Gemini 2.5 Flash Image (Nano Banana): Deprecated, Now What

If you are reading about Gemini 2.5 Flash Image in October 2026, you are probably maintaining something built on it. The model that introduced the "Nano Banana" name is deprecated, Google recommends a replacement, and the useful questions are about migration rather than capability.

This guide covers what the model actually introduced, the confusion over its shutdown date, and how to move off it without breaking the parts of your pipeline that quietly depend on its behaviour.

What it introduced

Google announced Gemini 2.5 Flash Image on 26 August 2025 as a preview, and the Gemini API changelog records it reaching general availability as gemini-2.5-flash-image on 2 October 2025. Its pitch was not raw image quality. It was editing through conversation, with four capabilities Google's launch post singled out:

• Character consistency — keeping the same person or object recognisable across prompts and edits.
• Prompt-based local edits — blur a background, remove a stain, change a pose, in plain language.
• Multi-image fusion — put an object into a scene, or restyle a room from a reference.
• World knowledge — reading a hand-drawn diagram or following a multi-step edit using Gemini's general understanding.

Google was also candid about its weaknesses at launch: long-form text rendering, character consistency in harder cases and fine factual detail all needed work. Those are exactly the areas its successors were marketed on, which tells you where the original falls short.

Every image it creates or edits carries an invisible SynthID watermark, according to the same post.

Where it stands in October 2026

Google has deprecated it, but its own pages disagree on when it goes. The Gemini API pricing page states that gemini-2.5-flash-image "is deprecated and will be shut down on October 2, 2026". The deprecations page lists 15 March 2027 as the shutdown date and describes such dates as the earliest possible retirement. The earlier gemini-2.5-flash-image-preview endpoint is already gone; it was shut down on 15 January 2026.

Do not plan around the later date. If a request against the Google endpoint fails tomorrow, that is within what Google has published. Third-party catalogues may keep listing the model for a while, but most of them ultimately depend on Google serving it, so a listing is not a guarantee of service.

For the full GA spec, Google's model page lists 65,536 input tokens and 32,768 output tokens, text and image in, text and image out, a June 2025 knowledge cutoff, and Batch API support. It does not support thinking, search grounding, function calling or structured outputs. Some catalogues list structured outputs and a smaller context; Google's page is the one to trust.

How it was priced

Image output is billed as tokens. Google prices it at $30 per million output tokens, and a 1024x1024 image costs 1,290 tokens, which is the origin of the much-quoted $0.039 per image. Input is $0.30 per million tokens. The Batch API halves both, to $0.0195 per image. In November 2025 Google also cut the input cost of an image from 1,290 tokens to 258, which made edit-heavy workloads cheaper.

On OpenRouter the model was listed at $0.30 input and $2.50 output per million tokens at the time of writing. That output figure covers text; images are billed at the image rate, which is why per-token listings and per-image costs rarely reconcile at a glance.

Picking the successor

Google's deprecations page names Gemini 3.1 Flash Lite Image (Nano Banana 2 Lite) as the replacement, and the pricing page also points to Gemini 3.1 Flash Image (Nano Banana 2). They solve different problems:

• Flash Lite Image is the like-for-like swap. It outputs at 1K only, which matches the 1024px images 2.5 Flash Image was priced on, and Google lists a 1K image at $0.0336, slightly below the old $0.039. It adds a thinking mode with minimal and high levels.
• Flash Image is the upgrade. It adds 2K and 4K output, search grounding and extreme aspect ratios, at $0.067 for a 1K image. Google deprecated it in October 2026 in favour of a 2.1 update, so check the current lineup before choosing it as a destination.

If you used 2.5 Flash Image for thumbnails, stickers or product shots at 1024px, start with Lite. If you were working around its text rendering or resolution limits, Flash or the newer update is where those limits were addressed.

What to retest when you migrate

A successor model with the same nickname will not produce the same images. Before switching traffic, check the behaviours your product depends on without saying so:

1. Aspect ratios. Confirm every ratio you request is supported, and that output dimensions match what your layout expects.
2. Edit fidelity. Run your twenty most common edit prompts. A newer model may "improve" parts of the image you asked it to leave alone.
3. Character consistency across turns. If your product keeps a character across a session, test a full session, not single edits.
4. Refusals. Safety filtering differs between generations. Prompts that worked may now be refused, and the reverse.
5. Response shape. If thinking is on by default in the successor, check your parser handles any extra parts in the response.

Model deprecation and migration covers running old and new side by side, and regression testing a model upgrade covers turning those checks into a repeatable suite.