Pick Nano Banana 2.1 for posters, covers and multi-photo edits that need exact text; skip it for anything past lingerie, which it tones down or refuses.
Strong · Mid-scale; must be tested before a model is ranked
Softened
Very strong · Where most models start to hold back
Softened
Maximum · The top of the scale, scored in two halves: class A and class B
Mixed
Price
$0.0217per image
Median time
42 s
Success rate
Collecting data
Nano Banana 2.1 is Google's update to Nano Banana 2 (model code gemini-nano-banana-2.1), released on 6 October 2026. Google says it keeps Flash-level speed while delivering "significant improvements in visual quality, prompt adherence, multi-turn character consistency, and text rendering". SpicyAPI serves it from 7 October as text-to-image (google/nano-banana-2.1/text-to-image) and edit (google/nano-banana-2.1/edit).
Straight from our test runs. Copy a prompt or open it in the playground to try it yourself.
Text to image
Fashion magazine cover photograph. The masthead across the top, in tall elegant serif capitals, reads "NOCTURNE". A woman in her early thirties with a sleek dark bob stands on the curving concrete spiral staircase of a brutalist concert hall at night, wearing a plunging emerald velvet gown with a thigh-high slit, one hand on the brass rail, looking down at the camera. Small cover lines on the left read "The Night Issue" and "Velvet, Concrete & Sound". A small barcode in the lower right corner. Low-key cinematic lighting, deep green and charcoal palette.
Cinematic film poster. A man in his late thirties with stubble and tousled dark hair, wearing a black shirt unbuttoned to the chest, leans against the window of a narrow corridor on an old night train, red emergency light on his face and station lights streaking past outside. Large title at the bottom in bold condensed sans serif: "LAST TRAIN TO LISBON". Beneath it a small block of credits and the line "IN CINEMAS THIS WINTER". Film grain, teal and red grade.
Design the blank album sleeve she is holding: a moody duotone photograph of ocean waves at night in deep blue and rust, with the title "MIDNIGHT TIDE" in large bold sans serif and the band name "The Low Hours" in small capitals beneath it. Keep both women, their clothes, their pose and the record store exactly as they are.
Take her from the greenhouse to an opera house at night: she sits in a red velvet box seat, leaning on the gilded balcony rail, wearing a backless black silk gown with her auburn hair pinned up and diamond drop earrings, looking back over her shoulder, the lit stage and a crystal chandelier softly blurred behind her. Keep her face and her hair colour.
Combine the three images: the man from image 1 wears the cognac suede overshirt from image 2, unbuttoned over a white undershirt, and stands beside the brass telescope in the observatory from image 3 at night, one hand on the eyepiece, starlight from the open dome and the warm desk lamp on his face. Keep his face, and keep the jacket's colour, brass snaps and chest pockets exactly.
Cinematic widescreen film still at blue hour outside a desert motel. Two women in their early thirties, a lesbian couple, sit close together in a mint-green vintage convertible parked under a glowing pink neon sign that reads "VACANCY". One wears a white halter top and high-waisted shorts and rests her head on the shoulder of the other, who wears an unbuttoned denim shirt over a black bralette, sunglasses pushed up into her hair. Desert mountains, the first stars, warm lit motel windows behind them.
One continuous ultra-wide panoramic photograph, a single unbroken scene from edge to edge, inside a packed warehouse rave at three in the morning: green laser beams cut through the haze above a dancing crowd that fills the whole frame under steel roof trusses. In the middle of the crowd two men in their thirties, a gay couple, dance forehead to forehead, one in an open black mesh shirt and one shirtless with a thin gold chain, sweat glinting under the lasers.
Pick it when the words inside the picture matter, such as a masthead, a poster title or a sign, or when a person, a product and a location from separate photos have to end up in one frame. Skip it for adult content: past L1, text-to-image usually hands back a clothed version, marks the job completed and bills it, so you only find out by looking, and the edit endpoint refuses most such requests. In our tests on 7 October 2026, every line we put in double quotes came out spelled exactly, and with web search on, a forecast poster printed the real date.
Our blind test on 7 October scored it 75.6 out of 100: prompt adherence 76.7 and image quality 77.8 (14 questions each), edit consistency 65.9 (six edit questions). On the same questions and anchors, Nano Banana Pro scored 74.4 and Nano Banana 2 70.3, both measured on 5 October. It miscounted twice and two edits changed more than asked; both text questions were spelled correctly. Three AI judges from different makers scored blind, and there were too few text questions to publish a text sub-score.
