Nano Banana 2

Nano Banana 2 is the public nickname for Gemini 3.1 Flash Image, an image generation and editing model released by Google DeepMind on 26 February 2026 ^[1][2]. It is the third model in Google's "Nano Banana" line and was built to bring the visual quality of the earlier Nano Banana Pro to the faster, cheaper Flash tier of the Gemini family ^[1][3]. On launch day Google made it the default image engine across the Gemini app, Google Search's AI Mode, Google Lens, the Flow filmmaking tool, and Google Ads, where it replaced Nano Banana Pro for most everyday image tasks ^[1][4].

Overview

Nano Banana 2 sits between the original Nano Banana, which was Gemini 2.5 Flash Image, and Nano Banana Pro, which was Gemini 3 Pro Image ^[4][5]. The pitch is straightforward: roughly the output quality people had been getting from the Pro model, but generated about four times faster and at close to half the price per image ^[3][6]. Google positions it as the model most users should reach for by default, reserving the slower, more deliberate Pro tier for jobs that need the very highest fidelity ^[1][4].

Like the rest of the line, the model handles both text-to-image generation and conversational editing of existing images. It draws on the broader Gemini model's world knowledge and can pull in real-time information and images from web search, which Google says lets it render specific real-world subjects more accurately and turn rough notes into infographics or diagrams ^[1][7]. Every image it produces carries an invisible SynthID watermark and is interoperable with C2PA Content Credentials so that the output can be identified as AI generated ^[1][2].

Naming and identity

The consumer-facing name "Nano Banana" began as an unofficial nickname that Google later adopted as a brand. The underlying technical model behind Nano Banana 2 is Gemini 3.1 Flash Image, and on developer platforms it is exposed under the model ID gemini-3.1-flash-image-preview ^[2][8]. The "Flash" in the name marks its place in the Gemini tier system: Flash models trade some of the depth of the Pro tier for much higher speed and lower cost ^[3][8].

The "3.1" reflects that the model belongs to the same point-release wave as Gemini 3.1 Pro, the language model Google shipped in February 2026 as an upgrade over Gemini 3 Pro ^[9]. The numbering can be confusing because the three Nano Banana releases do not map cleanly onto a single version sequence. The table below lays out the relationship between the marketing names and the technical model names.

Scroll sideways for more →

Release

Google announced Nano Banana 2 on 26 February 2026 through the Google blog and Google DeepMind, with simultaneous availability across consumer and developer surfaces ^[1][2]. The original Nano Banana had launched in August 2025 and drove users to generate millions of images in the Gemini app, and Nano Banana Pro followed in November 2025 with higher detail and quality ^[4][5]. Nano Banana 2 was the next step in that cadence and was framed as combining the speed of the first model with the capabilities of the Pro release ^[4][6].

A few months after the initial preview, on 29 May 2026, Google announced that both Nano Banana 2 and Nano Banana Pro had reached general availability on its enterprise platform, with 1K and 2K output generally available and 4K output remaining in preview ^[10]. That same update added a preview capability for Nano Banana 2 to accept video files as input for context-aware image generation ^[10].

Capabilities

Nano Banana 2 accepts text and images as input, with a context window of up to one million tokens, and produces images at resolutions from 512 pixels up to 4,096 by 4,096 pixels across a range of aspect ratios ^[3][8]. The 512-pixel tier was added for cheap, rapid iteration, and the release also introduced wide and tall aspect ratios such as 4:1, 1:4, 8:1, and 1:8 alongside the previously supported shapes ^[7].

Key capabilities reported by Google and reviewers include:

• Subject consistency. The model can hold character resemblance for up to five characters and object fidelity for up to fourteen objects within a single workflow, which is meant to support storyboarding and multi-image narratives without the inputs drifting in appearance ^[1][4].
• Text rendering. It renders legible text directly into images with control over fonts, styles, and sizes, and can translate in-image text across languages. Accuracy is strongest on short text in Latin scripts and degrades on long passages and non-Latin scripts ^[3][7].
• World knowledge. Because it is grounded in Gemini's knowledge and live web search, it can render specific real subjects, build infographics, and annotate diagrams more reliably than a model working from the prompt alone, though Google notes this knowledge is "extensive but not infallible" ^[1][3].
• Editing and upscaling. It performs conversational edits in under about twenty seconds and can upscale images to 2K and 4K and switch aspect ratios quickly ^[3][6].
• Configurable thinking. Developers can set thinking levels, trading latency for deliberation depending on whether the job needs a fast turnaround or more careful composition ^[7].

On the LMArena leaderboards used by reviewers, Nano Banana 2 led the text-to-image ranking at roughly 1,280 Elo, ahead of OpenAI's GPT Image 1.5, and placed near the top of the image-editing ranking ^[6]. Reported end-to-end generation time was on the order of four to six seconds per image ^[6].

