GPT Image 2.5 handled our strict infographic brief better in the matched second pair. Nano Banana 2 produced readable posters and useful label-color edits. The better choice depends on the asset you need.
- Infographics: GPT preserved the five-step sequence; Nano added unwanted content in the second pair.
- Local edits: both changed the label color while retaining its wording in the matched first pair.
- Product images: convincing materials from both, with camera-angle and lighting trade-offs.
Explore image creation and focused editing through GlobalGPT, or compare the original outputs below against your own brief.
GPT Image 2.5 vs Nano Banana 2 at a Glance
- GPT Image 2.5: Documentos OpenAI Sunburst for image capability and editing precision, and Sinal de fogo for fast everyday generation.
- Nano Banana 2: Google Imagem em Flash do Gemini 3.1, with generation, editing, reference inputs, and search grounding.
- Names matter: Nano Banana Pro is a separate model. Our GPT tests use Broly’s
gpt-image-2.5alias; the Sunburst/Flare mapping is unspecified.
The table describes documented features, not test scores. OpenAI size controls e Google resolution options differ; native dimensions are labeled on every test image.
Public Benchmarks: How Do They Compare?
Sunburst leads both cited Arena rankings. Its preliminary status and smaller vote sample are part of that result.
Arena: generation and editing
September 7, 2026 snapshot · Scores on a 0–1,600 display scale
Texto para imagem
Single-image editing
- What it measures: anonymous user preferences, rather than exact spelling or cost per usable image.
- Configuration: the Nano entry uses web search; our supplied-content tests did not request it.
- Âmbito: Sunburst’s rank does not identify the Broly alias tested below. Arena’s preliminary-result policy also applies.
Same-Prompt Tests: Five Real Design Tasks
Escopo do teste: five briefs, identical prompts, and the same reference-image bytes. Tests ran through Broly on September 9, 2026.
- Requested settings: 1024 × 1024, medium provider label, no search requested. “Medium” is not a verified equivalent quality tier across models.
- Sample: 10 GPT and 11 Nano outputs, including a new two-model poster replacement pair. This is a small sample, not a reliability estimate.
- Fair comparisons: five pairs returned 1024 × 1024 on both sides. Five GPT originals returned 1254 × 1254; those pairs are excluded from quality ranking.
- Referências: a Nano-generated bottle with locally added label text and a separately generated fictional adult. Identical inputs control the source, but do not remove source-model bias.
- Judging criteria: wording, omissions, invented content, composition, edit preservation, and resemblance. Original PNGs are linked under the same WebP display treatment.
1. Text-Heavy Posters and Small Print
Create a polished square poster for a fictional design event. White background, charcoal typography, cobalt blue and coral accents. Use crisp editorial typography and generous margins. Include exactly these text strings, with no additions or substitutions: DESIGN AFTER HOURS Ideas worth staying up for FRIDAY, OCTOBER 16 6:30 PM - 9:00 PM NORTH HALL, STUDIO 08 TALKS / DEMOS / OPEN CRITIQUE EARLY ENTRY $18 Reserve your seat at designafterhours.example Make the event name largest, the date and time easy to find, and the URL small but legible. Use abstract geometric shapes only. No logos, photos, extra text, or decorative fake lettering.
Pair 1: native dimensions differ; excluded from head-to-head ranking.
New replacement pair: native dimensions differ; excluded from head-to-head ranking.
No matched-size winner. GPT run 1 and the new GPT/Nano replacement pair reproduce all eight required strings legibly. GPT uses a flatter geometric layout; Nano often presents a sheet with a shadow. These are useful design differences, but the GPT originals are 1254 pixels square, so we do not use small-print detail to rank them against 1024-pixel Nano outputs.
The original second Nano attempt remains below: its wording is intact, but the time separator changes from the specified hyphen to a longer dash. The corresponding GPT run ended with an upstream 524 and no image. We kept the failure record and generated a new pair on both models, rather than replacing only the weaker or failed side.
More layouts are covered in the Guia rápido do Nano Banana 2.
2. Product Photography and Material Detail
Create a square commercial studio photograph of one unbranded cylindrical clear-glass perfume bottle with a brushed silver metal cap and a blank matte white rectangular paper label. The bottle contains pale green liquid. Place it upright on a pale gray stone surface against a soft white background. Use one large softbox on the left, a narrow rim reflection on the right, and a believable contact shadow. Eye-level three-quarter view, entire bottle visible, precise edges, realistic glass refraction, fine brushed-metal texture. No other objects, no lettering, no logos, no watermark-like graphics.
