OpenAI shipped ChatGPT Images 2.5 on September 8, positioning the update directly against Google’s Nano Banana 2 (Gemini 3.1 Flash Image) and describing it as delivering sharper details, richer textures and more natural lighting. Image-generation latency dropped by as much as 50% compared with Images 2.0, and the release introduces two new API models, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst. Flare is the fast default while Sunburst is built for premium editing precision.
Earlier comparisons showed GPT Image 2 exhibited oversharpening when prompts contained too many constraints, and GPT Image 1 carried a yellow color cast that was nicknamed the ‘piss filter’. In this round of testing, neither the yellow cast nor the oversharpening appeared in Images 2.5, and Images 2.5 maintained color balance while preserving detail. The report explicitly notes that those earlier artifacts were absent in Images 2.5 across the evaluated scenes. These observations are presented as direct comparisons to prior GPT Image model behavior rather than as new feature claims.
Steampunk and portrait tests showed Images 2.5 were more photographically coherent across three model generations. The comparison in this round was against Google’s Nano Banana 2 (Gemini 3.1 Flash Image), not Nano Banana Pro. In the complex intersection test described as a punishing 2 a.m. scene, Nano Banana 2 rendered most elements cleanly, and a payphone sticker that garbled its own text was noted as a minor flaw. The write-up reports those scene-specific outcomes alongside the generational coherence measurements for Images 2.5.
The section compiles image-quality test outcomes across multiple scenes and model generations. It also presents head-to-head results comparing Images 2.5 with a Google image model.
ChatGPT Images 2.5 introduces several user-facing features for image creation and collaboration. The release adds a Sketch tool, prompt sharing for reproducible prompts, and inline comments tied to regions within images to facilitate feedback. It also includes format templates aimed at output-ready designs for posters and merchandise, enabling standard sizes and layouts. These additions address editing and review workflows and are listed among the release features.
The API now includes updated quality tiers that add new high and max levels above Images 2.0. The report lists these “high” and “max” tiers as additions to the prior set of API quality options. The listing presents the new tiers without additional implementation details.
Nano Banana 2 was evaluated in a demanding 2 a.m. intersection test that contains multiple surfaces with readable text. In that test, Nano Banana 2 rendered most elements cleanly. The write-up notes a payphone sticker that garbles its own text as a minor flaw in the scene. The intersection test is described as punishing because nearly every surface carries readable text.
The section compiles these feature additions and the scene-level test observations for both models. It presents the features and test notes as separate items in the report.
ChatGPT Images 2.5 is presented as an iterative update that addresses earlier image artifacts while introducing user-facing tools such as Sketch, prompt sharing, inline comments on image regions, format templates for posters and merchandise, and expanded API quality tiers including new high and max levels. Comparative testing reported improved photographic coherence for Images 2.5 across steampunk and portrait scenes, while Google’s Nano Banana 2 rendered most elements cleanly in the demanding intersection test despite a minor payphone-sticker garble noted in that scene.
