On this page · 10 sections
- What Google announced, and how firm the date is
- What breaks if you do nothing
- The Nano Banana family on 1 November 2026
- The official replacement is cheaper, with real limits
- Which replacement fits which marketing workflow
- Price comparison: 40,000 ad variants a month
- Migration checklist
- India: synthetic-content labelling and consent
- How eCorpIT can help
- References
Summary. Google lists 15 March 2027 as the shutdown date for gemini-2.5-flash-image, the original Nano Banana, on both the Gemini API and Gemini Enterprise Agent Platform, and names gemini-3.1-flash-lite-image (Nano Banana 2 Lite) as the replacement. The date has already moved once, from 2 October 2026, but Google's pricing page now says the model will shut down that day, so plan for it. The official swap is a small saving: $0.039 per image ($0.0195 on Batch) against $0.0336 per 1K image ($0.0168), about 14% less. The real limits are a 4,096-token output cap, newly billed thinking tokens and weaker multi-turn editing. The less obvious option is gemini-nano-banana-2.1: the same $0.0336 at 1K, plus 2K ($0.0504), 4K ($0.113), search grounding and four character references, but input at $1.50 per 1M tokens against $0.25 and medium thinking by default. For high-volume single-product variants, take Lite. For recurring people, headline text or larger assets, test Nano Banana 2.1. Switch by mid-February.
About this article. This guide is for marketing engineers and CTOs running image pipelines on Gemini. On 1 November 2026 we checked dates, capabilities and prices against Google's Gemini API and Agent Platform documentation and release notes, archived copies of those pages and OpenAI's pricing and deprecation pages. For India we read the government's consolidated IT Rules (updated as on 10 February 2026) and the DPDP Act 2023 as published. This is not legal advice, vendors change prices often, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What Google announced, and how firm the date is
Google's deprecations page, last updated on 22 October 2026, lists gemini-2.5-flash-image as released on 2 October 2025, with a shutdown date of 15 March 2027 and gemini-3.1-flash-lite-image as the recommended replacement. A standing note says listed dates are the earliest a model might be retired, with advance notice of the exact date.
Several signals say to plan for that day. The pricing page says the model will be shut down on 15 March 2027, and the image guide calls it legacy and recommends Nano Banana 2 Lite. Google Cloud's Agent Platform (formerly Vertex AI) lists the same date and replacement, and says its listed dates can be extended but not brought forward.
The date has moved before, but do not plan on it. Agent Platform's release notes of 14 September 2026 extended retirement from 2 October 2026. The Gemini API table still showed 2 October in a capture from that day and 15 March by 7 October. That same week, Google's 6 October release note gave Nano Banana 2 a shutdown date of 29 October 2026, just 23 days out, and by the next day said no date had been announced. Other image dates held, including Imagen 4's 17 August 2026. Our AI model retirement policies comparison adds context.
Teams that moved off Imagen 4 onto the original Nano Banana now face a second image migration within about seven months. Agent Platform's retired-models table still names this model as the Imagen replacement, and our August note on the Imagen 4 shutdown warned about the trap when the date was still 2 October.
Prices below are for the Gemini Developer API. On Agent Platform, Lite is listed for 28 June 2027 or later, Nano Banana 2 and Pro for 28 May 2027 or later, and Nano Banana 2.1 has no date.
What breaks if you do nothing
Once a model is shut down, Google says it is fully off and its endpoint is gone.
- Scheduled creative jobs fail. Catalogue refreshes and variant generators stop, and ad sets run on stale creative.
- No-code and agency tools break quietly. Workflow tools, plug-ins and agency scripts often pin the model ID and log failures nobody reads.
- Late Batch jobs are a risk. Google targets 24 hours, and a job still pending or running after 48 hours expires with no results.
- AI Studio prototypes stop working. Apps that call the model directly need the same change as production code.
The listed date is 134 days from 1 November 2026. A workable timeline:
| Date | Action |
|---|---|
| By 30 November 2026 | Inventory every caller, including tools outside engineering |
| By 31 January 2027 | Golden-set tests complete; replacement chosen per workflow |
| By 15 February 2027 | Production on the new model; old model kept only as fallback |
| By 10 March 2027 | No new Batch jobs on the old model |
| 15 March 2027 | Listed shutdown; remove the old model ID everywhere |
The Nano Banana family on 1 November 2026
Google's guide lists five Nano Banana models. Paid-tier rates are from the pricing page, dates from the deprecations page.
