and limitations

Benchmarks document both the model's strengths and its characteristic failures. GPT-ImgEval found GPT-4o significantly surpassing existing methods in generation control, output quality, and knowledge-informed semantic synthesis, while also cataloging synthetic artifacts and examining detectability by image forensic models.⁹ The empirical study of GPT-4o image generation found weaknesses in culturally specific elements, such as rendering a western-style crown for a Chinese general, and in maintaining boundary continuity in large scenes.¹⁰ OpenAI's own release notes for GPT Image 1.5 state that results remain imperfect, with significant room for improvement in future iterations.⁵

References

Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Model families and named models › Image generation models

Initially written Sep 17, 2026 · Reviewed: Oct 15, 2026 · Edited: Oct 15, 2026; Oct 25, 2026 · Last review: Oct 15, 2026

Notice something wrong?