Nano Banana 2 Lite

Nano Banana 2 Lite is a text-to-image generation and image editing model released by Google DeepMind on June 30, 2026. Its formal product name is Gemini 3.1 Flash-Lite Image and its Gemini API identifier is gemini-3.1-flash-lite-image. It is the smallest and cheapest member of the Nano Banana image model family, positioned below Nano Banana 2 and Nano Banana Pro, and Google explicitly presents it as the recommended migration target for developers still calling the original Nano Banana (gemini-2.5-flash-image).^[1][2]

Google's documentation describes the model as "our fastest and cheapest Gemini image model, engineered for velocity and scale where speed and cost are the primary operational constraints."^[3] The headline vendor claims are text-to-image generation in about four seconds and a list price of $0.0336 per 1K-resolution image, roughly half the per-image cost of Nano Banana 2 at the same resolution.^[1][4] The model was announced in the same blog post as Gemini Omni Flash, Google's conversational video generation and editing model, under the title "Start building with Nano Banana 2 Lite and Gemini Omni Flash," credited to product managers Alisa Fortin and Anish Nangia of Google DeepMind.^[1]

Naming

Google uses three overlapping labels for this model, which has caused persistent confusion in secondary coverage:

• Nano Banana 2 Lite is the consumer and marketing name, inherited from the community nickname that attached to the original Nano Banana model in 2025.
• Gemini 3.1 Flash-Lite Image is the formal model name used on the Google DeepMind model page and in Google Cloud documentation.^[2][5]
• gemini-3.1-flash-lite-image is the stable API model code used in the Gemini API and Google AI Studio.^[6]

Despite the "2 Lite" branding, the model is not a distilled version of Nano Banana 2 in any sense Google has documented. Google frames it as the successor to the first-generation Nano Banana rather than a cut-down Nano Banana 2, describing it as "our recommended replacement for developers currently using our first version of Nano Banana (gemini-2.5-flash-image)."^[1]

Specifications

Figures below are from Google's own model page and image generation guide.^[6][3]

Scroll sideways for more →

The single-resolution restriction is the sharpest break from the rest of the family. Nano Banana 2 accepts requests from 512px through 4K, and Nano Banana Pro renders up to 4K; Nano Banana 2 Lite renders only at 1K.^[3][7]

Family comparison

Scroll sideways for more →

Prices are Google's published standard-tier list prices for the Gemini API.^[4] Release dates are the dates of Google's own announcement posts.^[1][7][8][9]

Pricing

Google's pricing page lists Nano Banana 2 Lite at $0.25 per million input tokens (text, image or video), $1.50 per million output tokens for text and thinking, and $30.00 per million image output tokens, which it summarizes as "$0.0336 per 1K resolution image." Batch API pricing is exactly half of each figure, giving $0.0168 per 1K-resolution image.^[4]

At $30.00 per million image output tokens, the $0.0336 figure works out to 1,120 output tokens for a 1024x1024 image. Nano Banana 2's $0.067 per 1K image against its $60.00 per million rate implies the same 1,120-token count, so at 1K resolution the entire price gap between the two models comes from the token rate rather than the image size.^[4] The image output rate is identical to the original Nano Banana's $30.00 per million, but the newer model's smaller per-image token count (1,120 versus 1,290) makes it slightly cheaper per image than its predecessor ($0.0336 versus $0.039).^[4][8]

The "per 1,000 images" misreading

Google's announcement post compressed the figure to "$0.034 per 1K image," where "1K" refers to the output resolution.^[1] Several outlets and aggregators read "1K" as one thousand and reported the price as roughly three cents per thousand images, a figure about a thousand times too low. TestingCatalog, for example, published the price as "$0.034 per 1,000 images."^[10] Google's pricing page is unambiguous on the point, spelling out "per 1K resolution image," and the arithmetic against the per-million token rate confirms the per-image reading.^[4] Artificial Analysis, which tracks list prices independently, states the same rate the other way round as $33.60 per 1,000 images through the Gemini API and $16.80 through the Batch API.^[11]

Availability

At launch Google made the model available to developers in Google AI Studio, the Gemini API and the Gemini Enterprise Agent Platform, and said it was "rolling out today in Google consumer surfaces including AI Mode in Search, Gemini app and many other products."^[1] Google Cloud confirmed general availability on the Gemini Enterprise Agent Platform in a July 1, 2026 post, which also notes that "image generation offers the fastest latency" while "image editing may experience slightly higher response time."^[5] The Google DeepMind model page lists the Gemini app, Google AI Studio, the Gemini API and the Gemini Enterprise Agent Platform as its surfaces.^[2]