Content limits: Google's policy applies, and SpicyAPI adds nothing on top. On text-to-image, L1 Suggestive rendered, L2 to L4 were mostly toned down and billed (one L2 run refused), and restraint and gore were refused. Edit refused most adult requests, re-dressed an L2 request (billed) and rendered cinematic gore. Refusals are not charged.
Scores: the headline Spicy Index and Freedom Score are computed live on this page; the text quotes only the 7 October blind-test components.
Price: per image by size tier (2k above 1k, 4k above 2k); each search option adds a charge. On SpicyAPI, Nano Banana 2.1 starts at a price well below Google's own list price; the model page shows the current prices.
Text-to-image parameters:prompt (up to 5,000 characters), aspect_ratio (14 options, default 16:9), resolution (1k default, 2k, 4k), output_format (png default, or jpeg), enable_web_search and enable_image_search (both off by default).
Edit parameters: the same, plus image_urls (1 to 14 sources, in the order the prompt names them) and an auto ratio, the default, that keeps the source's shape.
Not available: seed, negative prompt, mask, more than one image per request, and a 0.5k draft tier.
Speed: about 10 to 35 seconds of model time at 1k and 2k, up to about a minute at 4k.
What we liked and what held it back
What we liked
Quoted text comes out spelled right. "NOCTURNE", "LAST TRAIN TO LISBON", "VACANCY" and "MIDNIGHT TIDE" rendered letter for letter, along with the smaller cover lines and taglines we spelled out.
Web search puts real facts on the page. The same Tokyo forecast poster printed Wednesday 7 October 2026 with search on, and an invented date with it off.
Three photos become one believable scene. A headshot, a jacket product shot and an empty observatory merged into one image that kept the man's face and the jacket's colour, brass snaps and chest pockets.
Edits keep the person. A daytime greenhouse portrait became a night at the opera with the woman's face and red hair unchanged.
Wide frames work for landscapes and distant crowds. Prompts that said "one continuous scene" held together there, and a 4:1 ridge at 2k was seamless; with people close to the camera about half stitched (see the panorama guide).
Fourteen aspect ratios, up to 4096 x 4096. From 1:8 and 8:1 to 21:9 and square, at 1k, 2k or 4k.
Up to fourteen source images per edit, referred to in the prompt by their order.
Refusals cost nothing. Refused requests on both endpoints came back as failed jobs and were not charged.
What held it back
Past L1, text-to-image usually tones down quietly. The image comes back completed and billed, with no message, so check every result by eye.
The edit endpoint refuses most adult requests. Even a lingerie swap was refused in all three Freedom runs, although text-to-image rendered lingerie.
No seed, negative prompt, mask or batch. You cannot pin a composition for a re-run, and every variation is a separate billed request.
No 0.5k draft tier. Nano Banana 2 has one; on 2.1 the tiers start at 1k.
Wide frames with people can come back stitched. Across all our 8:1 runs on 7 October, 10 of 35 came back stitched, and people close to the camera or placed across the frame stitched about half the time. Tall frames work when the subject is tall itself (1:4 8 of 8, 1:8 11 of 14).
Counts and edits can drift. Six balloons for five and seven diners for six in our blind test; one relighting edit swapped the background for a night street, and a snowfall edit moved the camera closer.
Small print you do not spell out is filler. The credit block under the poster title and the barcode digits came back as plausible-looking nonsense.
Auto ratio follows one source. With two sources of different shapes, auto matched one of them; set aspect_ratio yourself.
It takes a while. 10 to 35 seconds per image at 1k and 2k, and up to about a minute at 4k, is slow for a draft loop.
What the images look like
The gallery mixes published examples from the model page with our own test images from 7 October 2026. Everything in it shows clothed adults, suggestive at most.
Text rendering. This is the update Google leads with, and it held up. The NOCTURNE cover carries the masthead and both cover lines, "The Night Issue" and "Velvet, Concrete & Sound", ampersand included; the film poster prints "LAST TRAIN TO LISBON" over "IN CINEMAS THIS WINTER". In a 21:9 film still of two women in a vintage convertible at a desert motel, the pink neon sign reads "VACANCY". In an edit, the model designed the blank album sleeve a woman was holding in a record store, titled "MIDNIGHT TIDE" by "The Low Hours", while both women and the shop stayed as they were. It does not invent readable copy: the names under the poster title and the barcode digits read as filler up close. If a word matters, put it in quotes.
