Nano Banana 2 Lite: Inside Google's 4-Second, $0.034 Image Model

Google shipped a new image model on June 30, 2026, and it did not arrive as a splashy keynote demo. It landed as a single API string — gemini-3.1-flash-lite-image — quietly enabled in Google AI Studio, then confirmed by Logan Kilpatrick on X, then written up by the Google DeepMind product team a few hours later. The name is Nano Banana 2 Lite. The pitch is speed and cost. The interesting part is what it says about the shape of Google's image stack going forward.

TLDR Nano Banana 2 Lite (gemini-3.1-flash-lite-image) is Google's fastest and cheapest Gemini image model, delivering text-to-image outputs in about 4 seconds at $0.034 per 1K-resolution image, per Google's official launch post. It replaces the original Nano Banana (gemini-2.5-flash-image) and is positioned as the entry tier below Nano Banana 2 and Nano Banana Pro. It launched alongside Gemini Omni Flash, a $0.10-per-second video preview, and the two are clearly meant to chain into an image-to-video pipeline.

Key Takeaways

• Nano Banana 2 Lite is exposed as gemini-3.1-flash-lite-image in the Gemini API and Google AI Studio, live as of June 30, 2026.
• Google states 4-second text-to-image latency and roughly $0.034 per 1K-resolution image, with API pricing listed at $0.25 per 1M input tokens and $1.50 per 1M output tokens.
• The Nano Banana family now spans three tiers: Lite (speed and cost), 2 (generalist), and Pro (highest fidelity, up to 4K).
• It is the recommended migration target from the original Nano Banana (gemini-2.5-flash-image), now described as legacy.
• Trade-offs are honest: 1K output cap, weaker small text rendering, no Search grounding, and character-consistency wobbles across scene changes.
• Gemini Omni Flash launched the same day at $0.10 per second of video, and Google frames the pair as a chained image-to-video pipeline.

What Actually Shipped

The concrete facts, dated and sourced.

On June 30, 2026 at 15:38 UTC, TestingCatalog spotted gemini-3.1-flash-lite-image live in Google AI Studio, flagging it as Nano Banana 2 Lite. The initial pricing screen listed image input at $0.25 and image output at $0.0336. About twenty-six minutes later, Logan Kilpatrick confirmed the launch, describing the model as "extremely fast (<4s image) & cheap ($0.034 / 1K image)" and announcing Gemini Omni Flash alongside it.

The Google AI Studio account then posted its own copy: "our fastest, most cost-effective gemini image model yet," built for "high-velocity developer pipelines," delivering "text-to-image outputs in 4 seconds at just $0.034 per 1K-resolution image." Google's official Keyword blog post from DeepMind product managers Alisa Fortin and Anish Nangia followed, and Google Cloud published a parallel announcement confirming general availability in the Gemini Enterprise Agent Platform.

Third-party access came the same afternoon. Replicate hosted the model as google/nano-banana-2-lite, with a first-run measured at 5.4 seconds and support for up to 14 reference images.

Source: @OfficialLoganK

The Model Family Reshuffle

The more interesting story is not the Lite tier itself. It is that Google now has a clean three-tier image family with named boundaries. According to Google's launch post, the family reads:

• Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image): Built for speed. Optimized for near-real-time, high-volume workflows.
• Nano Banana 2 (Gemini 3.1 Flash Image): The "generalist workhorse," described as the best balance of performance and cost.
• Nano Banana Pro (Gemini 3 Pro Image): The high-fidelity tier.

Legacy Nano Banana, powered by Gemini 2.5 Flash Image, is being pushed toward retirement. Google explicitly recommends Lite as the drop-in replacement, and TechCrunch's coverage notes the company now refers to the original as its "legacy model."

Product-tier reshuffles like this rarely come with fanfare, but they matter for anyone with a bill. If a workflow is still on gemini-2.5-flash-image, the migration path is now Lite by default, and the price and speed math will change. Nano Banana 2 Lite is not just a new SKU. It is the new floor of the family.

Pricing, Precisely

Google's numbers are unusually specific, so worth pinning down. Per the launch materials and Ars Technica's independent write-up:

• Per-image headline: $0.034 per 1K-resolution image.
• API token pricing: $0.25 per 1M input tokens, $1.50 per 1M output tokens under standard pricing.
• Per-image list price: $0.0336 per 1K-resolution image (the underlying figure that rounds to $0.034 in marketing copy).
• Comparison to sibling: Ars Technica reports that Lite's token pricing is "half the rate for Nano Banana 2."
• Comparison to Pro: Nano Banana Pro is priced at $2 per 1M input tokens (only marginally higher) but $12 per 1M output tokens — roughly 8× the Lite output rate.

The Fast-Cheap Floor — the cost per finished 1K image at the lowest reasonable latency — is now $0.034 for Google. That is the number every image-heavy pipeline will benchmark against.

Latency and the Low-Thinking Mode

Speed is where Lite earns its name. Google states 4 seconds for a text-to-image output, and Ars Technica measured the standard Nano Banana 2 taking about 20 seconds on the same prompt — a roughly 5× gap. Kilpatrick's post uses "<4s image" as the headline.