How it differs from Nano Banana and Nano Banana Pro

Scroll sideways for more →

The most important practical change is that Nano Banana 2 took over as the default model in the Gemini app's Fast, Thinking, and Pro modes, displacing Nano Banana Pro for routine use, while Pro remained available for the highest-fidelity work ^[1][4]. The headline trade is quality parity with the Pro tier at roughly a quarter of the latency and about half the cost ^[3][6].

Availability and pricing

At launch Nano Banana 2 rolled out across both consumer products and developer platforms ^[1][2]. In Search it became the default for image results through Google Lens and AI Mode across 141 countries on the Google app and on the web ^[4].

Scroll sideways for more →

Developer pricing is metered by input tokens and by output image resolution. Reported rates were $0.50 per one million input tokens, with per-image output prices that scaled with resolution ^[6][8].

Scroll sideways for more →

At those rates the model came in at roughly half the per-image cost of Nano Banana Pro and noticeably cheaper than GPT Image 1.5 at comparable quality ^[6].

Reception

Coverage framed Nano Banana 2 as Google pushing its viral image generator further down the price-performance curve rather than as a radical new capability set ^[4][6]. TechCrunch and CNBC reported it as a faster, more precise update that produced more realistic images and followed instructions more closely than the original Nano Banana ^[4][6]. Reviewers at DeepLearning.AI highlighted that it topped the LMArena text-to-image ranking while costing less and running faster than rivals, calling the combination of quality, speed, and price the main story ^[6].

Google also cited early production users. A product manager at HubX reported a 74 to 76 percent reduction in latency, which they described as making image workflows about four times faster without giving up Pro-level quality, and an engineering co-founder at Emergent pointed to stronger multilingual prompt understanding and more legible in-image text ^[3].

Nano Banana 2 Lite release

Google expanded the family on 30 June 2026 with Nano Banana 2 Lite, formally Gemini 3.1 Flash-Lite Image. Its stable model ID is gemini-3.1-flash-lite-image. Google positions it as the efficiency specialist for rapid ideation, interactive applications, and high-volume pipelines, while Nano Banana 2 remains the generalist model balancing quality, intelligence, latency, and cost. TechCrunch independently described the release as a faster, cheaper companion aimed at high-volume work rather than a replacement for every Nano Banana 2 use case ^[11][18].

The model launched in Google AI Studio, the Gemini API, and Gemini Enterprise Agent Platform. Google also began rolling it out to AI Mode in Search, the Gemini app, NotebookLM, Google Photos, Stitch, Flow, and Google Ads. In the Gemini app it is exposed through a Flash-Lite mode rather than silently replacing Nano Banana 2 in every image workflow ^[11][17].

Google AI for Developers marks gemini-3.1-flash-lite-image as a stable model, and Google Cloud lists its launch stage as generally available with a global endpoint. Google Cloud supports Standard and Flex pay-as-you-go access, batch inference, and Provisioned Throughput. The Cloud surface also lists enterprise controls for data residency, customer-managed encryption keys, VPC Service Controls, and Access Transparency ^[12][15][16].

Capabilities and service limits

On the Gemini Developer API, Nano Banana 2 Lite accepts text and images and returns text and images. It supports text-to-image generation, conversational image editing, 14 aspect ratios, Batch API, function calling, and minimal or high thinking levels. The current endpoint limit is 65,536 input tokens and 4,096 output tokens, and image output is restricted to 1K resolution. The same page marks caching, code execution, file search, Search grounding, Maps grounding, Live API, structured output, and URL context unsupported ^[12].

The Google Cloud surface adds several platform-specific details. It accepts up to 14 input images, supports multi-turn editing and interleaved image and text responses, and can sample a video at one frame per second as context for image generation. Audio inside a video file is not used. Its documentation does not support grounding, function calling, code execution, tuning, or chat completions, so tool support differs between the Gemini Developer API and Google Cloud interfaces ^[16].

The official model card describes a broader underlying model that can comprehend text, images, audio, and video with a context of up to one million tokens. Those are model-level properties, not the limits exposed by every endpoint. For deployed applications, the current Gemini API and Google Cloud pages specify the lower 65,536-token input cap and do not list direct audio input ^[12][14][16].

Google reported four-second text-to-image generation at launch, while its current Gemini API page describes a target of under two seconds end to end. These figures are not latency guarantees. The model card notes that occasional slowness and timeouts can still occur ^[11][12][14].

Comparison with Nano Banana 2

Scroll sideways for more →

Lite is therefore about half the price of Nano Banana 2 at 1K and half the standard input-token price, but it gives up higher-resolution output and real-time Google Search grounding. Google recommends Lite as the upgrade from the original gemini-2.5-flash-image, not as the universal replacement for Nano Banana 2 or Nano Banana Pro ^[11][12][13].