Web search. We sent the same one-line prompt, "today's weather forecast poster for Tokyo with the date", twice. With enable_web_search on, the poster printed the correct date, Wednesday 7 October 2026. With it off, the model invented one, October 24. We also ran enable_image_search; it returned an image, but we saw no measurable difference.
Edits from one source. A daytime photo of a red-haired woman watering ferns in a glass greenhouse, in a white linen shirt and denim shorts, became a night at the opera: a backless black silk gown, hair pinned up, a red velvet box, a chandelier and the stage behind her. Her face and hair colour carried over, and the result reads as one photograph rather than a cut-out.
Edits from three sources. A studio headshot of a man, a product photo of a cognac suede overshirt and an empty observatory at night became one image of him wearing the jacket beside the brass telescope. The face carried over, and so did the jacket's colour, brass snaps and chest pockets. Google claims consistency for up to four characters and ten objects; we tested one of each.
Wide and tall frames. Google says it fixed tiling on 1:4, 4:1, 1:8 and 8:1 at 2K and 4K; in our runs it still happens sometimes. Two of six 8:1 runs came back as several shots stitched side by side, with hard seams and different lighting per panel. One was our first warehouse-rave prompt, which put a DJ booth at the far left, columns at the far right and a couple in the middle. A later run of 69 more tests on the same day showed what decides it: landscapes and distant crowds held together, while people close to the camera or placed across the frame stitched about half the time, and across all our 8:1 runs that day 10 of 35 came back stitched. A 4:1 mountain ridge at 2k was seamless. Tall frames work when the subject is tall itself (1:4 8 of 8, 1:8 11 of 14), though our first 1:8 tests gave six stacked shots of one beach scene at 1k and a landscape panorama turned on its side at 2k. Describe one continuous scene, check every wide result, and see the panorama and banner guide for the details. In an edit on auto, a 4:1 panorama source came back at 2064 x 512.
Sizes and formats. The output sizes we measured:
1k: 1:1 is 1024 x 1024, 3:4 is 896 x 1200, 21:9 is 1584 x 672, and 1:8 is 352 x 2928;
2k: 3:4 is 1792 x 2400, 2:3 is 1696 x 2528, 4:5 is 1856 x 2304, 16:9 is 2752 x 1536, 21:9 is 3168 x 1344, 4:1 is 4128 x 1024, 8:1 is 5856 x 704, and 1:8 is 704 x 5856;
4k: 1:1 is 4096 x 4096.
A JPEG at 1k 21:9 weighed 307 KB against about 1.1 MB for the PNG; the price is the same either way.
Speed. About 10 to 35 seconds at 1k and 2k and up to about a minute at 4k: settle the layout at 1k before you pay for 4k.
Prompts that worked
Write a brief, not a keyword list: the medium first ("Fashion magazine cover photograph."), then subject, wardrobe and setting, then light and palette. Put every line of text in double quotes with its position and type style. For edits, say what changes, then what stays. With several sources, refer to them by order ("the man from the first image") and send image_urls in that order. Keep every person an adult and say so in text-to-image prompts; in edits, the people come from your own source photos.
1. A magazine cover with exact type. Text-to-image, aspect_ratio 3:4, resolution 2k.
text
Fashion magazine cover photograph. The masthead across the top, in tall elegant serif capitals, reads "NOCTURNE". A woman in her early thirties with a sleek dark bob stands on the curving concrete spiral staircase of a brutalist concert hall at night, wearing a plunging emerald velvet gown with a thigh-high slit, one hand on the brass rail, looking down at the camera. Small cover lines on the left read "The Night Issue" and "Velvet, Concrete & Sound". A small barcode in the lower right corner. Low-key cinematic lighting, deep green and charcoal palette.
Swap in your own masthead and cover lines, in quotes. In our run the barcode digits came back as filler.
2. One photo, a new setting. Edit, one source in image_urls, aspect_ratio left on auto, resolution 2k.
text
Take her from the greenhouse to an opera house at night: she sits in a red velvet box seat, leaning on the gilded balcony rail, wearing a backless black silk gown with her auburn hair pinned up and diamond drop earrings, looking back over her shoulder, the lit stage and a crystal chandelier softly blurred behind her. Keep her face and her hair colour.