Google Cloud's suggested workloads for the model are high-volume and operational rather than creative: rapid-firing ideas, A/B testing ad variations, powering social apps for millions of users, storyboarding tools, virtual try-ons for ecommerce, and localized ad variations.^[5]

Provenance and watermarking

Every image the model produces carries a SynthID watermark, Google DeepMind's imperceptible pixel-level watermark. The Gemini API image generation guide states that "all generated images include a SynthID watermark," and the launch post says that "Gemini Omni and Nano Banana 2 Lite use SynthID watermarking," with verification available through the Gemini app, Gemini in Chrome and Search.^[1][3]

C2PA Content Credentials are a separate matter. Google's Nano Banana 2 announcement in February 2026 described pairing SynthID with "interoperable C2PA Content Credentials" and reported that SynthID verification had been used more than 20 million times since its November 2025 launch, with C2PA verification coming to the Gemini app.^[7] Neither the June 30 launch post nor the Gemini API image generation documentation mentions C2PA in connection with Nano Banana 2 Lite specifically, so the extent to which Content Credentials attach to this model's output is not documented in Google's own materials as of July 2026.^[1][3]

Evaluations

Vendor-reported

Google published no head-to-head benchmark scores with the launch. Its claims are qualitative plus two numbers: four-second text-to-image latency and the per-image price. Google says the model retains "reliable prompt adherence, strong character consistency and legible in-image text rendering" despite prioritizing speed, and the DeepMind model page claims "the control and accuracy you expect from Nano Banana, accelerated."^[1][2]

Independent

Artificial Analysis added the model to its image arenas on release. It reported that Nano Banana 2 Lite "debuts at #5 on the Artificial Analysis Text to Image Leaderboard, slightly ahead of the base Nano Banana 2, but ranks substantially lower in Image Editing at #18 vs. Nano Banana 2's #3."^[11] Its own end-to-end timing put a 1K generation at a median of 3.4 seconds, faster than Google's advertised four seconds and the quickest of the top ten models by text-to-image quality.^[11]

As of late July 2026 the Artificial Analysis leaderboards placed the models as follows:^[12][13]

Scroll sideways for more →

The split is the most informative independent result on the model: on single-shot text-to-image prompts it scores within noise of Nano Banana 2 at half the price, but on image editing it falls fifteen places behind it, consistent with Google's own documented caveat that the model is "not optimized for multiple reference inputs or multi-turn sequential editing."^[3]

Decrypt ran a five-test hands-on comparison against Nano Banana 2 on July 11, 2026, covering cinematic portrait realism, prompt adherence on a steampunk cityscape, spatial awareness, in-image text generation and style transfer. It found that text rendering in complex scenes actually came out ahead of the full model, attributing this to brighter default lighting improving legibility where Nano Banana 2's darker aesthetic cost smaller text its readability, and that spatial composition was adequate for most professional work. It found portrait realism visibly weaker, including anatomical errors, and prompt adherence degraded specifically on in-image labeled text accuracy, with the Lite model rendering "1942" in place of the requested "1842." Its verdict was that "it's a focused tool with a specific ceiling," best suited to "signage mockups, branded graphics, editorial composites with text-heavy elements" rather than cinematic or detail-dependent work. Decrypt measured generation at roughly four seconds, matching Google's claim.^[14]

Simon Willison tested the model in Google AI Studio on release day with a "Where's Waldo" style prompt asking for a hidden raccoon holding a ham radio, and judged the output better than what the earlier Nano Banana models had produced for the same prompt in April, while noting that signage in the generated image spelled the festival name wrong in two different ways ("FOREE'S FESTIVAL" and "FOREST FIVAL"), a reminder that the in-image text rendering Google advertises is not reliable at this tier.^[15]

Limitations

The documented and observed constraints are:

• 1K only. No 512px, 2K or 4K output, which rules out print work and any pipeline that needs high-detail typography.^[6]
• Weak at editing relative to its own family. Google states it is "not optimized for multiple reference inputs or multi-turn sequential editing," and independent editing-arena placement is far below Nano Banana 2 and Nano Banana Pro.^[3][13]
• No grounding. Neither Google Search nor Google Maps grounding is supported, so the web-grounded rendering of specific real-world subjects that Google promoted for Nano Banana 2 is unavailable here.^[6][7]
• No context caching and no structured outputs.^[6]
• Higher latency for editing than for generation, per Google Cloud.^[5]
• Photographic realism lags the full model, and style transfer loses the surrounding visual context rather than rendering quality, per independent hands-on testing.^[14]

