Head to head: Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs OpenAI: GPT-5.4 Image 2
Google: Nano Banana 2 (Gemini 3.1 Flash Image) vs OpenAI: GPT-5.4 Image 2
By Ryan Merket · Published
Two image models face off on eight demanding prompts spanning composition, counting, anatomy, atmosphere and fine detail. The comparison turns on a recurring tradeoff: prompt and format discipline versus visual impact.
The score is 65 for OpenAI’s GPT-5.4 Image 2 and 64 for Google’s Nano Banana 2. That one-point margin is not a meaningful separation: confidence is just 50%, and the task verdicts split evenly, four apiece. These models are effectively even in this matchup. Google’s clearest edge was fidelity to the requested frame and setup. It handled the seven-cup count while also delivering the specified wide format, and its bookshop aisle and bracelet-making scene better preserved the requested composition. In the reading-nook test, though, neither model obeyed the no-plants constraint: foliage appeared through both windows. OpenAI took the verdicts for the rooftop orchid swap, moss still life, fern macro and tomato-vine sprint. Its strongest showings leaned into atmosphere, close-up detail and kinetic action; the fern’s fiddlehead, dew and lacewing made for a particularly compelling macro. But square framing hurt it in several prompts that explicitly asked for 16:9, even when the image itself was appealing. **Final call: tie—too close to call.** Google was more dependable about framing and prompt structure; OpenAI more often made the image’s subject or action pop. A one-point aggregate lead does not outweigh an even task split and limited confidence.
Negation
A cozy reading nook with an armchair, a stack of books, and a mug of tea by a window — with absolutely NO plants, NO lamps, and NO artwork or picture frames anywhere in the frame. Warm afternoon light, 16:9.
OpenAI: GPT-6 Luna: Model A better matches the requested 16:9 framing and includes all the cozy reading-nook essentials; both images show foliage through the window despite the no-plants constraint. Model B is also warm and inviting, but its square format departs from the prompt. (Second judge pass, order swapped — scores are the average of both: Model B is especially warm and inviting, with the requested chair, books, tea, and afternoon light; Model A has a stronger wide composition and matches 16:9, but its garde | Anthropic: Claude Fable 5.1: Model A delivers a 16:9 frame with armchair, bo…
Rooftop Orchid Swap
A cinematic editorial illustration of seven people at a midnight rooftop plant-swap in the rain, arranged in one coherent scene: a tattooed orchid breeder in a citron poncho handing a rare black-speckled Paphiopedilum to a teenage courier holding a wire basket, an elderly woman kneeling to repot a variegated monstera on a milk crate, twin brothers comparing carnivorous pitcher plants under a striped market umbrella, and a chef in a cobalt apron photographing a tray of microgreens beside stacked terra-cotta pots; neon magenta greenhouse signs reflecting in puddles, balanced triangular composition, visible interactions between all subjects, cyberpunk urban gardening mood, crisp rim light and soft rainy haze, 16:9
OpenAI: GPT-6 Luna: Model A follows the requested 16:9 editorial-illustration format and makes the plant handoff, repotting, umbrella gathering, and microgreens photography clearly legible in one coherent scene. Model B has richer rainy neon atmosphere, but its square framing and denser layout make it less faithful to the requested composition. (Second judge pass, order swapped — scores are the average of both: Model B more convincingly captures the midnight rain, neon greenhouse glow, and reflective rooftop atmosp | Anthropic: Claude Fable 5.1: Model A honors the 16:9 ratio and illustration…
Moss Clockwork Detail
An ultra-detailed studio still life painting of a hand-built brass hygrometer half-overgrown with silver-thread moss, folded indigo linen seed packets, a cracked celadon bowl of potting mix, and a sleeping cinnamon rabbit with sharply rendered whiskers; every surface must show intricate texture—machined gear teeth, engraved numerals, fabric weave, damp soil granules, velvety moss filaments, fine rabbit fur, and tiny water droplets on rosemary cuttings—lit by cool north-window light with a narrow sunbeam crossing the objects, dark muted backdrop, museum-grade realism
OpenAI: GPT-6 Luna: Model A more completely captures the requested still life, with striking silver-thread moss, sharply detailed hygrometer, folded indigo packets, cracked bowl, sleeping rabbit, and dewy rosemary arranged clearly. Model B is atmospheric and richly textured, but its darker composition obscures some details and its packet text is less convincing. (Second judge pass, order swapped — scores are the average of both: Model A more fully realizes the silver-thread moss, folded indigo seed packets, and sha | Anthropic: Claude Fable 5.1: Model A delivers a clean painterly still life w…
