How to Use Nano Banana 2 Lite: Speed, Price, and Limits

Google shipped Nano Banana 2 Lite on June 30, 2026, the fastest and cheapest tier of its Nano Banana image line^[1]. Its official API name is Gemini 3.1 Flash-Lite Image, and the entire point is speed: a 1K image in about four seconds, at roughly half the cost of the full Nano Banana 2.

Knowing how to use Nano Banana 2 Lite well comes down to one number that is easy to miss. On image generation it gives up about 19 Elo points to base Nano Banana 2, roughly a 1.5% quality gap, while running five times faster and costing half as much^[1]. That trade is excellent for most work and wrong for a specific kind of work, and this guide is about telling those apart.

One naming trap first: the text-only gemini-3.1-flash-lite is a different model and does not generate images. The image model is the separate -image variant^[1].

TL;DR

• About four seconds per 1K image, against 20 seconds for base Nano Banana 2^[1].
• Roughly 3.4 cents per 1K image at Google's list rate, or about 1.7 cents in batch mode. On reAPI it is 22 credits per image, which is $0.022, about 35% below Google's published rate.
• 1K is the ceiling. No 2K, no 4K. Those need base Nano Banana 2 or Nano Banana Pro^[1].
• Generation quality nearly matches the workhorse tier (1251 vs 1270 Elo) but editing does not (1308 vs 1387)^[1].
• Ten aspect ratios, including 16:9, 9:16, and 21:9^[1].
• Rated "Low reasoning" by Google, so complex multi-constraint scenes belong on a higher tier^[1].

The three Nano Banana tiers

Nano Banana 2 Lite replaces the first-generation Nano Banana, which Google now labels legacy, so the lineup is a clean three-tier stack^[1].

Note the version detail. Lite and base Nano Banana 2 both run on Gemini 3.1 Flash, while Pro runs on the Gemini 3 Pro Image line, a different and more capable model built for reasoning-heavy generation^[1]. Lite is explicitly the medium-quality, low-reasoning tier. You trade fidelity and complex-instruction handling for the lowest latency and cost in the family.

Benchmarks

Google published a four-panel chart comparing Lite against base Nano Banana 2, the legacy model, and three competitors. Elo scores are sourced to LMArena, latency to Artificial Analysis^[1].

The trade is favorable. On generation, Lite gives up about 19 Elo to base Nano Banana 2 while running five times faster at half the price. On editing the gap is wider, 1308 against 1387, so edit-heavy workloads still favor the full model. Against the outside field, Lite beats Flux 2 Klein 9B and Seedream v5 Lite on both quality axes, and beats Seedream on speed by more than ten times. It trails Grok Imagine Image slightly on editing while leading it on generation^[1].

Two caveats belong with that table. Google picked its own comparison set and did not include GPT Image, so no official Lite-versus-GPT-Image number exists. And Google reports Elo values without a published leaderboard rank, so any "Lite is number one" claim is unverified^[1].

Capabilities and limits

• Resolution: 1K (1024px) and 512px only. No 2K or 4K, which are exclusive to base Nano Banana 2 and Pro. On reAPI the model is exposed as a single 1K output, so there is no resolution field to set.
• Aspect ratios: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9. That covers social, web, and widescreen including the 16:9 slide ratio^[1].
• Core features: text-to-image, image editing, prompt adherence, character consistency, and legible in-image text, all inherited from the family. Multi-image composition works, though exact character and object limits are not clearly documented^[1].
• Text rendering languages: Google did not enumerate a language list for Lite specifically. Treat non-Latin scripts and long strings as something to test rather than assume^[1].

The honest framing is that Lite is a good-enough-for-volume model. It handles clean single-subject generation, straightforward edits, and standard aspect ratios at speed, and steps back from high fidelity, complex multi-constraint scenes, and dense factual infographics, which is what the low-reasoning rating is telling you.

Pricing and the cost math

This is where the model earns its place, and it is worth stating carefully because secondary coverage got it wrong.

Google bills Lite at $0.25 per million input tokens and $30 per million output tokens. A 1K image is 1,120 output tokens, which works out to $0.0336, about 3.4 cents per single image, not per thousand images as some outlets reported. Batch mode halves the output rate to $15 per million, dropping a 1K image to roughly 1.7 cents. There is no free tier^[1].

Within the family, base Nano Banana 2 bills output at $60 per million (a 1K image is about $0.067, 2K about $0.10, 4K about $0.15) and Pro at $120 per million (4K about $0.24)^[1].

The detail that makes the math simple: Lite and base Nano Banana 2 use the same 1,120 tokens for a 1K image. The entire saving comes from the cheaper per-token output rate, not from a smaller image^[1]. For 1K work, Lite is a straight 2x discount over the workhorse tier.

On reAPI the rate is 22 credits per image, which is $0.022, roughly 35% below Google's published $0.034. One image per request, one flat rate, no resolution tiers to reason about.

How it compares to competitors

Against the field Google chose to show, Lite is quality-competitive but not the cheapest. Flux 2 Klein 9B at $0.015 and Grok Imagine Image at $0.020 both undercut it. Seedream v5 Lite is close on price at $0.035 but eleven times slower^[1].

Where Lite wins is the balance of three axes at once: strong Elo, four-second latency, and mid-tier pricing. If your only metric is cost per image, cheaper models exist. If you need decent quality and real-time latency together, Lite is the sweet spot.

How to use Nano Banana 2 Lite: prompting that works

The low-reasoning rating means it rewards clear, concrete prompts over clever multi-constraint ones.

