Skip to main content

​
Overview

Nano Banana 2 (codename) is Google’s latest image generation model released on February 26, 2026, with model ID gemini-3.1-flash-image-preview. It redefines image generation cost-effectiveness with Pro-level quality at Flash-tier speed and cost, making it the latest flagship of the Nano Banana series.
🔥 Released February 26, 2026: Nano Banana 2 is live! Pro-level quality, Flash-tier speed, token-based billing as low as 36% of Google’s pricing! 512px from just $0.025/image! Supports 4K output, 14 aspect ratios, Image Search Grounding, and more exclusive features.
🆕 Update May 29, 2026 (-preview dropped): Google updated its official docs and released the stable model name gemini-3.1-flash-image (without -preview). APIYI already supports it.
  • The old name still works: gemini-3.1-flash-image-preview keeps working as usual, pricing unchanged, no code changes needed.
  • Both names work: use either the new gemini-3.1-flash-image or the original -preview name.
Heads-up: Google hasn’t clarified whether the stable version differs from the preview in output quality, safety filtering, or other behavior. We welcome you to test and share feedback.
All image APIs are synchronous — there is no task ID to poll, and if your client disconnects the result is lost while the request is still billed. Set a generous timeout for this model; see Image API Essentials & Best Practices.

Text-to-Image API

Generate images from text prompts. Includes an interactive playground for online testing.

Image Editing API

Upload an image + edit instructions to generate new images. Includes an interactive playground.

​
Let an AI Agent Do the Integration

If you build with Codex / Claude Code / Cursor, copy the prompt below and hand it to your agent. It first fetches the plain-text version of this page (append .md to any docs URL), then writes code in your project’s own stack — timeout, defensive parts parsing, upload compression and the resolution parameters are already baked into the requirements.

Have a coding agent integrate or troubleshoot Nano Banana 2 text-to-image and image editing. Copy and paste into Codex, Claude Code, Cursor and similar tools.

​
Why APIYI’s Nano Banana 2?

Nano Banana Pro / 2 is the #1 model by usage volume on APIYI — stable, reliable, and fast. If you want to work with a professional team, APIYI is the right choice. A freshly released Google flagship still capacity-constrained at the source — APIYI deeply optimizes the experience across reliability, cost, and integration:

Official Channel · Same as Gemini

100% compatible with Google’s native Gemini API (/v1beta/models/.../generateContent) and the OpenAI SDK pattern — same request body, response fields, and error codes. Zero-code migration.

No Concurrency Limits

Not bound by Google AI Studio’s RPM/RPD ceilings. Enterprise-scale batch generation and peak traffic scale linearly without quota interruptions.

28-36% of Google's List Price

$0.055/image per-call (vs Google $0.151), 512px from $0.025/image token-based (vs $0.045). Stack with top-up bonuses for as low as 30.3% of the list.

Global Zero-Barrier Access

No overseas server or proxy required — connect directly to api.apiyi.com from mainland data centers, residential networks, or overseas nodes. Stable latency, no cross-border re-architecture.

Full Model Lineup

Same series covers Nano Banana Pro (ultimate quality), Nano Banana 2 (best value), and the legacy Nano Banana — mix and match per scenario.

Professional Enterprise Support

Our team specializes in production image-generation deployments, with deep experience in model selection, tuning, and integration — end-to-end support from PoC to production.

​
Core Features

Pro-Level Quality

Vibrant lighting, rich textures, sharp details - quality rivaling Nano Banana Pro at much faster speeds

4K Ultra-HD Output

Supports 512px, 1K, 2K, 4K resolutions, up to 4096×4096

14 Aspect Ratios

New additions: 1:4, 4:1, 1:8, 8:1 - total 14 aspect ratios covering more use cases

Image Search Grounding

Nano Banana 2 exclusive - pulls visual context from Google Image Search

Accurate Text Rendering

Clear, legible text in images with multilingual support - perfect for posters and marketing materials

Multi-turn Editing

Conversational image editing with iterative refinement through chat

Thinking Mode

Configurable minimal or high thinking levels for more precise complex prompt handling

Subject Consistency

Maintains resemblance across up to 5 characters and 14 reference objects

​
Version Comparison

Selection Guide:
  • 🔥 Best Value → Nano Banana 2 (token-based from $0.025/image, Pro-level quality + Flash-tier speed)
  • 🎨 Ultimate Quality → Nano Banana Pro ($0.09/image, highest fidelity)
  • ⚡ Lowest Cost → Nano Banana ($0.025/image, fast and stable)

​
Pricing

Billing Mode Selection: Nano Banana 2 supports two billing modes, selected via the “Billing model” setting when creating your API token:
  • Select Pay-as-you-go or Pay-as-you-go Priority → Token-based billing
  • Select Pay-per-request or Pay-per-request Priority → Per-call billing (same as Nano Banana Pro)
  • ⚠️ Do NOT select Hybrid billing

​
Per-call Billing

Enterprise HA Channel: NanoBananaEnterprise is available at 1.4x the standard Nano Banana Pro rate ($0.126/image), offering a dedicated high-availability fallback for enterprise workloads.