Ars Technica adds a useful detail: the 4-second figure is for the "default low-thinking mode." That framing suggests Lite exposes a thinking-budget knob and defaults it low, which is consistent with a model targeted at high-throughput drafting rather than one-shot final art. Low-Thinking Mode is worth watching as a coined idea across Google's stack — if it appears as a documented parameter, it will shape how teams tune quality-vs-cost on the API side.

Replicate's playground shows a first generation at 5.4 seconds, which is close to Google's claim once round-trip and cold-start overhead are factored in.

What It Is Good At

Nano Banana 2 Lite is positioned squarely at throughput. Both Google's launch post and independent coverage agree on the sweet spot:

• Rapid prototyping and ideation.
• A/B testing ad variations.
• Product mockups and marketplace thumbnails.
• Localized creative at scale.
• Feeding downstream video generation models with cheap reference images.

Rohan Paul frames the workflow well: "Chaining both models is the real product shape, not either model alone. Nano Banana 2 Lite makes reference images, then Gemini Omni Flash animates them." This chained pipeline — call it the Image-to-Video Assembly Line — is Google's clearest product bet in the release, and the pricing math is the tell. A 4-second $0.034 image feeding a $0.10-per-second video means a 10-second animated product clip costs roughly $1.034 in raw model spend, before any regeneration. That is a number an e-commerce ops team can actually plan against.

Google's own launch demo, "Omni Product Studio," turns static product shots into cinematic e-commerce videos, which is exactly this pipeline in a wrapper.

Where It Falls Short

Google's launch post and Ars Technica are refreshingly direct about the limits. The Lite tier is honest about what it gives up:

• Text rendering wobbles on small text, per Google's own documentation.
• Infographics are more likely to include incorrect data.
• Character consistency across iterations is weaker than the generalist tier.
• Output resolution caps at 1K (~1 megapixel). For 2K or 4K, developers must move to Nano Banana 2 or Pro.
• Search grounding is not supported. Nano Banana 2 has Google Search and Image Search grounding for real-time web context; Lite does not.

For any use case that requires legible signage, infographic data, or narrative continuity across a sequence of frames, Lite is not the right pick. The Prompt-Adherence Floor — how faithfully the model follows a complex prompt — is presumably lower here, though Google claims Lite still "retains reliable prompt adherence, strong character consistency and legible in-image text rendering" for typical cases.

Google's launch post references Arena.ai Elo scores placing Lite "almost as highly" as the non-Lite version, but Ars Technica notes that "vibemarking doesn't always focus on the details that can make AI images look silly upon closer inspection." That skepticism is warranted until independent evals land.

Availability and Distribution

Nano Banana 2 Lite is live across an unusually wide surface for a launch day:

• Developer surfaces: Google AI Studio, Gemini API, Gemini Enterprise Agent Platform.
• Consumer surfaces: AI Mode in Search, Gemini app, NotebookLM, Google Photos, Stitch, Google Flow, Google Ads.
• Third-party: Replicate hosted the model within hours of launch.

Google Cloud's post confirms enterprise availability, and quotes Adobe's Matt Chotin saying the models are being brought into Adobe Firefly. That partnership matters — Firefly integration means Lite will end up in professional creative pipelines regardless of individual developer preferences.

Nano Banana 2 Lite vs Nano Banana 2: What the Signal Says

The bundle contains enough data to draw a direct comparison against the sibling model in the same family. Community sentiment against models from other vendors (Midjourney, FLUX, GPT Image) is absent from this signal window, so the honest comparison is intra-family.

• Text-to-image latency. Lite: ~4 seconds, Google-stated. Nano Banana 2: ~20 seconds for the same prompt, per Ars Technica. Roughly a 5× speedup.
• Price per 1K image. Lite: ~$0.034 per 1K-resolution image. Nano Banana 2: roughly double, per Ars Technica's note that Lite is "half the rate for Nano Banana 2."
• Output resolution. Lite: 1K (~1 megapixel). Nano Banana 2: up to 2K and 4K.
• Search grounding. Lite: not supported. Nano Banana 2: supports Google Search and Image Search grounding.
• API model ID. Lite: gemini-3.1-flash-lite-image. Nano Banana 2: gemini-3.1-flash-image.

The gap is engineered, not incidental. Google appears to be deliberately preserving Nano Banana 2 as the default for anything that needs high-resolution output, grounding, or maximum character consistency, while Lite is the throughput floor. Head-to-head against non-Google models — FLUX.2, Ideogram V3, GPT Image 2, Qwen Image 2.0 — is not addressable from this signal set, and independent Elo comparisons should be treated as pending.

What We Know vs. What We Don't

The signal is unusually rich for a same-day release write-up. It is also full of Google-stated numbers that have not yet been independently verified.