• Lead with subject and style, then details. "A minimalist product hero shot of a ceramic coffee mug, soft studio lighting, warm neutral background" beats a paragraph of layered conditions.
• State the aspect ratio explicitly. All ten are supported; saying so up front avoids awkward crops.
• Keep in-image text short. Legible text works, but quality on long strings and non-Latin scripts is undocumented. Headlines and single labels are safe; dense paragraphs are risky.
• Iterate cheaply, then finalize. At four seconds and a couple of cents per image, generate a dozen variations here, then re-render only the winner on a higher tier if you need 2K or 4K.
• Use editing for controlled changes. For same-scene background swaps, Lite's 1308 editing Elo is capable, though the full model's 1387 is stronger for demanding edits.

Lite, base Nano Banana 2, or Pro

Three questions settle it.

Do you need output above 1K? If yes, skip Lite. It caps at 1K.

Is the task edit-heavy or reasoning-heavy? Lite's editing Elo trails base Nano Banana 2 by a real margin, and complex multi-constraint infographics belong on Pro. For straightforward generation, Lite is nearly as good at half the cost.

Is latency or volume the constraint? Lite is the only tier that delivers sub-five-second turnaround. Base Nano Banana 2's 20-second latency rules it out of interactive, high-volume loops.

The healthy pattern is two-tier: draft on Lite, finalize on a larger model only for the specific assets that earn the extra cost. For most teams the overwhelming majority of generated images never need that upgrade, which is exactly why a fast, cheap tier changes the economics of shipping custom visuals at all.

Calling Nano Banana 2 Lite on reAPI

reAPI exposes the model on an async task endpoint. Submit returns a task_id; poll until it is ready.

curl https://reapi.ai/api/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-2-lite",
    "prompt": "a hand-illustrated travel postcard of Kyoto at dusk, glowing paper lanterns, warm cinematic light",
    "aspect_ratio": "16:9"
  }'

Editing uses the same endpoint. Add image_urls with up to ten reference images and the request becomes a prompt-based edit rather than a text-to-image generation. There is no mode field to set; the presence of image_urls is what switches it.

curl https://reapi.ai/api/v1/images/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nano-banana-2-lite",
    "prompt": "replace the background with a quiet stone lane at dusk, keep the subject unchanged",
    "image_urls": ["https://example.com/source.png"],
    "aspect_ratio": "16:9"
  }'

Two platform rules worth knowing before you write the integration. Media inputs are public http(s) URLs only, with no base64 or data: accepted on any model. And billing is a flat 22 credits per image with one image per request, so cost forecasting is multiplication rather than token estimation.

Full request and response shapes are in the reapi.ai/docs/nano-banana-2-lite reference, and current rates are on reapi.ai/models/nano-banana-2-lite.

FAQ

What is Nano Banana 2 Lite's official name?

Gemini 3.1 Flash-Lite Image. "Nano Banana 2 Lite" is Google's marketing name for that image model. The text-only gemini-3.1-flash-lite is a different model that does not generate images^[1].

How much does Nano Banana 2 Lite cost?

About 3.4 cents per 1K image at Google's list rate, 1,120 output tokens at $30 per million, or roughly 1.7 cents in batch mode. On reAPI it is 22 credits, which is $0.022. Ignore the "per 1,000 images" figure some outlets published; the rate is per single image^[1].

How fast is Nano Banana 2 Lite?

Around four seconds per 1K image, against 20 seconds for base Nano Banana 2, roughly five times faster^[1].

What resolutions does it support?

1K (1024px) and 512px only. No 2K or 4K; those need base Nano Banana 2 or Nano Banana Pro. On reAPI the model is exposed as a single 1K output^[1].

Can Nano Banana 2 Lite edit images?

Yes. Send image_urls alongside the prompt and the same endpoint performs a prompt-based edit, with up to ten reference images. Its editing Elo of 1308 trails base Nano Banana 2's 1387, so demanding edits still favor the larger model^[1].

How does it compare to GPT Image?

There is no official head-to-head. Google did not include GPT Image in its comparison set^[1].

Does it accept base64 images?

Not on reAPI. Every model on the platform takes public http(s) URLs only for media inputs.

What aspect ratios are supported?

Ten: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9^[1].

Picking the tier that matches the job

Nano Banana 2 Lite is a well-judged addition rather than a headline model, and that is the point. It keeps almost all of base Nano Banana 2's generation quality while running five times faster and costing half as much, in exchange for a 1K ceiling, medium fidelity, and weaker editing and reasoning.

For high-volume, latency-sensitive, cost-sensitive work it is the right default, with the larger tiers reserved for assets that genuinely need 2K or 4K output or maximum quality. If you are wiring it into a product, remember the corrected math and the platform rules: about 3.4 cents per single 1K image at list, 22 credits on reAPI, public URLs only for reference images, and one image per request. That is how to use Nano Banana 2 Lite without paying workhorse-tier prices for draft-quality work.

References

1. Google. Gemini API — image generation, models, and capabilities. Retrieved July 2026 from ai.google.dev/gemini-api/docs/image-generation
2. Google. Gemini Developer API pricing. Retrieved July 2026 from ai.google.dev/gemini-api/docs/pricing
3. Google DeepMind. Gemini image generation models. Retrieved July 2026 from deepmind.google/models/gemini-image

Further reading

• reAPI. Nano Banana 2 Lite endpoint reference. reapi.ai/docs/nano-banana-2-lite
• reAPI. Model catalog. reapi.ai/models