The distinction also appears in Google's own evaluation. In side-by-side human evaluation, the thinking configuration of Lite scored 1,059 +/- 7 on general text-to-image work, compared with 1,080 +/- 6 for Nano Banana 2. On general editing it scored 983 +/- 9 versus 1,062 +/- 8, and on text editing it scored 961 +/- 10 versus 1,107 +/- 10 ^[14].

These scores are vendor-reported Elo estimates from Google's curated evaluation sets, not independent leaderboard results. They support the claimed efficiency tradeoff, but they do not support describing Lite as equal to Nano Banana 2 in every quality category ^[14].

Pricing

Nano Banana 2 Lite has no listed free Gemini API tier. Standard paid use costs $0.25 per million text, image, or video input tokens, $1.50 per million text and thinking output tokens, and $30 per million image-output tokens. A 1K image consumes 1,120 output tokens, for an equivalent price of $0.0336 ^[13].

Batch prices are half the standard rates: $0.125 per million input tokens, $0.75 per million text and thinking output tokens, and $15 per million image-output tokens. Google gives the equivalent 1K batch image price as $0.0168. The pricing table states that paid-tier data is not used to improve Google's products ^[13].

Safety, provenance, and limitations

Google says training and deployment controls included dataset filtering, data labeling, supervised fine-tuning, reinforcement learning from human and critic feedback, red teaming, trust-assurance review, and product-level safety filters. The model card lists policy testing for child sexual exploitation, hate speech, dangerous content, harassment, sexually explicit content, and medical misinformation. Its frontier-safety conclusion is indirect: Google says the image model is less capable than Gemini 3.1 Pro, which did not reach the company's Critical Capability Levels ^[14][17].

Images receive an always-on invisible SynthID watermark, and Google Cloud says C2PA Content Credentials are enabled by default. These provenance signals identify Google-generated content when detected; they do not guarantee that an image is accurate, harmless, or free of third-party rights concerns ^[12][15].

Google documents weaknesses in small or lengthy rendered text, character consistency, masked and doodle editing, spatial directions, image blending, 3D reasoning, and factual accuracy. The model may copy part of an input image during an edit, confuse left and right, or produce visual artifacts. Its knowledge cutoff is January 2025, and it lacks Search grounding, so claims embedded in infographics or location-specific scenes require separate verification ^[12][14][17].

Limitations

Google and reviewers flagged several caveats. In-image text rendering remains reliable mainly for short strings in Latin scripts; error rates rise sharply for long passages and for scripts such as Arabic and Hindi ^[3]. Because the model leans on real-world knowledge and web search for data-heavy outputs like infographics, Google warns it can misinterpret information or produce factually incorrect results, so the knowledge grounding is helpful but not authoritative ^[3]. At launch the highest 4K resolution was offered only in preview on the enterprise platform even after the lower tiers reached general availability ^[10]. As with the rest of the line, every output carries a SynthID watermark, which is a safety measure rather than a limitation but does mean images are designed to be detectable as AI generated ^[1][2].

References

1. ^Nano Banana 2: Google's latest AI image generation model. Google. 26 February 2026.
2. ^Gemini 3.1 Flash Image (Nano Banana 2). Google DeepMind.
3. ^Gemini 3.1 Flash Image (Nano Banana 2) model page. Google DeepMind.
4. ^Google launches Nano Banana 2 model with faster image generation. TechCrunch. 26 February 2026.
5. ^Google launches Nano Banana 2, updating its viral AI image generator. CNBC. 26 February 2026.
6. ^Nano Banana 2, aka Gemini 3.1 Flash Image, Makes Edits Easier and Faster. DeepLearning.AI, The Batch.
7. ^Build with Nano Banana 2, our best image generation and editing model. Google. 26 February 2026.
8. ^Google: Nano Banana 2 (Gemini 3.1 Flash Image Preview). OpenRouter.
9. ^Gemini 3.1 Flash Image (Nano Banana 2). Google AI Studio.
10. ^Nano Banana 2 and Nano Banana Pro are generally available. Google Cloud Blog. 29 May 2026.
11. ^Start building with Nano Banana 2 Lite and Gemini Omni Flash. Google. 30 June 2026.
12. ^Gemini 3.1 Flash Lite Image. Google AI for Developers.
13. ^Gemini Developer API pricing. Google AI for Developers.
14. ^Gemini 3.1 Flash-Lite Image Model Card. Google DeepMind. 30 June 2026.
15. ^Nano Banana 2 Lite and Gemini Omni Flash available. Google Cloud Blog. 30 June 2026.
16. ^Gemini 3.1 Flash-Lite Image (Nano Banana 2 Lite). Google Cloud Documentation.
17. ^Gemini 3.1 Flash-Lite Image: Nano Banana 2 Lite. Google DeepMind.
18. ^Google introduces a faster, cheaper image generator with Nano Banana 2 Lite. TechCrunch. 30 June 2026.