Pair 1: native dimensions match.
Pair 2: native dimensions match.
No overall product winner. Both pairs match at 1024 × 1024. GPT gives the glass edges, metal brushing, and stone texture a crisp treatment; Nano also separates glass, metal, and paper convincingly. Nano run 1 includes an unwanted dark diagonal in the background, while run 2 is cleaner.
Neither model consistently demonstrates the requested three-quarter angle: the bottles look largely frontal. GPT run 2 also places stronger light on the right and shadow toward the left, which weakens compliance with the specified left softbox. A premium-looking bottle is useful for concepts, but exact camera and lighting direction still need checking.
3. Precise Edits Without Changing the Rest
Edit the supplied product photograph. Change only the label's background color to a muted teal. Preserve every printed word and its exact spelling, font, size, placement, and color. Keep the bottle, cap, liquid, reflections, lighting, shadows, background, camera angle, crop, and image dimensions unchanged. Do not redesign the label, add objects, sharpen the whole image, or restyle the photograph.
Pair 1: native dimensions match.
Pair 2: native dimensions differ; excluded from head-to-head ranking.
No clear winner in the matched first pair. Both models turn the label muted teal and retain FIELD NOTES, EAU DE PARFUM, and 50 mL. The bottle, cap, crop, and broad composition remain close to the source. Neither result is certified pixel-identical outside the label.
The second GPT edit preserves the wording but returns 1254 × 1254 despite the instruction to keep image dimensions unchanged. That is a concrete dimension-compliance failure, not evidence of superior detail. Nano retains 1024 × 1024 and produces a lighter teal. An exact brand-color workflow would benefit from a supplied color reference.
O GlobalGPT Image Edit guide explains how to specify a controlled change.
4. Character Consistency Across Scenes
Use the supplied fictional adult reference as the same person in both panels of one square image. Create two equal vertical panels with a clean narrow white divider. Left panel: the person standing outside a modern bookstore in daylight, waist-up, holding a closed notebook. Right panel: the same person seated at a cafe table in soft evening light, waist-up, with the notebook on the table. Preserve the reference face, hairline, hairstyle, glasses, jacket color, and shirt details in both panels. Keep a realistic photographic style. No captions, lettering, extra people, or changes of identity.
Pair 1: native dimensions differ; excluded from head-to-head ranking.
Pair 2: native dimensions match.
GPT follows the exclusions better in the matched second pair. Both retain the main reference features: short black hair, round glasses, a green jacket, and a white shirt. GPT shows the notebook held in the first panel and on the cafe table in the second, with no obvious storefront lettering or extra people. Nano adds shop text and background pedestrians despite explicit exclusions.
This is an instruction-following advantage in one matched pair, not a biometric identity score. GPT also introduces a visible shoulder bag in the bookstore panel, which the brief did not require. The first pair is shown for completeness but excluded from ranking because its native sizes differ. Both references originated with Nano, so the test does not eliminate source-model bias.
O subject consistency guide covers the wider reference-image workflow.
5. Infographics and Complex Layouts
Create a square, professional infographic on a white background titled exactly "FROM BRIEF TO LAUNCH". Below it, arrange five equally sized steps in one clear left-to-right sequence, connected by four right-pointing arrows. Use charcoal text with teal and coral accents, simple familiar icons, and clean sans-serif typography. Render these exact step titles and supporting lines: 1. DEFINE / Set one measurable goal 2. RESEARCH / Check audience needs 3. CREATE / Build the first draft 4. REVIEW / Fix factual and visual errors 5. LAUNCH / Publish and measure results Include every word exactly once. The arrows must connect 1 to 2 to 3 to 4 to 5, with no loops or branches. Keep every supporting line readable. Do not add charts, percentages, extra steps, decorative text, or a footer.
Pair 1: native dimensions differ; excluded from head-to-head ranking.
Pair 2: native dimensions match.
GPT wins content compliance in the matched second pair. Its five numbered steps, supporting lines, and four connecting arrows form the requested sequence without an extra row. Nano includes the main sequence but then adds duplicate REVIEW and LAUNCH blocks and invented text, violating the instruction to include every word exactly once.
