OpenAI released ChatGPT Images 2.5 on September 8, presenting it as a step forward in detail, texture, lighting, and edit fidelity. The company said image-generation latency is down by as much as 50% from Images 2.0. Two new API models are now available as well: GPT-Image-2.5 Flare, positioned as the fast default, and GPT-Image-2.5 Sunburst, aimed at higher-end editing precision.
Decrypt framed the new review as a follow-up to its May comparison, when GPT Image 2 and Nano Banana 2 split results across eight categories. GPT Image 2 had won more of them, but it also showed a distinct oversharpening artifact whenever prompts piled on too many constraints. The question going into this round was simple: did OpenAI fix that weakness?
To test that, the outlet ran another head-to-head using the same evaluator’s eye, six categories carried over or adapted from earlier rounds, and the same prompts for both models. On Google’s side, the comparison used Nano Banana 2, which Decrypt identifies as Gemini 3.1 Flash Image, not the slower Nano Banana Pro. The piece describes this as a same-tier contest: OpenAI’s new fast-and-precise model against Google’s fast-and-precise model.
What changed in GPT-Image 2.5
According to Decrypt, OpenAI’s image models have tended to ship with a recognizable flaw. The original GPT Image model was known for a warm yellow cast that internet users nicknamed the “piss filter,” a tint OpenAI never fully explained and never fully removed. GPT Image 2 lost that problem but picked up another one: prompts with too many stacked constraints could push the model into crunchy, overprocessed oversharpening.
Neither issue appeared in this round, the review says. Every image generated with Images 2.5 maintained color balance and detail even at high prompt complexity, without the yellow cast or the brittle oversharpened texture seen in prior generations. Decrypt calls that a visible fix and says it shows up most clearly in the steampunk and portrait tests, which it describes as the most photographically coherent ChatGPT outputs it has seen across three generations.
OpenAI also added workflow features that do not show up in a side-by-side visual comparison but still matter in use. Those include Sketch for drawing a rough layout directly inside ChatGPT as a generation reference, prompt sharing, inline comments tied to specific image regions, and format templates for posters and merchandise. On the API side, quality tiers now extend from low to new high and max settings above the Images 2.0 ceiling.
Lettering density: Kellerman's Hardware
This category pushed text rendering hard. The prompt called for a gritty 2 a.m. intersection where nearly every surface contains readable wording: a ghost sign, graffiti, storefront lettering, a torn concert poster, a stenciled curb, and a payphone covered in stickers.
Decrypt says Nano Banana 2 rendered nearly everything cleanly. Its main mistake was a payphone sticker that duplicated its own text in a garbled way, a flaw the reviewer treated as minor and easy to miss.
ChatGPT Images 2.5 did add one detail that Nano Banana 2 skipped completely: a lamppost covered with overlapping stapled flyers, torn and weathered just as the prompt described. But OpenAI’s model introduced its own legibility issues. A street-art tag read like “STILLL HERE,” with an extra L, and the apostrophe in “KELLERMAN'S” on the ghost sign was missing or unreadable.
Decrypt’s take was that OpenAI produced the more realistic image, while Google paid closer attention to the text itself. The category went to Nano Banana 2.
Spatial awareness: the steampunk clock tower
This prompt called for a complex aerial composition with five planes of depth, a giant clock tower, legible Roman numerals on clock faces showing different times, and six additional text elements distributed from foreground to background.
Decrypt found ChatGPT Images 2.5 clearly stronger on atmosphere. Steam rising from rooftops was visible, a river cut through the middle ground, and tonal separation across the five depth planes looked richer. The lettering was also easier to read in this version.
Nano Banana 2 produced a flatter scene, though its two visible clock faces did show readable Roman numerals—XII, III, VI, and IX. The issue was that the hands sat in similar positions, which did not match the prompt’s request for different times.
Decrypt gave this round to GPT Images 2.5 because it followed the instructions more closely without giving up realism.
Illustration: the anime spirit medium
The illustration test asked for a Studio Ufotable-style key visual showing a girl at a torii gate, mid-transformation into spiritual energy, with a nine-tailed kitsune and a twilight sky painted in the style of Makoto Shinkai.