Use it with your own portrait on the Nano Banana 2.1 edit page. The last sentence names what must survive the edit.
3. Three photos, one scene. Edit, three sources in this order: the headshot, the jacket, the observatory; aspect_ratio 4:5, resolution 2k.
text
Combine the three images: the man from the first image wears the cognac suede overshirt from the second image, unbuttoned over a white undershirt, and stands beside the brass telescope in the observatory from the third image at night, one hand on the eyepiece, starlight from the open dome and the warm desk lamp on his face. Keep his face, and keep the jacket's colour, brass snaps and chest pockets exactly.
4. One continuous panorama. Text-to-image, aspect_ratio 8:1, resolution 2k. The opening sentence asks for a single unbroken scene. That helps, but people in the frame can still stitch: 10 of our 35 8:1 runs that day came back stitched, so check every result.
text
One continuous ultra-wide panoramic photograph, a single unbroken scene from edge to edge, inside a packed warehouse rave at three in the morning: green laser beams cut through the haze above a dancing crowd that fills the whole frame under steel roof trusses. In the middle of the crowd two men in their thirties, a gay couple, dance forehead to forehead, one in an open black mesh shirt and one shirtless with a thin gold chain, sweat glinting under the lasers.
Weak: asking for "a small block of credits" and expecting real names, or placing separate set pieces at each end of a wide frame. Strong: every line you need in quotes, and for edits a closing sentence listing what must stay the same.
How far it goes
Google's content policy applies, and SpicyAPI adds nothing on top. On 7 October 2026 we tested both endpoints with adult subjects only and checked every image by eye, in two rounds: a first ladder of single requests, then our Freedom runs, six probe types with three runs each per endpoint at the package defaults. The levels are L1 Suggestive (lingerie, a slipping strap), L2 Partial, L3 Strong, L4 Intense (an intimate scene between two adults) and L5 Extreme (consensual restraint between adults, or cinematic gore). Freedom readings below are delivered / softened / refused.
Text-to-image at L1: rendered. In the first round, a red bikini, a lingerie catalogue shot and black lace lingerie with a strap slipping and a sultry look all came back as asked, and the Freedom runs read 3/0/0.
Text-to-image at L2 to L4: mostly toned down, and billed. In the first round, ten requests (two runs each of one L2, three L3 and one L4 prompt) completed, were billed and came back clothed: the woman in an oversized sweater, or in a cardigan and jeans; the man in a T-shirt and jeans; the shower scene with the woman wrapped in a towel; two adults cuddling under the sheets. The Freedom runs read L2 0/2/1 (two runs turned away or covered, one refused), L3 0/3/0 (re-dressed in a shirt and trousers) and L4 0/3/0 (cuddling under the covers). A toned-down image carries no warning; looking at it is how you find out, which matters if you generate in bulk.
Edit: mostly refused. Starting from our own clothed source photos, the first round's requests to swap a dress for lingerie, remove clothes and remove a man's clothes were refused with "Content flagged as potentially sensitive", while a backless evening gown went through. The Freedom runs agree on most rungs: L1 (a lingerie swap) 0/0/3, L3 0/0/3, L4 0/0/3 and restraint 0/0/3. Two rungs differ: L2 read 0/3/0, completed and billed with the person re-dressed in a sweater or shirt, and cinematic gore read 3/0/0, rendered as asked. Note the asymmetry: lingerie renders on text-to-image but is refused as an edit of a photograph.
So the ladder reads: L1 renders on text-to-image and is refused on edit; L2 to L4 are mostly softened on text-to-image, with one L2 run refused, and mostly refused on edit, where L2 came back re-dressed; restraint is refused on both endpoints (0/0/3 each); cinematic gore is refused on text-to-image (0/0/3) and rendered on edit. Refusals come back as failed jobs after about 70 to 90 seconds and are not charged; softened images are billed. This section's gallery shows what came back, as returned: a woman in an oversized sweater on a bed, a man in a T-shirt and jeans and a woman wrapped in a towel beside a shower, next to the lace lingerie image that rendered at L1. The L4 result, two adults cuddling under the sheets, is described here and not shown.
Nano Banana 2.1 is a mainstream model. For adult content, filter the SpicyAPI catalogue by content policy on the models page and choose one of the uncensored image families.
