OpenAI released ChatGPT Images 2.5, a new image generation model rolling out today to all ChatGPT, ChatGPT Work, and Codex users on desktop, mobile, and web. The company also shipped two new API models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, aimed at developers building generation workflows into products.
The headline number is latency. OpenAI says image generation is up to 50 percent faster than Images 2.0. That figure is a ceiling, not a typical result, and OpenAI did not publish a distribution of generation times across prompt types. Flare, which the API uses by default, is pitched as producing better images than GPT-Image-2 while taking half as long, the same baseline the consumer claim leans on.
Every quality claim beyond that number (sharper detail, more natural lighting, better preservation of subjects in reference photos, more reliable editing across multiple turns) comes from OpenAI’s own description of its model. The announcement includes no independent benchmark results and no third-party evaluation. That is standard for a first-party launch post, but the claims should be read as a vendor’s account of its own product until outside testing catches up.
The more consequential change is reference-image handling. OpenAI says Images 2.5 keeps a subject recognizable when it is placed in new settings or styles, and that edits made across a multi-turn conversation are less likely to drift or degrade the image over successive passes. Image generation has largely stopped competing on the quality of a single output and started competing on control. A model that will not hold a reference image steady across five edits is unusable in a production pipeline no matter how good any one frame looks. If the reliability gain holds up, it changes the calculus for teams that abandoned reference-based workflows after earlier models let subjects drift after two or three edits. Existing single-shot prompts do not need to change. Multi-turn editing prompts are the ones worth re-testing first.
For builders, the API split matters more than the consumer features. Flare is positioned as the default choice for creator content, product experiences, visual search, and high-volume generation, at lower latency than GPT-Image-2. Sunburst trades that speed for tighter control on premium workflows such as campaign creative and takes longer to generate per image. OpenAI did not publish per-image pricing for either model in the announcement. Teams evaluating a switch will need to check the API pricing page directly before estimating cost at volume.
The consumer-facing additions are incremental rather than architectural. Sketch lets a user draw a rough reference directly in ChatGPT; prompt-format templates, inline comments on images, and the ability to share the prompt behind a generated image round out the release. They lower the friction of getting a first draft closer to what a user intended, but none of them bear on whether the underlying model holds up across a repeatable pipeline. OpenAI also says the release carries its existing safety checks, C2PA metadata, and invisible watermarking, with more detail in an accompanying system card.
Teams currently running reference-based editing pipelines on GPT-Image-2 or an earlier ChatGPT image version should re-test a multi-turn edit sequence against Flare before assuming the latency gain is free. A faster model that still drifts after three edits solves the wrong problem.
OpenAI announced ChatGPT Images 2.5 in a company blog post published September 9, 2026.
