Updated October 15, 2026

OpenAI announced 4o image generation on March 25, 2025. This article preserves that launch history and explains subsequent changes.

Editorial note: This update is based on official OpenAI documentation. We did not run image-generation or API tests. The featured image is the article’s original conceptual cover, not a product screenshot or a verified model output.

What launched in March 2025

The launch announcement emphasized conversational image creation, text within pictures, uploaded-image references, and iterative editing. Examples included menus, illustrated explanations, and game-character concepts. OpenAI described handling 10–20 objects; that was a company capability claim, not a guaranteed count for every prompt. Some demonstrations were selected from multiple attempts.

OpenAI announced a rollout starting with Free, Plus, Pro, and Team, with Enterprise and Edu to follow. That described the rollout plan, not immediate access for every account.

How the API model was named

On April 23, 2025, OpenAI released the image model in the API as gpt-image-1. It identified this as the model powering the then-current ChatGPT image experience. The initial release used the Images API; Responses API support was still forthcoming. Some developers needed organization verification. The March announcement’s promise of API access therefore became a dated release, rather than an indefinitely pending feature.

The GPT Image 1 model page describes text and image inputs with image outputs. A model’s ability to understand an uploaded picture should not be confused with its ability to generate one. Developers should use the documented image model and endpoint, rather than assume that a request naming GPT-4o provides equivalent image output.

The launch system-card addendum described 4o image generation as autoregressive. That description is sufficient here; this article does not infer an undocumented layer-by-layer architecture.

Availability as of October 2026

OpenAI’s API deprecation schedule, checked for this update, lists gpt-image-1 for shutdown on October 23, 2026. It is deprecated, with shutdown still scheduled in the future as of this article’s update date. The table recommends gpt-image-2.5-sunburst or gpt-image-2.5-flare as replacements. Its December 1, 2026 entries concern other image models, including gpt-image-1-mini and gpt-image-1.5; they should not be substituted for GPT Image 1’s deadline.

Separately, GPT-4o’s retirement from ChatGPT began on February 13, 2026, with a temporary Custom GPT exception for Business, Enterprise, and Edu ending April 3. OpenAI explicitly said the image-generation model was different from the retiring text model and that ChatGPT Images was unaffected by that retirement announcement. That historical statement should not be read as a promise that image products will never change.

OpenAI subsequently introduced ChatGPT Images 2.5 on September 8, 2026, alongside its Flare and Sunburst API models. Readers choosing a tool now should consult current product documentation rather than the plan names or access instructions in a 2025 launch report.

Capabilities and practical limits

In its launch limitations, OpenAI acknowledged cropping, hallucinations, object-binding errors, graphing difficulties, multilingual text problems, imprecise edits, and dense small-text limitations. These qualifications matter when interpreting its selected demonstrations.

The current image-generation guide still warns about text clarity and placement, consistency across generations, and precise composition. It says complex prompts can take up to two minutes. These are documented limitations of the current GPT Image guidance, not measurements we made of the March 2025 release.

For publishing work, inspect every word, number, relationship, and cropped edge. A visually convincing diagram still needs a factual review. A mockup should be checked at its intended display size, and a requested local edit should be compared with the whole original image for unintended changes.

Safety and provenance

The 2025 system card describes layered safeguards: chat-model refusals, prompt screening, and output screening with classifiers and a safety-focused reasoning monitor. It also describes C2PA metadata and internal tools for assessing whether an image came from OpenAI. These were mitigation and provenance measures, not guarantees of harmless output or factual authenticity. The internal tool described at launch was not presented as a public image-search service.

OpenAI’s current provenance guidance explains that Content Credentials can be lost through screenshots, conversions, or editing. Missing metadata does not prove human authorship; detected signals do not establish factual accuracy, ownership, or who created the content. Coverage varies with the model, product, export, and creation date. The page now discusses newer watermarking and verification tools, with exceptions for API models facing imminent deprecation, including gpt-image-1. Those current descriptions should not be retroactively applied to every 2025 file.

Keep the original exported file when possible, retain any available credentials, and label illustrative AI imagery clearly. Do not use a realistic-looking generated picture as evidence that a depicted event occurred.

A hypothetical poster workflow

The following is an original editorial example, not a tested prompt or an OpenAI demonstration. Suppose a community group wants a square poster for a fictional book swap:

Create an illustrated square poster with three books beside a blue mug. Use a cream background and a large heading reading “Book Swap.” Leave a clear blank area beneath the illustration for event details. Keep the design simple and readable at phone size.

  1. Review the first result for the requested objects, exact heading, and usable blank space.
  2. Request one change, such as moving the mug to the left. Compare the revised picture with the previous version before accepting it.
  3. Add the confirmed venue, date, and contact details in an editor with editable text. Proofread the exported poster at its actual size.
  4. Save the original and final files, and describe the illustration accurately when publishing it.

This workflow keeps factual copy reviewable and makes accidental changes easier to catch. It does not establish a success rate for any model.

Reading the launch in context

The useful takeaway is a workflow: describe a visual goal, inspect the result, and refine it with explicit checks. Historical capability claims and today’s availability need separate dates. For the preceding product generation, see our historical overview of DALL·E 3; it is background reading, not a recommendation to use a retired API.