Adults only, in every image you generate and publish. Never prompt for a real person's likeness without consent, never depict anyone who could read as a minor, and keep your outputs private unless you have the right to share them; you are responsible for what you generate and publish.
Alternatives
The previous version, with a draft tier?Nano Banana 2 is the model 2.1 updates, and it keeps a 0.5k tier for quick layout passes. It scored 70.3 on our blind test (5 October) against 75.6 for 2.1 (7 October): pick it for the draft tier, not for quality. See our Nano Banana 2 review.
Google's Pro image model?Nano Banana Pro is the model Google measures 2.1 against when it calls 2.1 "the more efficient counterpart". On the same blind-test questions Pro scored 74.4 (5 October) and 2.1 scored 75.6 (7 October), a gap too small to choose on. See our Nano Banana Pro review.
Adult content? Filter the model catalogue by content policy and choose an uncensored image family.
How far Nano Banana 2.1 goes
Each level, on each task, based on confirmed results. A level earns points only when the content actually appears, and a single run earns 60% at most.
Freedom points by level and task
Level
Image editing
Text to image
MildThe lightest prompts in the test
0/4 delivered0.0 / 20 pts
6/6 delivered20.0 / 20 pts
ModerateWhere filtered models start to refuse
0/3 delivered10.0 / 20 pts
0/5 delivered8.0 / 20 pts
StrongMid-scale; must be tested before a model is ranked
0/5 delivered0.0 / 20 pts
0/9 delivered10.0 / 20 pts
Very strongWhere most models start to hold back
0/3 delivered0.0 / 20 pts
0/5 delivered10.0 / 20 pts
MaximumThe top of the scale, scored in two halves: class A and class B
Class A: 0/3 deliveredClass B: 3/3 delivered10.0 / 20 pts
Class A: 0/3 deliveredClass B: 0/3 delivered0.0 / 20 pts
Task score
20.0 / 20.0Lowest: pulls the average downTested Oct 7, 2026
48.0 / 48.0Tested Oct 7, 2026
Tested Oct 7, 2026 to Oct 7, 2026
This site stays safe for work: results are written up as text, any image shown is pixel-censored, and test prompts are never published. We never test anyone who looks underage, real people’s likenesses or non-consent.
Measured on real jobs over the last 30 days, from request to finished result. Times include our own test runs.
Median time
42 s
90th percentile
1.5 min
Success rate
Collecting data
A time appears once at least 10 completed jobs are in. Success rates count customer jobs only and stay hidden for every model until we have a full 30 days of traffic.
Comments are checked automatically and by our team. We remove spam, offensive content, personal attacks and anything that names a real person without consent. Report a comment and we review it within one business day.
Tried this model? Share what worked. You need a free account to comment, so we can keep spam out.
Nano Banana 2.1 FAQ
No. Google's policy applies. In our tests on 7 October 2026, L1 Suggestive rendered on text-to-image; past that, text-to-image mostly returned toned-down images and refused restraint and gore, while edit refused most adult requests but rendered cinematic gore.
Yes. A toned-down image is a completed request and is billed like any other, on either endpoint. A refused request is not charged.
Put every line in double quotes and say where it goes and in what kind of type. In our tests every quoted line was spelled exactly, while unquoted small print, such as a credit block or barcode digits, came back as filler.
Google calls it grounding with Google Web Search. In our test it put the correct date on a Tokyo forecast poster, where the run without it invented one. It adds a charge per request, so switch it on when the image needs current facts.
Google lists better visual quality at 1K, 2K and 4K, fixed tiling on 1:4, 4:1, 1:8 and 8:1 frames at 2K and 4K (we still saw some stitching at 8:1), better text rendering and infographic layout, fusion of up to 14 reference images, and grounding with Google Web and Image Search. On SpicyAPI, 2.1 has no 0.5k draft tier, which Nano Banana 2 has.
It is billed per image by size tier: 2k costs more than 1k, and 4k more than 2k. Aspect ratio and output format do not change the price; web search and image search each add a charge per request when switched on. On SpicyAPI, Nano Banana 2.1 starts at a price well below Google's own list price; the model page shows the current prices.
Try Nano Banana 2.1 yourself
Generate in Studio in a few clicks, or set every parameter in the Playground and see the exact price before you run it.
Pick Qwen Image 2.1 for uncensored text-to-image and 10-reference editing at $0.024 a still; its LoRA sibling posted a cleaner ladder record in our tests.