Reception

Coverage of the June 30 launch treated the model as a pricing and throughput move rather than a capability advance. VentureBeat framed it as aimed at "low-cost, 4-second fast enterprise image generations."^[16] TestingCatalog presented it as the recommended model for users of the earlier Nano Banana, and Google Cloud's own developer post announced general availability on the Gemini Enterprise Agent Platform.^[10][5] Trade coverage generally read the release as Google segmenting its image lineup into three explicit tiers, with Nano Banana Pro for quality, Nano Banana 2 as the default, and Nano Banana 2 Lite for volume, in a market where OpenAI's GPT Image 2 held the top of both Artificial Analysis image arenas and open-weight competitors such as Black Forest Labs' FLUX.2 occupied the self-hosted end.^[12][13]

See also

• Nano Banana
• Nano Banana 2
• Nano Banana Pro
• Gemini Omni
• Gemini 3
• Gemini 3.1 Pro
• Google DeepMind
• GPT Image 2
• DALL-E 3
• FLUX.2

References

1. Alisa Fortin and Anish Nangia, Google DeepMind, "Start building with Nano Banana 2 Lite and Gemini Omni Flash," Google Blog, June 30, 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni-flash-nano-banana-2-lite/ ↩
2. "Gemini 3.1 Flash-Lite Image, Nano Banana 2 Lite," Google DeepMind. https://deepmind.google/models/gemini-image/flash-lite/ ↩
3. "Image generation," Gemini API documentation, Google AI for Developers. https://ai.google.dev/gemini-api/docs/image-generation ↩
4. "Gemini Developer API Pricing," Google AI for Developers. https://ai.google.dev/gemini-api/docs/pricing ↩
5. "Nano Banana 2 Lite and Gemini Omni Flash available," Google Cloud Blog, July 1, 2026. https://cloud.google.com/blog/products/ai-machine-learning/nano-banana-2-lite-and-gemini-omni-flash-available/ ↩
6. "Gemini 3.1 Flash Lite Image," Gemini API model documentation, Google AI for Developers. https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-lite-image ↩
7. "Nano Banana 2: Combining Pro capabilities with lightning-fast speed," Google Blog, February 26, 2026. https://blog.google/innovation-and-ai/technology/ai/nano-banana-2/ ↩
8. "Introducing Gemini 2.5 Flash Image, our state-of-the-art image model," Google Developers Blog, August 26, 2025. https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/ ↩
9. "Nano Banana Pro: Gemini 3 Pro Image model from Google DeepMind," Google Blog, November 20, 2025. https://blog.google/innovation-and-ai/products/nano-banana-pro/ ↩
10. "Google launches Nano Banana 2 Lite and Gemini Omni Flash," TestingCatalog, July 1, 2026. https://www.testingcatalog.com/google-launches-nano-banana-2-lite-and-gemini-omni-flash/ ↩
11. Artificial Analysis, post on X, June 2026. https://x.com/ArtificialAnlys/status/2074917317591564619 ↩
12. "Text to Image Leaderboard," Artificial Analysis. https://artificialanalysis.ai/image/leaderboard/text-to-image ↩
13. "Image Editing Leaderboard," Artificial Analysis. https://artificialanalysis.ai/image/leaderboard/editing ↩
14. "Nano Banana 2 Lite vs. Nano Banana 2: When to Save Your Money and When to Upgrade," Decrypt, July 11, 2026. https://decrypt.co/373002/google-nano-banana-2-lite-vs-nano-banana-2-comparison-review ↩
15. Simon Willison, "Nano Banana 2 Lite," June 30, 2026. https://simonwillison.net/2026/Jun/30/nano-banana-2-lite/ ↩
16. "Google unveils Nano Banana 2 Lite aka Gemini 3.1 Flash-Lite for low cost, 4-second fast enterprise image generations," VentureBeat, June 30, 2026. https://venturebeat.com/technology/google-unveils-nano-banana-2-lite-aka-gemini-3-1-flash-lite-for-low-cost-4-second-fast-enterprise-image-generations ↩