OpenAI (DALL-E, GPT-Image, and GPT-Image-2)

OpenAI's image work is unusual among the labs in this document because it has never been led by a single dedicated image research team. From the beginning, image generation at OpenAI has been done by researchers who rotated between text, image, and multimodalMultimodalHandling more than one kind of data, for example text and images together, in a single model. work, treating images as one output modality of a broader research program. Aditya Ramesh led the original DALL-E and DALL-E 2 projects. Gabriel Goh led significant parts of the CLIPCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. work that underpinned those models. More recently, the image work has been folded into the broader GPT-4o and GPT-5 multimodal teams under researchers like Greg Brockman, Mira Murati (before her departure in late 2024), and various project leads whose names have not been publicly associated with specific releases. The result is that unlike Black Forest Labs (three founders with clear public identities), Midjourney (David Holz), or Stability (the chaotic Emad Mostaque era), OpenAI's image work has no defining public personalities. It is institutional work.

This matters strategically because OpenAI's image generation has always been positioned as a feature of ChatGPT rather than as a standalone product. You do not go to a dedicated DALL-E website the way you go to midjourney.com or flux1.ai. You go to ChatGPT and ask it to generate an image. The image model is a subsystem of the larger assistant. This is a deliberate choice that has shaped every decision OpenAI has made about image generation, from the pricing (included in ChatGPT subscriptions rather than sold per-image) to the distribution (through partners like Microsoft Copilot rather than through creative-tool integrations) to the product roadmap (tightly coupled to the GPT base model release schedule).

OpenAI's image architecture has gone through three distinct generations. The original DALL-E (January 2021) was autoregressiveAutoregressiveGenerating output one piece at a time, each conditioned on what came before, the way language models write text. Some image and video models work this way instead of by diffusion., meaning it generated images patch by patch the way GPT generates text tokenTokenA single unit of input, a word fragment for text or a patch for images, that a model processes. by token, built on a GPT-3-adapted architecture. DALL-E 2 (April 2022) abandoned autoregression and switched to a diffusion-based architecture built around CLIPCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. embeddings, which was dramatically higher quality. DALL-E 3 (October 2023) kept the diffusion architecture but added a critical new component: instead of using the user's raw promptPromptThe text description you provide to a model to specify what you want it to generate., the system rewrites the prompt using GPT behind the scenes to produce a much more detailed description, which is then fed to the diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image.. This 'prompt expansion' is why DALL-E 3 felt so much smarter than DALL-E 2 even though the underlying diffusion model was only modestly improved.

The current generation, sometimes called GPT-Image or GPT-4o image generation, takes this idea further. Image generation is now a native output mode of the GPT-4o and GPT-5 language models, which means the language model is not just rewriting the prompt, it is actually producing the image tokens directly as part of its own inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training.. This is the 'language-model-firstModel-firstA platform strategy built around training and improving one's own flagship generation model, exemplified by Runway and Google's Flow.' approach that Google later adopted with Nano Banana. The architectural details have not been fully disclosed, but the effect is that image generation inherits all of GPT's world knowledge, reasoning, and instruction following. When you ask GPT-Image to generate a historically accurate map of 1930s Europe, it knows the borders. When you ask it to render an infographic with correct statistics, it looks them up. The image model is whatever subset of the language model is responsible for visual tokens, and it is strong exactly where language models are strong.

DALL-E 1 (January 2021)

A 12-billion-parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. autoregressiveAutoregressiveGenerating output one piece at a time, each conditioned on what came before, the way language models write text. Some image and video models work this way instead of by diffusion. model trained on 250 million image-caption pairs. Never publicly released. Demonstrated that AI could generate recognizable images from arbitrary text prompts at quality that felt convincing for the first time. The technical approach (autoregression over image tokens) was eventually abandoned because diffusion turned out to be a much better fit for the image generation problem.

DALL-E 2 (April 2022)

A diffusion-based model built around CLIPCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. embeddings. Made available through a closed beta with a waitlist. The model that put AI image generation on the cultural map, DALL-E 2 outputs went viral, 'AI art' entered mainstream conversation, and most people's first exposure to what generative imagery could do came from DALL-E 2 examples shared on social media. Quality was dramatically higher than DALL-E 1, though it still failed on text, hands, complex compositions, and most of the same things later models would also fail on.

