A release aimed at the editing round

OpenAI released Images 2.6 on November 4, adding GIF creation and more control over image revisions. The company says it is available globally in ChatGPT and Codex across subscription tiers, with Flare and Sunburst variants in the API.

The practical focus is iteration: change one part of an image while keeping other details consistent. OpenAI also announces image markup and SVG exports in the API. It reports faster generation, but those speed and consistency claims have not been independently tested for this article.

Detail and complexity are different controls

The announcement distinguishes image quality from complexity. Quality governs visual detail and fidelity; complexity changes how much planning goes into satisfying a request. A small local edit and a scene with many constraints therefore need not use the same approach.

For readers assessing a workflow, the useful comparison is the finished result, including the corrections it needed. A quick generation is not necessarily a quick completed job. The release does not establish how consistently every unchanged detail will survive repeated revisions.

What the safety card does—and does not—measure

OpenAI’s system card evaluates saved generations from 3,077 deliberately challenging requests, comparing each variant with its corresponding earlier baseline. It separates outputs graded unsafe from those blocked by the evaluated safeguards. That is an important distinction: generating a problematic image and delivering it to a user are different events.

The card says these tests are not representative of ordinary production traffic. It also identifies risks that emerge across animated frames and acknowledges that motion-related violations remain difficult to detect. The results should not be read as a universal safety rate.

OpenAI describes C2PA metadata and SynthID watermarking as complementary provenance signals. Those disclosures concern where an image came from; they are not evidence that the scene it depicts actually happened. Human review of the result remains a separate editorial or production decision.

Sources

  1. OpenAI: Introducing Images 2.6November 4 launch announcement, product availability and editing controls. Speed and consistency are developer claims, not independently measured results.
  2. OpenAI: ChatGPT Images 2.6 System CardNovember 4 first-party evaluation account: 3,077 adversarial requests, generation versus moderation distinctions, GIF limitations and provenance tools. Not production-traffic failure rates.