Google releases Nano Banana 2.1 with 4K image editing

• Google has shipped Nano Banana 2.1 as the update to Nano Banana 2, with API, AI Studio, and Gemini app access called out in OfficialLoganK's launch post.
• The work centers on better prompt adherence, text, subject consistency, mask-based editing, and wide-format reliability, according to MayorKingAI's launch note, while the API supports up to 14 reference images.
• Image-output pricing is roughly half the previous model's at 1K, 2K, and 4K, but ozansihay's pricing note says input and thinking charges rose too.
• The model is spreading through Google's creative stack and third-party tools, with MayorKingAI's rollout list naming Google surfaces and higgsfield_ai's announcement confirming a 4K rollout on launch day.

Google's API model page gives the model a stable slug, gemini-nano-banana-2.1, and names Gemini 3.6 Flash underneath. A production pipeline's blind face test preferred 2.1 by 40 to 24, while techhalla's 20-prompt comparison put GPT Image 2.5 slightly ahead overall in one same-prompt test.

Google marked Nano Banana 2.1 generally available on October 6 in its Gemini API release notes. The release replaces the gemini-3.1-flash-image model ID used by Nano Banana 2 with gemini-nano-banana-2.1.

The concrete changes are:

• Resolution: 1K, 2K, and 4K output, with 1K as the default.
• Wide formats: 1:4, 4:1, 1:8, and 8:1 aspect ratios, with tiling artifacts fixed at 2K and 4K.
• References: Up to 14 reference images, including up to four characters and 10 objects.
• Text and layouts: Better text rendering and infographic layout accuracy.
• Thinking: Minimal, medium, and high levels, with medium as the default.
• Grounding: Google web and image search can inform generated visuals.

The 4K ceiling carries over from Nano Banana 2. The version bump is concentrated in consistency, instruction following, and fewer visual failures around layouts and panoramic frames.

The public workflow in Google's image-generation guide is conversational image-to-image editing. A creator sends an image with a text instruction to add, remove, or modify an element, then continues the conversation for more revisions.

One demonstrated workflow turns an uploaded photo into a split poster with the original subject preserved, a hand-drawn version below it, and tiny characters interacting with the scene, using icreatelife's doodle prompt as the recipe.

The mask promise sits alongside a caveat. Google's model card reports strong results on mask and ink editing, but also lists partial instruction following and ink persistence as known limitations. It separately flags blurry small text, imperfect character consistency, and occasional left-right confusion.

The independent ppl.studio price table reproduces Google's standard image-output rates like this:

• 1K: $0.034 for Nano Banana 2.1, down from $0.067 for Nano Banana 2.
• 2K: $0.050, down from $0.101.
• 4K: $0.076, down from $0.151.

Google's pricing documentation separates image output from input and text or thinking output. That makes the headline half-price claim accurate for the image component, while the total request can move with prompt size and thinking level. In the ppl.studio test, 2.1 took 17 seconds at minimal thinking, 24 seconds at the default medium level, and 2.0 took 15 seconds.

Google's model card lists the Gemini app, AI Studio, Gemini API, Search AI Mode, Google Ads, Flow, and Stitch. MayorKingAI's rollout list also names the Gemini Enterprise Platform in MayorKingAI's rollout list.

Third-party access is already broader than the Google stack:

• Higgsfield: Its launch post promises faster generation, more control over complex prompts, and output up to 4K in higgsfield_ai's announcement.
• Pika: Nano Banana 2.1 is live in Pika's creative tooling and API playground, according to pika_labs' announcement.
• ElevenCreative: AIwithSynthia's update puts generation and editing in the same place.
• Magnific: The platform says it added 2.1 alongside its image and video stack in magnific's announcement.
• Figma Weave: figmaweave's offer advertised 50% off credits for a limited period.

Techhalla ran the same prompts through Nano Banana 2.1 and GPT Image 2.5 on Magnific, scoring prompt adherence, aesthetics, and composition. The 20-prompt thread found that neither model fumbled the text in its text, diagram, and still-life round, but it put GPT slightly ahead on aesthetics.

The split was prompt-dependent. Nano Banana 2.1 won a flat-background composition, while GPT Image 2.5 showed more detail in the photography rounds, according to techhalla's follow-up. A separate Istanbul street-food comparison used the same scene and signage across four models, including Nano Banana 2.1, GPT Image 2.5, Grok Imagine 2.0, and Ideogram 4.5, in ozansihay's same-prompt test.

Google's model card reports higher human-preference Elo for 2.1 than its listed predecessors: 1,050 versus 990 overall text-to-image, 1,048 versus 961 for infographic design, and 1,106 versus 978 for multi-character consistency. Those are Google's own side-by-side evaluations, not a cross-vendor benchmark.

The production test from ppl.studio was narrower but operationally specific. Two AI judges preferred 2.1's reference face in 40 of 64 comparisons, and its exact-text test came out 9 of 9 versus 7 of 9 for Nano Banana 2. The authors describe the result as a small lean rather than a landslide.

The old API model is on a clock. Google's release notes say gemini-3.1-flash-image will shut down on October 29, 2026, making gemini-nano-banana-2.1 the migration target.

• The 512px option is gone. An independent side-by-side developer test found only 1K, 2K, and 4K outputs in 2.1.
• The default thinking level changed from minimal to medium, which made 2.1 slower in that test. Minimal and high remain available.
• The model card says small text can still blur, long paragraphs remain difficult, character identity can drift, and masked edits can preserve unwanted structure or pose.

The billing mix changes with the slug too: ozansihay's pricing note describes lower image-output charges alongside higher input and thinking charges.

Brand continuity is one of the clearest creative targets. LukeW tested Nano Banana 2.1 in a consistent brand-assets maker, while LukeW's test also shows a separate before-and-after visual review workflow for judging UI changes.

The reference-image workflow also travels into stylized work. ProperPrompter's pixel-art prompt keeps a Pokémon battle interface and character language while moving the fight to a new location with new Kanto Pokémon.

Other tests push the model toward production-like constraints:

• A night photograph of an Istanbul food stall keeps the Turkish sign, the cook, the waiter, and the Bosphorus setting across variations in ozansihay's same-prompt test.
• A Flow test asks for a reverse camera angle while preserving the living-room layout in ai_artworkgen's Flow test.
• A close-up skin test reports progress in texture and realism through icreatelife's test.

One tutorial uses Nano Banana as part of a larger product workflow. The app teaches Japanese through live voice calls, uses a spec skill to create key designs, connects Gemini Live APIs for voice, and generates diorama art with Nano Banana. The tutorial targets 100 travel phrases before a trip, with the model supplying the visual world around the lessons.