Head to head: Google: Nano Banana Pro (Gemini 3 Pro Image) vs OpenAI: GPT-5.4 Image 2
Google: Nano Banana Pro (Gemini 3 Pro Image) vs OpenAI: GPT-5.4 Image 2
By Ryan Merket · Published
This matchup tests more than image polish: it asks both models to handle strict framing, exact counts, legible text, named art direction and a specific emotional portrait. The results turn on how consistently each model follows those different kinds of instructions.
On the saved task sample, Google scored **64.3** to OpenAI’s **60.8**, with six task wins to two and no ties. The editorial assessment is a **clear result (not statistical)**—a verdict on these tasks, not a claim about broader model capability. Google’s advantage came from reliable prompt fidelity across varied scenes: it better matched the requested wide framing in the rainy tram stop and tabletop scene, followed the restricted palette, and delivered the requested counts in the cup and helmet tasks. It also won the kiosk-poster task. Across these results, getting the specified composition and details right mattered more than surface polish alone. OpenAI’s two wins were more focused. Its ukiyo-e image made the fishing boat and dramatic wave the clear subjects, while its night-bus portrait better captured the intimate pose and relief in the prompt. Those are meaningful strengths, but they did not outweigh Google’s consistency across the other six tasks. **Verdict: Google: Nano Banana Pro (Gemini 3 Pro Image) wins this saved task sample.** The assessment applies only to these eight tasks; it should not be read as a general ranking of the models.
Rainy tram-stop realism
Photorealistic 16:9 street-level scene at a narrow tram stop on Vesper Lane just after rain, a brushed stainless-steel ticket machine beside a wet varnished oak bench, a commuter in a charcoal wool coat adjusting a teal umbrella, a cyclist passing in the background, puddles reflecting amber tram lights, believable skin, fabric, metal, glass, and rain-darkened concrete, low evening light from a break in the clouds casting soft shadows and cool reflections, shot like a high-end documentary photograph with natural perspective and physically convincing materials.
OpenAI: GPT-6 Luna: Model A better matches the requested 16:9 documentary photograph, with convincing wet reflections and all key scene elements, while Model B adds a clear Vesper Lane sign but is square and its ticket machine looks less like brushed stainless steel. Model A also has the stronger natural perspective and cohesive rainy-evening atmosphere. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested 16:9 documentary framing, with convincing wet reflections, | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 frame with…
Named art style
A ukiyo-e woodblock print of a fishing boat riding a large cresting wave at dawn, faithful to the flat color planes, bold outlines, and stylized foam of the tradition, muted indigo and cream palette.
OpenAI: GPT-6 Luna: Model B more clearly centers a fishing boat riding a dramatic cresting wave and captures the requested ukiyo-e outlines, flat indigo-and-cream planes, and stylized foam. Model A is also attractive and stylistically convincing, but its multiple boats and expansive open sky make the scene less focused. (Second judge pass, order swapped — scores are the average of both: Model B more fully matches the prompt, with a prominent fishing boat, a towering stylized wave, dawn light, and a muted indigo-and | Anthropic: Claude Fable 5.1: Model B delivers a single fishing boat with vis…
Restricted palette
A flat-vector illustration of a mountain campsite at dusk using ONLY four colors — cream, burnt orange, teal, and charcoal — with no gradients and no stray off-palette colors, balanced composition, 16:9.
OpenAI: GPT-6 Luna: Model A follows the four-color, flat-vector treatment and 16:9 format while clearly depicting a complete campsite at dusk. Model B is polished, but its square format and visible gradients conflict with the prompt’s explicit constraints. (Second judge pass, order swapped — scores are the average of both: Model A better follows the 16:9 format and presents a clear, balanced campsite scene in a restrained flat-vector style. Model B has strong visual appeal and the requested subject matter, but it i | Anthropic: Claude Fable 5.1: Model A delivers the required 16:9 aspect ratio…
Kiosk poster typography
Photorealistic 16:9 urban sidewalk scene centered on a backlit poster inside a glass bus-shelter kiosk at dusk, the poster cleanly printed in bold modern sans-serif typography and fully legible, reading exactly: NIGHT MARKET / FRIDAY 19 OCT / ELM QUAY / 7–11 PM, with no other prominent text competing for attention, passersby softly blurred in the background, cool blue ambient light contrasting with the warm poster glow, straight-on composition for easy reading.