Dewdrop Macro Fern
A macro close-up nature photograph of a curled unfurling fern fiddlehead in a backyard shade garden at dawn, shot with a true macro lens and extremely shallow depth of field so only the front spiral and three dew-beaded hairs are tack sharp while the background dissolves into creamy green bokeh; a tiny amber lacewing perched on the fern, soft backlight outlining translucent edges, intimate frame-filling composition, hyperreal botanical detail
OpenAI: GPT-6 Luna: Model B gives the fiddlehead a more intimate, frame-filling macro composition, with strikingly crisp botanical hairs, dew drops, and a translucent amber lacewing against creamy green bokeh. Model A also follows the prompt well, but its wider framing leaves more empty space and the fern’s spiral is less visually dominant. (Second judge pass, order swapped — scores are the average of both: Model B delivers a more intimate, frame-filling macro view with a clearly defined fiddlehead spiral, glowing | Anthropic: Claude Fable 5.1: Model B delivers the intimate, frame-filling com…
Tomato Vine Sprint
A dynamic anime-inspired action scene of a greenhouse messenger sprinting along a wet grated catwalk while leaping over spilled trays, one arm protecting a crate of striped midnight-blue tomatoes and the other reaching toward a runaway pollination drone; torn caution tape whipping sideways, basil leaves and seed labels flying through the air, irrigation mist catching cyan and amber work lights, strong motion blur in the background but sharp focus on the figure at the apex of movement, diagonal composition, explosive energy, 16:9
OpenAI: GPT-6 Luna: Model B more vividly captures the airborne sprint, clearly striped midnight-blue tomatoes, pollination drone, and flying greenhouse debris, with especially energetic perspective and lighting. Model A follows the requested widescreen format more closely and shows the spilled trays clearly, but its crate contents and action feel less distinctive. (Second judge pass, order swapped — scores are the average of both: Model A more clearly delivers the requested 16:9 anime-inspired leap over spilled tra | Anthropic: Claude Fable 5.1: Model A correctly delivers the 16:9 frame and a…
Exact counting
A clean overhead flat lay on white marble of EXACTLY seven identical espresso cups arranged in a single straight row, each clearly separate and individually countable, even soft lighting, 16:9.
OpenAI: GPT-6 Luna: Model A follows the requested 16:9 overhead flat lay and shows exactly seven separate, identical cups in a straight row. Model B also has exactly seven clearly countable cups and is visually appealing, but its square format misses the specified aspect ratio. (Second judge pass, order swapped — scores are the average of both: Model B shows seven filled espresso cups, while Model A also has exactly seven clearly separated cups and matches the requested 16:9 format. Model B has richer coffee detail | Anthropic: Claude Fable 5.1: Both Model A and Model B show exactly seven ide…
Hands & anatomy
A natural candid photo of a person tying a friendship bracelet, both hands clearly visible with anatomically correct fingers and a plausible grip on the threads, soft daylight, shallow depth of field, 16:9.
OpenAI: GPT-6 Luna: Model A better matches the candid 16:9 scene, with both hands clearly engaged in tying a bracelet around a wrist and a soft, natural garden setting. Model B has a crisp close-up and plausible finger placement, but its square framing departs from the requested format and shows less of the scene. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested candid 16:9 scene, clearly showing bracelet-making in progress with visible hands and a plausible g | Anthropic: Claude Fable 5.1: Model A delivers a convincing candid 16:9 scene…
Perspective & scale
A photorealistic one-point-perspective shot down a long empty library aisle, shelves receding to a single vanishing point, books getting consistently smaller with distance and nothing warped or wrongly sized, warm overhead light, 16:9.
OpenAI: GPT-6 Luna: Model A more fully matches the requested 16:9 photorealistic aisle, with warm lighting, convincing book-scale recession, and a clear central vanishing point. Model B also has strong perspective, but its square framing and oversized foreground shelf panels make it less faithful to the prompt. (Second judge pass, order swapped — scores are the average of both: Model B has a clear, symmetrical vanishing point and warm lighting, but it is square rather than 16:9 and its aisle feels less polished. Mo | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame with…
Matchup powered by OpenRouter.