What we know:

• Nano Banana 2 Lite is live as gemini-3.1-flash-lite-image in the Gemini API and Google AI Studio as of June 30, 2026, per Google's launch post.
• Google states 4-second text-to-image latency in default low-thinking mode, versus ~20 seconds for Nano Banana 2 per Ars Technica.
• Pricing is ~$0.034 per 1K-resolution image, with API rates of $0.25 per 1M input tokens and $1.50 per 1M output tokens.
• It is available in Google AI Studio, Gemini API, Gemini Enterprise Agent Platform, AI Mode in Search, the Gemini app, NotebookLM, Google Photos, Stitch, Google Flow, and Google Ads.
• Google explicitly recommends it as the replacement for the original Nano Banana (gemini-2.5-flash-image), now described as legacy.
• Gemini Omni Flash (gemini-omni-flash-preview) launched the same day at $0.10 per second of video output.
• All generated images carry SynthID watermarking.
• Documented weaknesses include small-text rendering, infographic accuracy, and character consistency across iterations.
• No Google Search or Image Search grounding is supported in the Lite tier.

What we don't know:

• No independent third-party benchmarks — Arena.ai Elo scores referenced in Google's materials come from Google's own framing; external evaluations from independent labs are not yet published in this signal window.
• Concurrency, requests-per-minute ceilings, or enterprise rate-limit tiers have not been publicly disclosed.
• Whether "low-thinking mode" is a user-controllable API parameter or an internal default is not documented in the materials available here.
• The exact prompt-adherence delta between Lite and Nano Banana 2 on identical prompts — Google shows examples, but no measured accuracy percentages appear in the launch post.
• Whether the $0.034 headline holds under complex prompts, multi-reference edits, or 14-image inputs on Replicate remains untested externally.
• Regional availability, data-residency options, and any enterprise-specific SLAs are unclear from public materials.
• No official statement on whether Nano Banana 2 Lite supports the full three sequential edits documented for the broader Nano Banana / Omni Flash Interactions API.

Why This Matters for Builders

The interesting question is not whether Nano Banana 2 Lite is a good image model. It is what a 4-second, $0.034 image does to the shape of a creative pipeline.

Three things change at this floor. First, the cost of iteration collapses. A creative director who was budgeting a dozen tries per concept can now budget a hundred. A/B testing shifts from a batch process to a live one. Second, the image-to-video assembly line becomes economically viable at operational scale — not just for demos, but for actual product catalog generation. Third, the legacy Nano Banana migration is now urgent for any team still on gemini-2.5-flash-image, because Google's positioning of that model as "legacy" is the standard prelude to eventual deprecation.

For engineering leads, the practical read is that Lite is the right default for anything drafting-shaped and the wrong default for anything final-shaped. The failure modes — small text, infographics, character consistency — are exactly the failure modes that matter most in polished output. A pragmatic pattern is Lite for the first 90% of the funnel, Nano Banana 2 for the final render, and Pro reserved for hero assets.

The SynthID watermark is also worth thinking about. Every Lite output carries it. For any pipeline that ingests Lite output and re-processes it downstream (video generation, further editing, human touch-up), the watermark's persistence claim is the interesting testable property.

How to Evaluate It Yourself

The community sentiment layer is thin so far — five evidence tweets, one Ars Technica review, one TechCrunch write-up, and Google's own materials. Independent evaluation is the next thing to watch.

A practical framework for teams evaluating Lite in the first week:

1. Prompt-parity test. Run 50–100 prompts through both gemini-3.1-flash-lite-image and gemini-3.1-flash-image and score the outputs blind. The 5× speed and 2× cost advantages of Lite only matter if the quality gap is smaller than the price gap.
2. Text-rendering audit. Pick prompts with small text (product labels, UI copy, signage). Lite is documented as weaker here — measure how much weaker in your domain.
3. Multi-image reference behavior. Replicate's implementation supports up to 14 reference images. Test whether increasing reference count degrades speed proportionally or Lite maintains its 4-second envelope.
4. Chained pipeline latency. If the goal is image-to-video, measure end-to-end time from prompt to finished 10-second Omni Flash clip. Google's marketing implies fast; real workflows tend to reveal queueing.
5. SynthID persistence. Round-trip a Lite output through common editing tools and check watermark survival.

None of these require Google's cooperation. All of them will produce numbers other blogs will want to cite.

What to Watch Next

Three concrete signals to track over the next week:

• Watch the model card. A full Nano Banana 2 Lite model card with formal benchmarks against FLUX.2, Ideogram V3, and GPT Image 2 has not appeared. When it does, the Arena.ai Elo claim becomes verifiable.
• Run your own latency eval. The 4-second claim is Google-stated and measured on Google's infrastructure. Replicate's first-run at 5.4 seconds already hints at real-world variance. Any team planning to bet a product on the number should measure it in their own region.
• Track legacy Nano Banana deprecation. Google has not yet posted a formal sunset date for gemini-2.5-flash-image, but calling it "legacy" is the first move. Watch the Gemini API changelog for a deprecation notice, especially if your production traffic still runs on it.

Building similar high-throughput image and video pipelines? On kie.ai you can try Nano Banana 2, Nano Banana Pro, and Gemini Omni.