​
Token-based Billing (Nano Banana 2 Exclusive)

​
Token-based Billing Price Estimates

💰 Token-based billing saves more! With token-based billing, 512px images from just $0.025 — only 36% of Google’s pricing! For low-resolution use cases, this is much cheaper than per-call billing ($0.055/image). For 4K, token-based pricing (~$0.07) is even cheaper than per-call billing, and both are far below Google’s official $0.151/image. Combined with top-up bonuses, actual costs are even lower.

​
Three parameters that affect billing

The figures below come from live testing against the production gateway on 2026-08-27 (12–20 runs per condition, all based on actual charges recorded in the console rather than estimates). Their impact on your bill differs by two orders of magnitude, so optimise in the order shown.

​
thinkingLevel: the default already is minimal, and high costs half as much again

gemini-3.1-flash-image supports thinking-level control (Pro does not). Same prompt, 20 runs each:
  • The default is identical to minimal — statistically indistinguishable (p=0.64), so passing minimal explicitly gains nothing.
  • Only at high does thoughtsTokenCount appear as its own field, and it is folded into completion_tokens and billed at the image rate.
  • Complex prompts only cost more once high is on: at the default level, a 3-token prompt and a 98-token reasoning prompt produced no significant difference in output tokens (p=0.37); with high, the reasoning prompt used 126% more thinking tokens and cost 38% more per call.
Leave this parameter alone for everyday generation. Enable high only when you have hard requirements on composition logic, in-image text layout, or chart proportion accuracy — at the cost of +54% spend and +55% latency.

​
Google Search grounding: works, but billed per search query

For images that need live information to be correct (weather cards, market charts, event posters), attach the googleSearch tool. Grounding triggered in 12/12 runs.
Output tokens are barely affected (−2.1% versus the no-tool control once the search fee is removed, not significant); the entire cost comes from the search calls themselves at $0.014 each.
The model decides how many searches to run; you cannot set it in advance. In testing a single image request issued 1–3 queries on its own, so with this tool enabled the per-call cost is a range of $0.050–$0.078, not a fixed number. Budget against the upper bound.

​
Image Search grounding: exclusive to Nano Banana 2, and currently free

searchTypes.imageSearch lets the model pull visual references from Google Image Search — useful for “make a collage from real photos” or “draw this from the actual object”.
  • Confirmed by imageSearchQueries appearing in groundingMetadata, with values like ["current weather in Tokyo"].
  • No extra charge at present — the per-call cost even came in slightly below the no-tool control.
  • It triggered in 7 of 12 runs: the model decides whether to search, it does not query every time.
  • Web search and image search can be enabled together, but the model chooses which to use, so the cost becomes unpredictable.
searchTypes must be an object, not an array.✅ {"searchTypes": {"imageSearch": {}}} — works❌ {"searchTypes": ["imageSearch"]} — no error, returns 200, and you still get an image, but image search never triggers once. Nothing in the response body reveals it. The upstream error for the array form is Proto field is not repeating, cannot start list.

​
responseModalities: saves nothing, but removes unwanted text

responseModalities: ["IMAGE"] declares that you only want images. 20 runs each:
  • All three are within 1% of each other, and ["TEXT","IMAGE"] is exactly equivalent to omitting the field (p=0.90).
  • With ordinary generation prompts, 60/60 responses already contained a single image part and no text at all, so the modality switch has nothing to act on. Only reasoning prompts (infographics, data charts) attach a summary paragraph roughly a third of the time; there ["IMAGE"] suppresses it and saves about 2.8% (p=0.09, not significant).
Why there is nothing to save: output tokens are not “image + text” but 1120 for the image plus roughly 400 tokens of invisible overhead. That overhead appears in no field of candidatesTokensDetails, yet it is billed at full rate and accounts for 28% of the per-call charge, and it does not vary with prompt length. The text is negligible next to it.So estimating per-image cost from 1120 tokens understates it by about 28% — reconcile against candidatesTokenCount or totalTokenCount instead.

​
Group Setup

Nano Banana 2 ships with two groups on APIYI. Switch in dashboard → Token Settings: Why 1.4x? Even at 1.4x, the price is still around 50% of Google’s list — far below official pricing. This is a fallback lane for higher-concurrency workloads and unexpected upstream risk-control events, providing high-availability guarantees for enterprise customers. When the default group is tight, switch your Token to NanoBananaEnterprise to ride out the spike. Recommended Billing model: pick Pay-as-you-go Priority — covers Nano Banana 2’s token-based billing AND Nano Banana Pro’s per-call billing, one Token for the whole series.
Token creation UI: Billing model 'Pay-as-you-go Priority' covers NB2 token-based + NB Pro per-call; primary group Default + fallback group NanoBananaEnterprise (1.4x lane)

Token settings: Billing model = Pay-as-you-go Priority, primary group = Default, fallback group = NanoBananaEnterprise (1.4x)

Going further: if your Token also covers other image models (e.g. GPT-image-2), keep the more stable Default as the primary group and put NanoBananaEnterprise in the fallback slot — 429s on the primary will auto-failover to the enterprise group without a token swap.