The first pair tells a more limited story: both models omit the required step numbers, and GPT returns 1254 × 1254. It cannot support a matched-size win. The useful finding is therefore specific: GPT solved the strict content brief in the second 1024-pixel pair; Nano required correction in both displayed attempts. Two attempts do not establish a general success rate.
Official API Pricing and Recorded Tokens
GPT has the lower image-output token rate; Nano has the lower input rate. The rates below come from OpenAI’s pricing documentation e Google’s API pricing, checked September 9, 2026.
Official image-output token rate
USD per 1 million image output tokens · Standard API rates
- Additional categories: OpenAI cached text/image inputs are $1.25/$2 per million tokens when applicable. Google text/thinking output is $3 per million tokens.
- Google’s 1K example: 1,120 image output tokens × $60 / 1M = $0.0672 for output alone, before input and any separately billed text/thinking. This is a documentation example, not our test bill.
- Compare complete jobs: resolution, quality, input references, and additional attempts determine cost per usable image. A lower token rate alone does not make one model cheaper.
Recorded GPT Image 2.5 token usage · 10 outputs
Broly-reported input and image-output totals. These are usage records, not verified official billing units or a cross-model cost comparison.
What the Results Mean: Quality vs. Value
Accurate text and accurate information design are different strengths. Both models rendered the event wording well, but neither the posters nor their small print can establish a size-matched winner. The infographic offers a firmer distinction: GPT’s second 1024-pixel output preserves the sequence and content, while Nano’s adds invented material. If the asset explains a process, those additions create more editing work than a minor layout preference.
Product photography is less one-sided. GPT’s crisp glass and metal can help a concept feel finished, while Nano’s cleaner second background is easier to use. Both missed the clearly requested camera angle, and GPT’s second lighting direction needs correction. For a specific product campaign, compliance with the shot brief should matter more than surface polish alone.
The narrow label edit was a practical success for both models in the first matched pair. Character continuity exposed a different trade-off: GPT avoided Nano’s unwanted shop text and people in the second pair, while adding a shoulder bag. Resemblance, preservation, and exclusion compliance should be checked separately. A recognizable person does not automatically make the whole composition usable.
Value depends on the finished asset. GPT’s cleaner second infographic could reduce content correction; both models produced useful label edits. Official token rates add pricing context, but these results do not establish a measured cost-per-usable-image winner.
Qual você deve escolher? Start with the tested GPT route for this strict infographic brief; either model is a reasonable candidate for the narrow label edit. For posters and product shots, choose against your exact layout and lighting requirements. These findings do not establish a universal quality or price winner.
Perguntas frequentes
Is GPT Image 2.5 better than Nano Banana 2?
In our matched second infographic pair, GPT followed the content requirements better. It also avoided unwanted lettering and people in the matched second character pair. Product and label-edit results do not support a clear overall winner.
Which GPT Image 2.5 version was tested?
We tested Broly’s gpt-image-2.5 alias. The route returned images successfully but did not disclose a Sunburst or Flare mapping. The Sunburst leaderboard is separate context, not verified identification of the tested model.
Were the two models tested with identical settings?
The prompts, 1024 × 1024 target, medium provider label, and reference-image bytes matched. Five pairs also match at native 1024 × 1024. Five GPT results returned 1254 × 1254; they are excluded from paired ranking. The medium label is not a verified equivalent quality tier.
O Nano Banana 2 é igual ao Nano Banana Pro?
No. Google identifies Nano Banana 2 as Gemini 3.1 Flash Image. Nano Banana Pro is Gemini 3 Pro Image; they are separate models.
Which model is better for text-heavy infographics?
GPT performed better in the second native-size-matched pair: all five numbered steps and supporting lines remained in sequence without extra blocks. Nano added duplicate steps and invented text. This is a small-sample result, not a general success-rate claim.
Is GPT Image 2.5 cheaper?
Its official image-output token rate is lower: $30 versus $60 per million tokens. Nano Banana 2 has lower input rates. Total image cost depends on token consumption, resolution, quality, references, and any additional attempts.
Are these original model outputs?
Yes. The article shows 10 GPT Image 2.5 route outputs and 11 Nano Banana 2 outputs, including a fresh poster replacement pair. Original PNGs are linked; two prepared references and the official Google screenshot are labeled separately.
Posso usar os dois modelos no GlobalGPT?
We have not verified access to both exact models on GlobalGPT. Its Nano Banana 2 content and a GPT Image 2 listing do not establish availability of GPT Image 2.5.
