Create and edit images with Nano Banana 2.1.
Describe what you want in plain language — precise text rendering, subject consistency, and results in seconds.
No design skills needed — just describe what you want.
Many pages describing Nano Banana 2.1 list resolution tiers, reference limits and prices as settled facts. None of that comes from Google — and if you are building on those numbers, you are building on someone else's guess.<br/><br/>What Google has confirmed is narrow: a model label called Nano Banana 2.1 is selectable in the Google Flow image-generation interface. Screenshots of that picker, listing Nano Banana Pro, Nano Banana 2 Lite and Nano Banana 2.1 with 2.1 selected, were published on October 5, 2026. That single screenshot is the entire confirmed set.<br/><br/>Everything else remains open. There is no Google announcement, no DeepMind model card, and no entry in the Gemini API changelog — as of its last update, Google lists only Nano Banana 2 (gemini-3.1-flash-image), Nano Banana 2 Lite (gemini-3.1-flash-lite-image) and Nano Banana Pro (gemini-3-pro-image). There is no price, no resolution table, and no benchmark for 2.1.<br/><br/>The screenshot below is the clearest single piece of evidence for that gap. It is the Artificial Analysis Text-to-Image v2.0 leaderboard, an independent evaluator. Nano Banana 2, Nano Banana Pro and Nano Banana 2 Lite each hold a ranked position with an Elo score, a sample count and a per-image price. There is no Nano Banana 2.1 row. If a model existed in the wild with measurable output quality, it would appear on a board that tracks exactly that. Screenshot sourced from Artificial Analysis.
Seen in community screenshots from October 5, 2026. Reachable, but undocumented.
No API identifier, no price, no resolution tiers, no benchmark, no vendor evaluation.
TestingCatalog observed early outputs resembling Nano Banana Pro and said it could not establish whether requests were routed to a genuinely new model. A menu entry is a label; routing happens server-side.
A Google Flow web build carried the string "Nano Banana 2.5 Flash" on September 24, 2026. By September 27 the same build relabelled it 2.1. By October 5 it was a clickable menu item.
The workspace sits on top of an image generator and exposes the controls that generator actually offers. Text-to-image and image-to-image are both supported, so you can start from a description or bring a reference image and describe the change.<br/><br/>Reference images are the useful part for consistency work. Upload a face, a product or a layout, then refer to them in your prompt so the model knows what must stay fixed. This is what makes a character or a product recognisable across a series rather than reinvented每 time.<br/><br/>Aspect ratio and resolution are both selectable from the generator. Decide the destination before you generate — 9:16 for stories, 1:1 for feed posts, 16:9 for covers — because the ratio is much cheaper to set correctly before generation than to fix afterwards.
Turn a plain-language prompt into a finished visual — posters, thumbnails, product shots, storyboards.
Bring an existing picture and describe the change: swap the background, adjust lighting, remove an object, restyle the whole scene.
Keep characters, products and visual style recognisable across a series instead of reinventing them each time.
Frame for the destination — vertical stories, square feed posts, wide covers — without recomposing later.
Choose what the current model offers. Raise it when the image will be seen large, not by default.
Follow up with short, single-variable edits instead of rewriting the whole prompt each time.
Put words in quotes and state where they belong — useful for posters, labels and packaging concepts.
Take the finished file when it looks right, at the ratio and size you chose.
Four steps, and the order matters more than it looks. Deciding ratio and references before you generate is cheaper than fixing them afterwards.<br/><br/>Step three is the one people skip, and it is the one that changes results most. Aspect ratio is baked into the composition — a subject framed for 16:9 does not survive being cropped to 9:16 without losing either the edges or the vertical emphasis you wanted. Same with resolution: raising it costs more, and the extra detail is invisible unless the image is actually seen large. Deciding both up front means you generate once instead of three times.<br/><br/>Step four assumes your first result will be close but not right. That is the normal case, not a failure. The mistake is rewriting the whole prompt in response — that discards whatever was already working and starts the search over. Change one variable, look, change the next.<br/><br/>Worth knowing before you start: this model is not documented, so there is no published resolution table or reference limit to design around. What you can rely on is what the generator itself offers at the moment you open it.