| Model | Model ID | Released | Shutdown | Output sizes | Image output, Standard | Image output, Batch | Input per 1M tokens |
|---|---|---|---|---|---|---|---|
| Nano Banana | gemini-2.5-flash-image | 2 Oct 2025 | 15 Mar 2027 | Up to 1024px | $0.039 | $0.0195 | $0.30 |
| Nano Banana 2 Lite | gemini-3.1-flash-lite-image | 30 Jun 2026 | None announced | 1K only | $0.0336 | $0.0168 | $0.25 |
| Nano Banana 2.1 | gemini-nano-banana-2.1 | 6 Oct 2026 | None announced | 1K, 2K, 4K | $0.0336 / $0.0504 / $0.113 | $0.0168 / $0.0252 / $0.0567 | $1.50 |
| Nano Banana 2 | gemini-3.1-flash-image | 28 May 2026 | Deprecated, no date | 0.5K, 1K, 2K, 4K | $0.045 / $0.067 / $0.101 / $0.151 | $0.022 / $0.034 / $0.050 / $0.076 | $0.50 |
| Nano Banana Pro | gemini-3-pro-image | 28 May 2026 | None announced | 1K or 2K, 4K | $0.134 / $0.24 | $0.067 / $0.12 | $2.00 |
The per-image figures are rounded token arithmetic. Nano Banana bills image output at $30 per 1M tokens and a 1024px image uses 1,290 tokens ($0.0387). Lite and Nano Banana 2.1 also bill $30 per 1M, but a 1K image uses 1,120 tokens ($0.0336). Nano Banana 2 bills $60 and Pro $120 per 1M, so they cost two and four times as much at 1K. The pricing pages give 3,780 tokens for a 4K image on Nano Banana 2.1, while the image guide still shows 2,520; we use the pricing figure.
Text and thinking output costs $1.50 per 1M tokens on Lite, $7.50 on Nano Banana 2.1, $3 on Nano Banana 2, $12 on Pro and $2.50 on the retiring model.
The official replacement is cheaper, with real limits
Google targets sub-two-second latency for Lite and positions it for high-volume generation. Against the model it replaces:
| Capability | Nano Banana (retiring) | Nano Banana 2 Lite | Nano Banana 2.1 |
|---|---|---|---|
| Inputs | Image and text | Text, image, video, PDF | Text, image, video, PDF |
| Input / output token limit | 65,536 / 32,768 | 65,536 / 4,096 | 131,072 / 32,768 |
| Output resolution | Up to 1024px | 1K only | 1K (default), 2K, 4K |
| Thinking | Not supported | Minimal (default) or high | Minimal, medium (default) or high |
| Google Search grounding | Not supported | Not supported | Web and Image Search |
| Reference images | Works best with up to 3 | Up to 14 objects; no character slot | Up to 10 objects and 4 characters |
On paper Lite gives up little: neither model goes above 1K or has search grounding, and Google says Lite keeps character alignment at the original's level. Five points matter.
Output cap and editing. Lite stops at 4,096 output tokens against 32,768, and Google says it is not tuned for multiple references or sequential multi-turn editing. Test workflows that refine one image over many turns.
Character consistency. Lite has no character-reference slot for a recurring model, founder or mascot. Nano Banana 2.1 holds up to four characters and Pro up to five.
Aspect ratios. Lite and the retiring model share ten ratios, including 1:1, 4:5, 9:16 and 16:9; Lite's page also mentions 14 but names only those ten. Neither Lite nor Nano Banana 2.1 offers 1.91:1, the horizontal ratio Google Ads recommends for responsive display ads, so keep a crop step. Nano Banana 2.1 adds 1:4, 4:1, 1:8 and 8:1. Pixel sizes shift too: 16:9 is 1344x768 on the old model and 1376x768 on Nano Banana 2.1, so check fixed-size templates.
Flex and Priority. Google's documents disagree. The pricing page lists Flex ($0.0195) and Priority ($0.0702) rates for Nano Banana, but its model page marks both unsupported and neither tier's guide lists an image model. Lite and Nano Banana 2.1 are also marked unsupported, so plan around Standard and Batch. On Agent Platform, Priority pay-as-you-go reached Nano Banana 2.1 on 14 October 2026.
Provenance. Every Nano Banana image carries SynthID. The Gemini API pages name C2PA only for Lite, but Agent Platform's model pages list it for all five, so it is not a Lite-only gain. Check your own output for the metadata.
Which replacement fits which marketing workflow
| Pick this | If your workflow is |
|---|---|
| Nano Banana 2 Lite | High-volume 1K variants of one product shot: background swaps, colour variants, thumbnails, A/B creative |
| Nano Banana 2.1 | Recurring people or mascots, legible headlines, infographics, 2K or 4K output, search grounding |
| Nano Banana Pro | Hero assets, brand-consistent scenes, style references and localisation worth four times the 1K price |
| Nano Banana 2 | Only for 512px output, which Nano Banana 2.1 drops; it is deprecated and notice can be short |
| OpenAI gpt-image-2.5-flare or gpt-image-2.5-sunburst | A second supplier |
OpenAI lists $8 per 1M image input and $30 per 1M image output tokens for both, with Batch at half, and its guide estimates cost by size and quality. Avoid gpt-image-1-mini and gpt-image-1.5 for a new build: OpenAI removes both on 1 December 2026. Our earlier Nano Banana 2 and Pro comparison predates Nano Banana 2.1 but still helps with hero assets.
