OpenAI has released a new generation of its image generator, ChatGPT Images 2.5. The company states that the GPT-Image-2.5 Flare model creates images 50% faster than the previous GPT-Image-2, with more natural lighting, richer textures, and better preservation of object features from reference photos. Two models have been released in the API — the fast GPT-Image-2.5 Flare and the editing-precision-focused GPT-Image-2.5 Sunburst — with identical pricing, while ChatGPT itself has gained Sketch, direct comments on images, templates, and generation with a transparent background.

image
image

What happened

OpenAI has moved its image generation lineup to a new generation. Two model identifiers are established in the API: gpt-image-2.5-flare for mass generation with low latency and gpt-image-2.5-sunburst for scenarios where maximum editing precision is important; both models are fixed as a 2026-09-08 snapshot. Their prices are identical: $5 per 1 million input text tokens, $8 for input images, and $30 per 1 million output tokens, with cached input costing $1.25 for text and $2 for an image. Developers connect both models via the Images API and Responses API.

Context

The release continues the GPT-Image line, and OpenAI names GPT-Image-2 as the comparison point for the new generation. The lineup structure looks like a product solution: two models with identical pricing but different profiles — speed versus editing precision — serve to segment tasks, rather than indicating different scientific content of the architectures. The pricing grid is built on tokens, and input caching reduces the cost of repeated generations with the same references. The main focus in describing improvements is moving image work from one-time generation to iterative editing in a dialogue.

Why this matters for the industry

The main statement for production pipelines is the editing mode: the model changes an individual scene element without affecting the rest, and, according to OpenAI, stops losing structure during a series of sequential edits in one dialogue. Marketing, retail, and media teams can now edit a product, background, or text without regenerating the entire asset, which directly reduces the cost of iterative edits in content pipelines. The identical pricing of both models simplifies budgeting: choosing between speed (Flare) and precision (Sunburst) does not require recalculating costs, and input caching makes repeated generations with the same references significantly cheaper. The likely automation pattern is routing: fast Flare for mass tasks, precise Sunburst for final edits.

Why this matters for users

The new tools are available to all ChatGPT, ChatGPT Work, and Codex users on the web, mobile, and desktop, with no separate connection required. You can write @Sketch, draw a sketch by hand, and get an image based on it; leave a comment directly on a specific area of the image so the model corrects only that area; generate an image immediately with a transparent background; use templates for posters, merch, flyers, and product photos. For pinpoint edits, it is enough to mark the desired fragment of the image, rather than describing the desired result from scratch.

What is still unknown / limitations

All key release characteristics are currently confirmed only by OpenAI's words. The stated 50% speedup of Flare relative to GPT-Image-2 is accompanied only by the formulation: the announcement materials contain no latency metrics, time per image, load conditions, or hardware base for measurement, so this figure cannot be reproduced or verified based on release data. There is no technical report, article, or description of the architecture and training methods in the release sources, so improvements in lighting, textures, and preservation of features from references remain marketing characteristics. Until the speedup is confirmed by independent measurements, it is premature to build economic conclusions on it, and the reliability of pinpoint editing in long series of edits should be verified with your own tests.

Sources

Author

Look at AI, editorial team