Subject, setting, lighting, composition, style. Put any text that must appear in quotes and say where it goes. Full templates are in the prompt section below.
A face that has to recur across panels, a product whose shape and packaging matter, a layout to match. More references are not always better — each one narrows the solution space, and unrelated details constrain the result in ways you did not intend.
9:16 for stories and reels, 1:1 for feed posts, 16:9 for covers and banners, 4:5 or 2:3 for product listings. Raise resolution only for print or large-format display — it is invisible on a phone and costs more.
If the result is close but wrong, change one thing per follow-up. Several simultaneous changes make it impossible to tell which one caused the difference, and you end up regenerating instead of converging.
Three documented models and one undocumented label. The comparison below is by evidence status, because for 2.1 that is the only honest axis available.
Reusable skeletons. Replace the braced variables with your own details.<br/><br/><strong>Standard template</strong> — {subject} in {setting}, {lighting}, {composition}. {style}. Mood: {mood}. If text must appear: "{exact wording}" placed at {location}.<br/><br/><strong>Text-free template</strong> — use when spelling must be exactly right: {subject} in {setting}, {lighting}, {composition}. {style}. Leave negative space at {location} for a headline. No text, no letters, no typography. Text rendering varies with phrasing and prompt length, so adding type afterwards is the more reliable production path.<br/><br/><strong>Consistency template</strong> — for series work: Keep the face, hair, clothing and color palette of Image 1 unchanged. Change only: {the single thing you are editing}. Using the same reference across every image in a set is what carries a character or product through a sequence.<br/><br/>Field Note (granularity demo, not a project claim): if exact spelling matters, generate text-free and set type in post-production rather than regenerating until the letters land. This is a workflow habit, not a measured finding about 2.1 — no independent test data exists.
Wrap text that must appear in quotes and state its position — "SALE" at the top of the poster. Removes the ambiguity about whether letters belong in the frame at all.
Refer to uploads as Image 1, Image 2 in the prompt so you can state what goes where instead of leaving the model to guess.
Altering a single element per follow-up keeps the causal chain readable. Several simultaneous changes hide which one moved the result.
9:16 stories and reels, 1:1 feed posts, 16:9 covers, 4:5 or 2:3 product listings. Set it before generating, not after.
Higher resolution is invisible on a phone and costs more. Worth it for print, large-format display, or when you need room to crop.
If the exact words matter, compose without lettering and set type afterwards. The most reliable path, and what most production teams do.
Most of these come down to one of three things: a partial rollout, a routing decision you cannot see, or a prompt asking for too much at once.
Rollout is partial and account-specific, so availability differs between accounts and regions. There is no waitlist and no setting to change. Check again later. This is the single most common reason people conclude 2.1 does not exist — it exists, it is just not everywhere.
A menu entry is only a label. Routing is decided server-side and nothing on screen reveals it — this is why the community could not agree on whether 2.1 is a new model or Nano Banana Pro under a new name. The practical fix: log your prompt, settings and date for every generation so comparisons are reproducible, and treat any A/B difference as provisional until a model card exists.
Usually long or indirect phrasing rather than a model fault. Quote the exact wording, state where it sits, keep the sentence short. If it still fails, generate text-free and add type in post — that path is deterministic in a way generation is not.
Almost always a different reference set per image. Reuse one reference for the character or product across the whole set, and change only the single element you are editing.
Multiple constraints in one sentence compete for attention. Split them across follow-ups: change one thing, look at the result, change the next. A prompt rewritten from scratch each time will never converge.
The request exceeds what the current model accepts. Check what the generator currently offers, step down to the nearest supported value, and add references one at a time to find the actual ceiling.
Grouped by what you are trying to find out. Every answer separates what is confirmed from what is reported, because for this model that distinction is the whole answer.