DALL-E 3 (October 2023)

Integrated directly into ChatGPT. The major architectural innovation was automatic promptPromptThe text description you provide to a model to specify what you want it to generate. rewriting using GPT-4 behind the scenes, which meant users could write short casual prompts and still get high-quality results. This also made DALL-E 3 much more accessible to casual users who did not know prompt engineeringPrompt engineeringThe craft of writing and structuring prompts to reliably get the result you want. tricks. The image quality was meaningfully better than DALL-E 2, though by this point Stable Diffusion XL and Midjourney V5 had caught up or surpassed it on pure visual quality.

GPT-Image-1 (March 2025)

The first model where image generation was a native output of the language model rather than a separate diffusion system. Launched as a feature of GPT-4o in late March 2025. This was the version that produced the 'Studio Ghibli moment' when the internet briefly flooded with Ghibli-style transformations of personal photos. GPT-Image-1 was unusually good at instruction following and consistent identity across multi-image generations, but it had a characteristic warm yellow color cast that became a recognizable signature of the model. It also had limitations around dense typography and complex spatial composition that the next generation aimed to fix.

GPT-Image-1.5 (late 2025)

An incremental update to GPT-Image-1 with improved quality, reduced yellow-cast bias, and better prompt following. Not a major architectural change. This was the version accessible through ChatGPT and the OpenAI APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. in early 2026, before GPT Image 2 replaced it.

GPT-Image-2 (launched April 21, 2026)

On April 4, 2026, three anonymous image models appeared on LM ArenaLM ArenaA public leaderboard where anonymous models compete in blind head-to-head human votes, increasingly the venue where labs quietly test new models before official release. under codenames maskingtape-alphaAlpha (LoRA alpha)A scaling factor that controls how strongly a LoRA's learned change is applied, effectively its volume knob, tuned separately from rank., gaffertape-alpha, and packingtape-alpha. They were identified by community testers as OpenAI internal variants within hours. On April 21, 2026, OpenAI officially launched GPT-Image-2 (gpt-image-2), confirming the leak was real. What made GPT-Image-2 matter was not the margin but the method. Unlike the diffusion predecessors DALL-E 2 and DALL-E 3, it is autoregressive, generating images tokenTokenA single unit of input, a word fragment for text or a patch for images, that a model processes. by token the way GPT generates text, and it reasons about an image's structure before rendering it. That architectural shift is why its text and instruction-following jumped so far ahead that its arenaArenaA public leaderboard that ranks models from large numbers of blind pairwise human preference votes, which tends to reward average appeal over fitness for a specific job. lead was the widest the board had ever recorded, a symptom of the change rather than the point of it. It features built-in reasoning before generating, meaning the model researches, plans, and self-corrects before producing the image. Text rendering accuracy is roughly 99 percent, native resolutionResolutionThe pixel dimensions of an image or video, for example 1024x1024. Higher resolution shows more detail but costs more memory and time. is 2K, and multilingual text rendering including non-Latin scripts works reliably. Pricing is token-based at $8 per million input tokens and $30 per million output tokens. DALL-E 2 and DALL-E 3 are both being retired on May 12, 2026, marking the formal end of the diffusion-based DALL-E lineage at OpenAI.

GPT-Image's strengths are instruction following, world knowledge, and the kinds of tasks that require the model to actually understand what it is rendering. If you need an image that depicts a specific historical event accurately, an infographic with correct statistics, a scene that follows complex multi-part instructions, or a visual that requires the model to reason about the objects in it, GPT-Image-2 is now the best choice by a wide margin, confirmed by its record-setting +242 point ArenaArenaA public leaderboard that ranks models from large numbers of blind pairwise human preference votes, which tends to reward average appeal over fitness for a specific job. lead. Text rendering accuracy of roughly 99 percent makes GPT-Image-2 the new leader on typography, surpassing Ideogram 3.0 and Recraft V3. Character identity across multiple generations is also unusually good because the language model can reason about consistency in a way diffusion models cannot natively.

