ChatGPT Images 2.5

In ChatGPT, Images 2.5 adds sharper-detail and more precise-editing claims plus Sketch, Templates, image comments, and prompt sharing. In the API, developers choose Flare or Sunburst directly through the Images API or as the model behind the Responses API image-generation tool. `ChatGPT Images 2.5` is not itself the API request ID.

• Flare: `gpt-image-2.5-flare` or dated snapshot `gpt-image-2.5-flare-2026-09-08`.
• Sunburst: `gpt-image-2.5-sunburst` or dated snapshot `gpt-image-2.5-sunburst-2026-09-08`.
• Both accept text and image input, produce image output, and support generation and edit endpoints.

OpenAI positions Flare as the default for most applications and its fastest high-quality everyday image model. Sunburst is the more capable choice for detailed generation and edits where precision matters most, with longer generation times. Both support low, medium, high, xhigh, max, and auto quality settings. Treat this as provider positioning until a matched independent workflow shows which model is better for your exact task.

OpenAI says Images 2.5 better preserves subjects from reference photos, makes more targeted edits, maintains consistency across multiple turns, handles complex layouts and transparent backgrounds, and reduces latency by up to 50% versus Images 2.0. Early community discussion is mixed and focuses on rollout visibility, editing drift, text, artifacts, and whether an account is actually using the new model. Use the same source assets, prompts, settings, and acceptance criteria for your own comparison.

OpenAI currently lists both 2.5 API models at $5 per million text-input tokens, $1.25 per million cached text-input tokens, $8 per million image-input tokens, $2 per million cached image-input tokens, and $30 per million image-output tokens. The model pages say those token rates match GPT Image 2, while the general pricing table separately lists lower GPT Image 2 rows. Because those first-party surfaces conflict, do not claim a reliable percentage price change; verify the live pricing and billing console before launch.

OpenAI applies prompt, input-image, and output checks and publishes fixed adversarial safety evaluations with stated limitations. Images 2.5 outputs use C2PA metadata and an invisible SynthID watermark across ChatGPT, Codex, and the API. Neither mechanism guarantees that every copied or transformed image will retain complete provenance.

OpenAI says Images 2.5 is rolling out across all tiers on desktop, mobile, and web. Release notes add that Templates are not yet available in Work mode, Sketch and image comments are mobile flows, and existing image-generation limits are unchanged. A launch-day rollout can still reach accounts at different times.

This GetLLMs example was generated and edited in Codex. The requested change was to turn the bottle cap from blue to terracotta orange while retaining the bottle, label, fruit, linen, framing, and lighting. The label remains readable and the composition stays similar, but small texture differences remain. Open both images at full size to judge the edit. The available generation receipt does not verify a Flare or Sunburst selection, so this illustrates an editing workflow rather than a matched comparison of the two API models.

Before: blue cap

After: orange cap

Silent before-and-after sequence: the blue-cap image appears first, followed by the orange-cap edit, with visible labels. · 8 seconds · No audio

• Before and after use the same 1254 × 1254 frame. Compare the cap, label lettering, bottle edges, fruit texture, and linen folds.
• The silent eight-second video shows the original for four seconds and the edit for four seconds. It is assembled from these two still images; it demonstrates neither native video generation nor generation speed.
• For a separately attributed Sunburst example, the Simon Willison source below includes his model command, reference chart, and raccoon edit. The OpenAI launch video is also linked below.