Price comparison: 40,000 ad variants a month
Assumptions. 40,000 images a month at 1K, each from a 400-token text prompt plus one product reference image, at Google's published per-image prices with no thinking. A reference image counts as 258 tokens on Nano Banana, 1,120 on Lite, Nano Banana 2 and Nano Banana 2.1, and $0.0011 on Pro. The 1,120 figure is from Agent Platform's pricing page and matches the Gemini API's Gemini 3 default, so confirm it in your usage data.
| Model | Image output | Text input (16M tokens) | Reference images | Monthly total | On Batch |
|---|---|---|---|---|---|
| Nano Banana (retiring) | 40,000 × $0.039 = $1,560 | $4.80 | $3.10 | $1,567.90 | $783.95 |
| Nano Banana 2 Lite | 40,000 × $0.0336 = $1,344 | $4.00 | $11.20 | $1,359.20 | $679.60 |
| Nano Banana 2.1 at 1K | 40,000 × $0.0336 = $1,344 | $24.00 | $67.20 | $1,435.20 | $717.60 |
| Nano Banana 2.1 at 2K | 40,000 × $0.0504 = $2,016 | $24.00 | $67.20 | $2,107.20 | $1,053.60 |
| Nano Banana 2 at 1K | 40,000 × $0.067 = $2,680 | $8.00 | $22.40 | $2,710.40 | $1,375.20 |
| Nano Banana Pro at 1K | 40,000 × $0.134 = $5,360 | $32.00 | $44.00 | $5,436.00 | $2,720.00 |
Batch uses Google's published Batch rates, including $0.034 per image on Nano Banana 2 and $0.0006 per input image on Pro. Exact token arithmetic moves these totals by about 1% or less.
Thinking changes the picture. The retiring model has none. On Gemini 3 image models it is on by default and cannot be disabled; thinking tokens are billed, interim thought images are not. Every 1,000 thinking tokens per image adds $0.0075 on Nano Banana 2.1 ($300 a month here) and $0.0015 on Lite ($60). Google publishes no typical counts, so set the level explicitly and read token usage in week one.
The result. Lite saves about $209 a month, roughly 13%. Nano Banana 2.1 at 1K is about $133 below today's bill and $76 above Lite before thinking, and passes today's bill at roughly 440 thinking tokens per image. Its search grounding is extra: 5,000 free requests a month shared across Gemini 3 and newer models, then $14 per 1,000, with a cap of 1,500 requests a day.
No image model shows a dated price change. Our note on the Gemini 3.8 Flash price increase covers the text models that change on 1 January 2027.
Migration checklist
- Inventory every caller. Search code, environment variables, no-code automations, creative tools and AI Studio apps for
gemini-2.5-flash-image.
- Change the model ID and set parameters explicitly. Set aspect ratio, image size (with an uppercase K) and thinking level, because defaults differ.
- Build a golden set of 50 to 100 live prompts: product fidelity, logos, brand colours, faces and headlines, scored blind with the brand team. Hindi is the only Indian language on Google's best-performance list, so test Tamil, Bengali and other scripts harder.
- Re-check rejection rates. A new model can refuse prompts the old one accepted, or the reverse.
- Preserve provenance. Check that resizing, compression and CDN steps do not strip C2PA metadata.
- Move non-urgent jobs to Batch for 50% off.
from google import genai
client = genai.Client()
# Before: model="gemini-2.5-flash-image"
interaction = client.interactions.create(
model="gemini-3.1-flash-lite-image", # or "gemini-nano-banana-2.1"
input="Studio shot of the steel water bottle on a terracotta background",
response_format={"type": "image", "aspect_ratio": "4:5", "image_size": "1K"},
generation_config={"thinking_level": "minimal"},
)
Google recommends the Interactions API for new work but says generateContent remains fully supported, so if your code uses it, change the model string there and run the golden set first.
India: synthetic-content labelling and consent
India's Ministry of Electronics and Information Technology notified amendments to the IT Rules on 10 February 2026, in force from 20 February 2026. Synthetically generated information means audio or visual content made or altered by computer that appears real and depicts a person or event in a way likely to be taken as genuine; routine good-faith edits such as colour adjustment are excluded. A realistic AI-generated model in an ad probably falls inside; a stylised product background may not. Take advice on borderline creative.
Intermediaries whose tools can create such content must label it prominently, embed permanent metadata or another provenance mechanism as far as technically feasible, and not enable its removal. Significant social media intermediaries must ask users to declare it before publication, verify the declaration and label confirmed content. Advertisers are not the direct addressees, but the platforms running your ads are, so keep SynthID and C2PA intact and answer declarations accurately. Our guide to the 2026 IT Rules has more.
Reference photos of identifiable people are personal data under the DPDP Act. Record consent for AI use, and see our DPDP checklist for the 13 May 2027 deadline. Since 26 October 2026, new paid-tier projects using the generateContent API store requests by default; opt out per request or in the logging settings if your policy requires it.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI and machine learning services team handles image-model migrations, from caller inventory and golden-set testing to Batch pipelines and provenance checks. Talk to us at /contact-us/.
Last updated: 1 November 2026.
