Nano Banana 2.1: Google Halves Image Prices and Retires Nano Banana 2

October 6, 2026 — Google released Nano Banana 2.1, a new image generation and editing model that runs at Flash-tier speed and costs roughly half as much as the model it replaces. Per Google's Gemini API documentation, the model ID is gemini-nano-banana-2.1, it supports 1K, 2K and 4K output, and it is positioned as an efficient alternative to the more powerful Gemini 3 Pro Image. The Decoder reports that a 1K image drops from 6.70 cents to 3.36 cents and a 4K image from 15.10 cents to 7.56 cents.

There is a deadline attached. The previous Nano Banana 2 is being retired, with shutdown on October 29, 2026, per The Decoder. If your app calls the older model ID, you have about three weeks to test and switch.

TL;DR: what you need to know

What actually changed from Nano Banana 2?

Google says the model improves "across the board" on its predecessor. The specific areas named in coverage are visual quality, prompt adherence, text rendering, character consistency across turns, panoramas and infographics. The Decoder reports the model can hold consistency for up to four characters and ten objects while fusing up to 14 reference images, which is the feature that matters most for product shots, storyboards and brand work.

The API docs list the capabilities developers will touch directly: image generation at three resolutions, multi-image fusion, search grounding through Google Web and Image Search, and configurable thinking levels of minimal, medium and high. Inputs can be text, images, video and PDFs; outputs are images and text. Input is capped at 131,072 tokens and output at 32,768.

The docs also list what is not supported, which is useful if you are porting an agent: audio generation, caching, code execution, file search, function calling, Maps grounding, the Live API, structured outputs, URL context, and flex or priority inference. The Batch API is supported. If your pipeline relies on function calling or structured outputs around the image step, plan to keep that logic in a separate text model call.

The Decoder additionally says the model is built on Gemini 3.6 Flash. Google's model page does not state this in the pages we reviewed, so treat that detail as the outlet's reporting. For background on that model family, see our post on the Gemini 3.6 Flash launch.

Is it really half the price?

Image output is. Output pricing for a 1K image fell by about half, and the 4K price fell by a similar proportion. Be careful with the total bill, though. The figures here cover per-image output only; check Google's pricing page for input-token pricing, which matters if you send many reference images or long prompts, because the input side could change the total saving.

A quick way to model it for your workload:

1. Count how many images you generate per month at each resolution.
2. Multiply by the new per-image output price.
3. Add input cost: average reference images per request times tokens per image times the input price.
4. Compare to your current Nano Banana 2 bill.

Edit-heavy workloads with many reference images may see a smaller saving than pure text-to-image if input costs differ.

Is it better than Nano Banana Pro?

Sometimes on paper, less often in practice. Google says 2.1 beats the previous Pro model in some benchmarks at a lower cost. The Decoder's hands-on comparison is more guarded: the previous model also scored well in tests, and Pro "still looks more realistic and natural, particularly in its colors and proportions," while 2.1 occasionally gets the scale of objects wrong.

That fits the usual tiering. Pro is about four times the price per 1K image, so the question is whether the extra realism pays for itself. A reasonable rule: use 2.1 for volume work such as marketing variants, social graphics, UI mockups and infographics, and keep Pro for hero images where photographic realism and exact proportions are the product.

If you are weighing Google against OpenAI's image stack, our coverage of ChatGPT Images 2 and GPT Image 2, the transparent background API preview and the GPT Image 2.5 stop-motion how-to gives the other side of the comparison. And OpenAI is testing image generation inside ads, covered in ChatGPT visual ads test.

How do I migrate before October 29?

1. Find every call site. Search your code, configs and prompts for the old model ID. Pinned IDs in cron jobs and batch scripts are the ones people miss.
2. Swap the model ID to gemini-nano-banana-2.1 in a staging environment.
3. Run a fixed regression set. Generate the same 30 to 50 prompts on both models and compare text legibility, identity consistency across turns, and object scale. The Decoder flagged scale as a weak spot, so include prompts with measurable proportions.
4. Re-check thinking level. The new thinking setting trades latency and cost for quality. Start at medium and move it per use case.
5. Audit unsupported features. If you used function calling or structured outputs on the old model, move that to a separate call.
6. Update cost alerts. Your per-image cost drops, but input cost may rise, so recalibrate budgets.
7. Ship before the shutdown. After October 29 requests to the old ID will fail.

What should you test first?

Google's own list of improvements is a decent test plan. Write one prompt for each claim and judge the result yourself rather than trusting a benchmark chart.

• Text rendering: ask for a poster with a headline, a subhead and a three-item price list, then check spelling and alignment.
• Character consistency: generate a character, then request the same person in three new scenes over several turns.
• Infographics: supply a small table of real numbers and ask for a labeled chart, then verify every figure against the source table.
• Panoramas: request a wide landscape and look for repeated patterns or seams.
• Multi-image fusion: upload several product photos and ask for a single catalog layout, watching for scale errors.

Keep the outputs next to the same prompts run on Nano Banana Pro, so the quality gap is visible in your own use case rather than in someone else's test.

What people are asking

Does the new model replace Nano Banana Pro? No. Pro remains the premium option at a higher price. Google also continues to test newer variants; our post on the Nano Banana 2.5 test on LMArena covers an earlier report about a different, unconfirmed model.

Where can non-developers use it? Google says it is rolling out across the Gemini apps, Google Search, AI Studio and enterprise platforms. Availability can vary by region and plan.

Is it good for infographics and text? Google highlights text rendering and infographic accuracy as improvements. As always with image models, check every word and number in the output before publishing; text rendering is better but not perfect.

Does it mark images as AI-generated? Google's image models have used SynthID watermarking in the past; the model page we reviewed does not spell this out for 2.1, so confirm in the docs before building a compliance workflow around it.

How does this connect to Google's other image work? Nano Banana is already used in other Google products, for example in Google Earth image generation. A cheaper model makes those embedded uses more economical to scale.

Why this matters

Price and deprecation schedules shape what builders ship more than benchmark deltas do. Cutting the cost of a 4K image to under eight cents makes high-resolution batch work, such as catalog images or personalized creative at scale, far easier to justify. Setting a hard shutdown date for the previous model, only seven months after its February 2026 release, also shows how fast Google is cycling its image line. Teams that wrap image generation behind their own interface, rather than hard-coding a model ID everywhere, will absorb these changes more easily.

The counterweight is quality. A cheaper default model that is slightly worse at proportion and realism pushes teams toward a two-tier setup: 2.1 for drafts and volume, Pro for finals. That is more operational complexity, but it is also more control over spend.

Limitations and caveats

• Pricing figures come from The Decoder's reporting and secondary sources; confirm them on Google's pricing page before budgeting.
• The Gemini 3.6 Flash base-model detail is The Decoder's reporting and is not confirmed by the API page.
• Google's benchmark claims are the company's own; hands-on comparison found Pro more natural.
• Availability across apps and regions is rolling out and may differ by plan.

Related reading

• Google DeepMind tests Nano Banana 2.5 on LMArena
• Google Earth gets Nano Banana image generation
• Gemini 3.6 Flash and 3.5 Flash-Lite cyber launch
• ChatGPT Images 2 and GPT Image 2
• GPT Image 2 transparent backgrounds API preview
• Krea 2 open-weights image model technical report
• Gemini 4 Argon launch, benchmarks and pricing
• Official: Gemini API docs, Nano Banana 2.1

Details reflect reporting and Google documentation as of October 6, 2026; pricing and availability may change.