Did OpenAI change GPT Image 2.5 since launch? Changelog check, Oct 3

No: as of 2026-10-03 OpenAI's changelog lists no change to the GPT Image 2.5 models after their September 8 release, and its October section has no image entries. The one image-related fix that month, on September 25, concerns image inputs to GPT-6 Sol and GPT-6 Luna, which read images and write text. It does not touch image generation.

This post reads OpenAI's API changelog and image generation guide, both read on 2026-10-03, and compares them with Sume's Image API docs.

What does the changelog say, entry by entry?

The table is a faithful reading of what the page lists and nothing more. If you saw a claim of a new GPT Image revision this week, it is not on this page.

What was in the September 8 launch?

The guide names two models: Sunburst, positioned for editing precision workflows, and Flare, positioned for fast, high-quality everyday generation. Both support xhigh and max quality in addition to low, medium, high and auto; earlier models stop at high. Recommended sizes are 1024x1024, 1536x1024 and 1024x1536, and custom sizes need both edges as multiples of 16, a maximum edge of 3840, a ratio from 1:3 to 3:1 and 655,360 to 8,294,400 pixels. Anything above 2560x1440 is experimental.

Outputs default to PNG, with JPEG and WebP available and a 0 to 100 compression value for those two. Transparent backgrounds need background: transparent with PNG or WebP output.

How does that map to Sume?

Sume lists ChatGPT Image 2.5 as openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst. Both take text-to-image, up to 16 references, an optional mask_url, and background of auto, transparent or opaque, with quality values auto through max. The Sume docs say an omitted quality defaults to high, which differs from a default of auto in OpenAI's guide, so set quality explicitly if you compare outputs across the two. The same docs say auto quality reserves the cost of max.

ChatGPT Image 2 remains selectable on Sume, and the docs say Auto model routing continues to use Flare.

What does the September 25 fix mean for you?

Only if you send images to GPT-6 for checking. A common pattern is to generate an image, then ask a vision model whether the headline spelling or the product placement is right. If your baseline for that check was recorded before September 25, OpenAI's note says image understanding was degraded in both GPT-6 models, so scores may have shifted. Rerun your evaluation set before comparing prompts or quality levels.

Image generation through POST /v1/images is a separate path. The changelog does not say the fix changed generated output, and neither do we.

What should you watch next?

Two places carry the signal: the changelog page above, and GET /v1/images/models on Sume, which lists the models and the parameters each accepts. A new revision would show up as a new model id there, and a request that sets a parameter the row does not list is rejected with 400 unsupported_parameter. If your integration pins a model id, review it when OpenAI posts a shutdown or a snapshot, not on a schedule.

A cheap safeguard is a golden set: ten prompts, a fixed quality and size, and the model id written down next to the outputs you approved. When a vendor ships something, rerun the set and compare. That tells you whether a change matters to your work without relying on anyone's marketing, and it gives you evidence if a cost or quality shift appears on an invoice later. Keep the set small so you actually run it, and include at least one prompt with quoted text and one edit with a reference image, since those are the cases where image models differ most.