GPT-Image's weaknesses are flexibility and ownership. You cannot download the weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable., you cannot fine-tune on your own data, you cannot train a LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models., you cannot use it offline, and you are subject to OpenAI's content moderation policies which are more restrictive than most competitors. For casual consumer use inside ChatGPT these are not big problems. For professional creative work where you need brand consistency, character persistence, or specific aesthetic control, the lack of customization is a real limitation and most serious operators use FLUX.2 or a fine-tuned Stable Diffusion variant for those use cases instead. The model also has the characteristic 'language-model-firstModel-firstA platform strategy built around training and improving one's own flagship generation model, exemplified by Runway and Google's Flow.' trade-off: it is much smarter about the world than a pure diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image. but it is also much more tightly coupled to OpenAI's base model release schedule, which means you cannot use it independently of whatever ChatGPT version is current.

OpenAI's strategic position in image generation is that it does not really care about winning image generation specifically. It cares about winning the broader assistant category, and image generation is a feature of that assistant. This is why OpenAI has never built standalone creative tools, has never offered fine-tuningFine-tuningContinuing to train an existing model on new data to specialize it for a style, subject, or domain, without training from scratch. for image models, has never released open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models., and has not courted the creative professional market the way Adobe, Midjourney, or Black Forest Labs have. When OpenAI does push forward on image generation, it is usually because the capability makes ChatGPT more useful as a general assistant, not because OpenAI wants to dominate image generation as a category.

The Sora shutdown announced in March 2026 (covered in the OpenAI Sora chapter) is the clearest expression of this positioning. OpenAI looked at the economics of running a standalone video generation product in a crowded competitive market and decided the compute was better spent elsewhere. The same logic could eventually apply to standalone image generation, though for now the image work remains cheap enough and strategically integrated enough with ChatGPT that no retreat is expected. The practical implication for operators is that OpenAI is probably not the long-term partner of choice for a generative imagery business, and that the value OpenAI brings is 'good enough image generation inside the most popular AI assistant,' not 'best-in-class creative tooling for professional use.'

OpenAI's image story has a clean before and after. Before was DALL-E, a standalone diffusion modelDiffusion modelA generative model that learns to produce images by training a neural network to reverse a noise-adding process. Generates by starting from random noise and iteratively denoising to produce a clean image. that DALL-E 3 wired into ChatGPT as a tool the assistant called. After was March 25, 2026's predecessor moment: on March 25, 2025 OpenAI folded image generation directly into GPT-4o as a native capability, so the same model that reasons over text and pictures now draws them, using the conversation, uploaded references, and world knowledge as context. That change is why the modern models follow instructions so well and render legible in-image text, and it is the seed of everything since, exposed to developers as gpt-image-1, refreshed through 1.5, and, in April 2026, gpt-image-2, an agentic model that plans an image's structure before generating and, as of mid-2026, tops the Artificial Analysis text-to-image arena.

You already have it if you have ChatGPT, where image generation is built into the GPT-4o-era and later models across the free and paid tiers. For developers the APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. exposes gpt-image-1, a cheaper gpt-image-1-mini for volume, gpt-image-1.5, and gpt-image-2, all billed on tokens: the original tier worked out to roughly two cents an image at low quality up to about nineteen cents at high. There is a moderation parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. you can set to low to cut false refusals where policy allows, and every image carries C2PAC2PA (Content Credentials)An open provenance standard, backed by Adobe, OpenAI, and Microsoft, that attaches cryptographically signed metadata proving a file's origin, though the metadata strips easily when a file is re-encoded. provenanceProvenanceA verifiable record of how a piece of media was made and by whom, increasingly required so that synthetic content can be traced. metadata. Because the model inherits the assistant's comprehension, it responds to plain, detailed natural language and conversational refinement rather than keyword soup.

It is best at exactly what diffusion models were worst at: legible in-image text, faithful edits to an uploaded photo, infographic and layout work, and complex compositional prompts. It is weakest on speed, often tens of seconds per image, on a warm yellow color cast that many people noticed on gpt-image-1 and promptPromptThe text description you provide to a model to specify what you want it to generate. against, on dense crowds of faces and some non-Latin scripts, and on moderation that can block benign prompts, especially around real people. If you need fast generation at scale, a cheaper diffusion API will beat it; if you need the image to actually say the right words and follow a precise instruction, this is the one to reach for.