​
Supported Resolutions & Aspect Ratios

​
Output Resolutions

​
Supported Aspect Ratios (14 total)

1:1, 1:4, 4:1, 1:8, 8:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9

​
Output Dimensions per Aspect Ratio (pixels)

The table below lists Nano Banana 2’s actual output dimensions for all 14 aspect ratios across the 512px / 1K / 2K / 4K resolution tiers (source: Google official docs). In your request, set aspect_ratio for the ratio and image_size (or resolution) for the tier:
Compared to Nano Banana Pro’s 10 aspect ratios, Nano Banana 2 adds 1:4, 4:1, 1:8, 8:1 - ultra-tall/ultra-wide ratios ideal for long-form images and infographics. It is also the only model with a 512px low-resolution tier, great for thumbnails and quick previews.

​
FAQ

Nano Banana 2 (gemini-3.1-flash-image-preview) is based on Gemini 3.1 Flash, while Nano Banana Pro (gemini-3-pro-image-preview) is based on Gemini 3 Pro. Key differences:
  • ✅ Speed: Nano Banana 2 is faster (Flash-tier speed)
  • ✅ Price: Nano Banana 2 token-based billing is cheaper (from $0.025 vs $0.09)
  • ✅ Aspect Ratios: Nano Banana 2 supports 14 (4 more)
  • ✅ Image Search Grounding: Nano Banana 2 exclusive
  • ⚠️ Ultimate Quality: Nano Banana Pro still slightly better
Recommended! Nano Banana 2 offers near-Pro-level quality at a lower price with faster speed. Unless you have extreme quality requirements, Nano Banana 2 is the better choice.Simply change the model name from gemini-3-pro-image-preview to gemini-3.1-flash-image-preview.
For code / API calls, we recommend always using the general model name gemini-3.1-flash-image-preview (without the -4k suffix), not gemini-3.1-flash-image-preview-4k.
  • The official name has no -4k: Google’s official model name is gemini-3.1-flash-image-preview. This is also the general channel we invest the most resources in maintaining, so it offers the best stability and compatibility.
  • Where -4k comes from: gemini-3.1-flash-image-preview-4k was originally a configuration prepared for “chat-to-image” scenarios in conversational clients like Chatbox—generating images directly through chat in a messaging UI.
  • Code + Gemini native format favors the general name: If you’re calling via code using the Gemini native format (/v1beta/models/.../generateContent), the regular model name gemini-3.1-flash-image-preview is more stable.
When you need 4K output, there’s no need to rely on the -4k model name—just specify the 4K resolution in your request parameters (see “Supported Resolutions and Aspect Ratios” above).
Image Search Grounding is a Nano Banana 2 exclusive feature. It pulls visual context from Google Image Search to generate images that better match real-world subjects. For example, when generating images of real landmarks, it can reference search results for improved accuracy.
Thinking Mode allows the model to reason and analyze before generating images, improving accuracy for complex prompts. Setting it to high produces the best results but slightly increases generation time. Best for tasks requiring precise composition, text rendering, or complex scenes.
Generation time depends on resolution and thinking mode:
  • 1K resolution: ~5-10 seconds
  • 2K resolution: ~10-15 seconds
  • 4K resolution: ~15-25 seconds
  • Enabling high thinking mode adds a few extra seconds
Recommend setting a longer timeout (at least 360 seconds) for occasional delays and peak congestion.
The API has no concurrency limit and does not process serially. You can safely fire concurrent requests yourself — requests don’t queue or block each other. For 20 users calling simultaneously, just issue 20 concurrent requests; no extra quota request or throttling needed.Unlike Google AI Studio, the APIYI channel has no hard RPM/RPD limits, so enterprise batch generation and peak traffic scale linearly.What actually matters is the timeout: image generation (especially 4K or during peak congestion) can take a while per request, so set your client timeout to 360 seconds to avoid cutting off requests that are still being processed normally.
If you occasionally hit 429 (throttled due to high concurrency), add NanoBananaEnterprise as the fallback group on your token (see “Group Setup” above). When the primary group is saturated it auto-falls back, further improving success rates under high concurrency.
Supports image/png and image/jpeg formats. Can be uploaded via base64 encoding or Files API.
All output images carry SynthID invisible digital watermarks (Google’s AI-generated content identification technology) - invisible to the naked eye and doesn’t affect usage.
The full error looks like:
This is usually caused by oversized image uploads — the request body gets too large and the connection collapses. Follow these best practices:
  • Limit the image count: stay within the official rules (max 14 images per prompt) — don’t pile on reference images.
  • Limit per-image size: keep each image under 5MB — the official per-image cap is 7MB, and base64 encoding inflates size by roughly 1/3, so leave headroom.
  • Compress on the frontend before uploading: compress images on the frontend (or a server-side relay) before sending them to the API — common practice is capping the longest edge, converting to JPEG/WebP, and tuning the quality parameter.
  • Switch to URL input: the Gemini native format supports passing an image URL via fileData.fileUri, sidestepping oversized base64 request bodies entirely — see the Nano Banana Dev Guide.
Nano Banana 2 is currently in Preview status. Features and pricing may be adjusted. Please follow documentation updates for the latest information.