Decrypt says ChatGPT Images 2.5 delivered the strongest sky seen in the entire series. The image included a visible sun disc, water reflection, and a mountain silhouette, enough for the reviewer to say the Shinkai comparison felt earned. The asymmetric eyes also worked well. Where the model drifted was in the instruction that the girl should be dissolving into energy. Instead of a flowing translucent dissolve, it turned that idea into an electric-crackle effect running through her hair.
Nano Banana 2 came closer on that one line, using a wispy blue-white energy trail. Neither model, however, produced a convincing nine-tailed fox.
Decrypt still awarded the category to ChatGPT Images 2.5 on overall visual impact, despite the interpretive drift.
Realism: the rooftop architect
This cinematic portrait prompt stacked a long list of independent constraints: a beige trench coat, round glasses, blueprints specifically in the left hand, golden-hour lighting, shallow depth of field, and film grain.
Decrypt praised ChatGPT Images 2.5 for its light. The sun disc appeared directly behind the subject, and skin micro-texture looked strong. The skin remained too smooth, but the broader scene came across as highly realistic, like a frame shot on an analog camera.
The review also notes an interesting effect: adding terms that would normally imply lower-quality photography actually increased the sense of realism. After introducing keywords such as “realistic, highlights, crushed shadows, uneven flash, blown out skin tones, candid moment, shot on a phone camera,” the generated result looked more like an actual photograph.
Nano Banana 2 preserved a fuller composition, though it placed the blueprints in the right hand instead of the left. It also added a readable blueprint label—“PROJECT: 124 DUANE ST”—that many renders omit.
Decrypt’s conclusion here was split: Nano Banana 2 won on a single shot, while GPT performed better over repeated iterations.
Agentic research: the Bitcoin timeline
Because both platforms can research before rendering, Decrypt asked for a widescreen Bitcoin history timeline in a children’s drawing style and set a strict requirement for factual accuracy. The publication says this was the most important category in the comparison, and the place where the biggest gap appeared.
ChatGPT Images 2.5 produced a clean two-row infographic with specific dates throughout, but one date was wrong. It marked 2023 as the year Bitcoin ETFs were approved in the United States. The U.S. Securities and Exchange Commission approved the first spot Bitcoin ETFs on January 10, 2024, one year later. The review adds that Bitcoin futures ETFs were approved in 2023, so the error may reflect an interpretation issue.
Nano Banana 2’s version was less structured but covered similar events. Decrypt says that also suggests the model shows little variance on repeated prompts. Still, it grouped ETF approval and the fourth halving under a “2023–2024” bracket instead of stating a single incorrect year, so none of its claims were technically false.
Decrypt gave the category to Nano Banana 2, arguing that a confidently wrong date matters more in a test built specifically to measure whether agentic reasoning leads to accurate output.
Abstract concepts: a prompt built from invented words
The final category used a prompt made almost entirely of words with no established meaning: “A woman eating shmfiyxl in Lyxin. Next to her, her Lymglsushing plays Lakishkark.” The dish, the place, the companion, and the activity all had to be invented by the models.
ChatGPT Images 2.5 responded by turning the nonsense terms into visible labels. “Lyxin” appeared on a neon sign above a sleek futuristic restaurant. “Lakishkark” was printed on the box of a board game being played by the woman’s alien tablemate. In Decrypt’s reading, the words stopped being abstract once they became readable signage.
Nano Banana 2 went the other way. It created a warm, culturally specific setting with a Guatemalan market stall, a woman in a traditional huipil, and an orc-like figure playing a hybrid string-and-pipe instrument. It interpreted “plays” as playing music rather than playing a game, which the reviewer treated as equally valid. What was missing was any visible tie back to “Lyxin” or “Lakishkark.” Those invented terms never appeared in the frame as text.
Decrypt awarded the category to ChatGPT Images 2.5, saying literal visible labeling was the stronger answer when the prompt offered no other grounding.
The verdict
Nano Banana 2 won three of the six categories in this round, leaving the result effectively tied, according to Decrypt. The publication says the better choice will depend largely on what a user expects and how that user works with the model.
It also says ChatGPT Images 2.5 fixed the oversharpening problem that affected its predecessor, and that its illustration output in this test may be the most visually striking single image produced by either model across the two full rounds of comparison.
In the end, Decrypt argues that the difference is not broad image quality. It comes down to small, verifiable misses: a spelling slip in a text-heavy scene, a wrong year in a research-based infographic, and similar details. On overall aesthetics and finish, the two models were described as roughly on par.