OpenAI: GPT-6 Luna: Model B centers the kiosk in a straight-on composition and renders all requested wording clearly, with the cool dusk and warm poster glow well balanced. Model A also has legible, accurate text, but its poster is more angled and dominates the frame, while Model B’s poster artwork is slightly busier than requested. (Second judge pass, order swapped — scores are the average of both: Model A has cleaner, more legible typography with the requested wording and matches the 16:9 format, though its kiosk | Anthropic: Claude Fable 5.1: Both render the exact text correctly in bold sa…
Attribute binding
A photorealistic tabletop scene: a matte green cube to the LEFT of a glossy red sphere, a small blue cylinder BEHIND both, and a single yellow rubber duck sitting ON TOP of the cube. Colors must stay on the correct objects. Soft studio light, 16:9.
OpenAI: GPT-6 Luna: Model A more convincingly matches the photorealistic tabletop brief and the requested 16:9 composition, with the correct object colors and placements. Model B is clean and clearly arranged, but looks more like a polished render and uses a square frame. (Second judge pass, order swapped — scores are the average of both: Model A better matches the requested 16:9 tabletop scene while keeping all objects and their color and spatial relationships correct. Model B has cleaner studio lighting, but its | Anthropic: Claude Fable 5.1: Model A nails every binding (matte green cube le…
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 matches the requested 16:9 overhead layout with exactly seven clearly separated cups in a straight row. Model B also has seven cups and appealing espresso detail, but its square format departs from the requested aspect ratio and leaves more unused space. (Second judge pass, order swapped — scores are the average of both: Model B clearly shows seven espresso-filled cups on marble, but its square format misses the requested 16:9 aspect ratio. Model A has exactly seven individually countabl | Anthropic: Claude Fable 5.1: Both show exactly seven identical, clearly sepa…
Relief on the night bus
A photorealistic close three-quarter portrait inside a nearly empty city night bus, a young paramedic in a navy uniform visibly showing relieved exhaustion after a difficult shift, eyes glossy, shoulders dropping, a small genuine exhale-smile, one hand still gripping a crumpled orange dispatch slip, the other resting on the vinyl seat, sodium streetlights streaking through rain-speckled windows behind them, subtle pores and skin texture, cinematic shallow depth of field, intimate composition that makes the emotion instantly recognizable.
OpenAI: GPT-6 Luna: Model B more closely matches the requested intimate three-quarter portrait, with a clear relieved expression, the orange dispatch slip in one hand, the other resting on the seat, and rain-lit night-bus details. Model A captures the setting and emotion too, but is framed wider and keeps the slip in both hands rather than showing the requested pose. (Second judge pass, order swapped — scores are the average of both: Model B better captures the requested close, intimate portrait, with a clear relie | Anthropic: Claude Fable 5.1: Model A delivers the more convincingly photogra…
Twelve lantern helmets
A detailed contemporary oil painting of an alley bicycle repair pop-up on Mercer Cut, viewed straight on in 16:9, with EXACTLY 12 distinct matte-yellow bike helmets hanging in a single row from a corrugated metal shutter, each fully visible and individually countable, no extra helmets anywhere else, a mechanic’s red toolbox on the ground, chalk marks on wet asphalt, muted overcast daylight, crisp spacing and composition designed so the count is unambiguous.
OpenAI: GPT-6 Luna: Model A shows exactly 12 individually countable matte-yellow helmets in a wide, straight-on oil-painted scene, with the red toolbox and chalk-marked wet pavement. Model B is visually strong but appears to show only 11 helmets and uses a near-square composition rather than the requested 16:9 framing. (Second judge pass, order swapped — scores are the average of both: Model A shows exactly 12 individually visible helmets in a single row and follows the requested wide, straight-on composition. Mode | Anthropic: Claude Fable 5.1: Model A delivers the requested 16:9 straight-on…
Matchup powered by OpenRouter.