Not as far as public documentation goes. The label has been reported as selectable in Google Flow's image picker as of October 5, 2026, but Google has published no announcement, no DeepMind model card, and no Gemini API entry for a 2.1 model. Treat it as a reported rollout, not a documented release.
The label is real in the sense that it is reachable in a shipping Google product and appears in multiple independent screenshots. What is not established is what the label points to. It is a genuine sighting with no accompanying documentation.
Third-party trackers including TestingCatalog reported the model inside Google Flow, and screenshots of the picker listing Nano Banana Pro, Nano Banana 2 Lite and Nano Banana 2.1 circulated publicly on October 5, 2026. Earlier sightings trace back to a Google Flow web build where a reference string first read Nano Banana 2.5 Flash before being relabelled 2.1.
No. Google's developer surface currently exposes gemini-3.1-flash-image (Nano Banana 2), gemini-3.1-flash-lite-image (Nano Banana 2 Lite), and gemini-3-pro-image (Nano Banana Pro). There is no 2.1 endpoint to call. Independent probes of five plausible 2.1 and 2.5 route names returned 404 on all of them.
Unknown, and any number you see quoted is a guess. One Flow account showed a note that generating with 2.1 would consume zero credits, but that describes that account's plan, not pricing. Google has published no per-image rate for 2.1.
Unconfirmed. Resolution tiers up to 4K are documented for Nano Banana 2 (gemini-3.1-flash-image). No equivalent documentation exists for 2.1, and pages describing 2.1 as supporting 4K are repeating Nano Banana 2's specifications under a new name.
Nano Banana 2 is a documented public model with published specifications and independent benchmark scores. Nano Banana 2.1 is not, so a real quality comparison is not yet possible. The numbering suggests an iterative release rather than a new generation, but Google has confirmed nothing about scope or capabilities.
TestingCatalog observed early 2.1 outputs looking similar to Nano Banana Pro and noted it could not establish whether requests were routed to a genuinely new model. That is a visual impression from a handful of outputs, not a routing trace.
Google has published no terms for 2.1, so there is nothing to review. If you need commercial certainty today, build on the documented Nano Banana 2 family endpoints, whose terms are published.
Social posts, posters, thumbnails, product shots, storyboards, infographics, and concept art. What the underlying model produces is not officially documented, so this workspace makes no specific performance promises on its behalf.
None. Describe what you want in plain language and the workspace handles composition, lighting, and detail. If exact spelling or exact brand colours matter, plan to set type in post rather than relying on the model.
Yes. Upload a picture and describe the change — swap backgrounds, adjust lighting, remove objects, or restyle the whole scene. Keep one reference constant across a series when the subject must stay recognisable.
Official Google sources first, then third-party reports, oldest to newest. Last reviewed October 6, 2026.<br/><br/>The independent leaderboard screenshot in the evidence section is from Artificial Analysis. Third-party reports are cited as reports, not as confirmations — they document what was seen in Flow, not what Google has committed to shipping.
The Nano Banana 2 post on blog.google, and Gemini image model documentation at ai.google.dev listing gemini-3.1-flash-image, gemini-3.1-flash-lite-image and gemini-3-pro-image. The API changelog was last updated October 1, 2026 and mentions only those three.
Reported that a Google Flow web build had relabelled a Nano Banana 2.5 Flash reference to Nano Banana 2.1, and noted the change pointed to an iterative release rather than a new generation.
Images of the Flow model picker listing Nano Banana Pro, Nano Banana 2 Lite and Nano Banana 2.1 circulated, with 2.1 selected. One account noted that generating with it would consume 0 credits — a property of that account's plan, not a published price.
Added that early 2.1 outputs looked like Nano Banana Pro output, so it could not establish whether requests were routed to a genuinely new model. This is the origin of the unresolved routing question.
SandBase probed five plausible 2.1 and 2.5 API routes and received 404 on all of them. This page was last reviewed the same day, and the leaderboard screenshot above was captured the same day.
Describe your first idea and see it come to life in seconds.