Google's Gemini 2.5 Flash Image is the original Nano Banana model for generating and editing images through an API. As of October 4, 2026, the Gemini Developer API deprecation schedule lists October 2, 2026 for this model. The table also describes its dates as the earliest possible retirement dates, so the calendar alone does not prove that every project's requests have stopped working. Applications still calling gemini-2.5-flash-image should check their actual responses and prepare to migrate now.
For cost-sensitive, single-turn 1K generation, evaluate Nano Banana 2 Lite first. For multi-turn editing or higher resolutions, evaluate Nano Banana 2 first, following the official image generation guide. Get a minimal request working, replay existing business examples, and then switch traffic. Do not copy a preview replacement ID directly from the retirement table or rerun a backlog before validating the new path.
This guide covers the Gemini Developer API. The Gemini web app and models on Google Cloud have their own lifecycle, permissions, and billing considerations; the same retirement table does not necessarily apply. If you are still choosing an image service, start with our AI image generation guide.
Why you should not copy the replacement ID directly
As of October 4, 2026, Google's deprecation table contains a replacement chain worth following to its end. It points gemini-2.5-flash-image to gemini-3.1-flash-image-preview, but the same page lists June 25, 2026 as that preview's retirement date and names gemini-3.1-flash-image as its replacement. The current model page also labels the version without the preview suffix as Stable.
For a migration to Nano Banana 2, therefore, evaluate gemini-3.1-flash-image. This choice comes from cross-checking the retirement chain against the current model page. Replacing the model string alone is not a compatibility test.
Check three places in your application:
- Request code and environment configuration: which model ID is actually sent?
- Queues, scheduled jobs, and retry jobs: was the model name saved into existing task records?
- Proxies and gateways: which model and service endpoint does a displayed alias resolve to?
Changing startup configuration may not change tasks already in a queue. Identify records containing the old model and decide how to handle them before resuming the queue.
Choosing between Nano Banana 2 and Lite
Evaluate Nano Banana 2 first if you need to preserve multi-turn editing, 2K/4K output, or search-assisted generation. If you only need single-turn 1K images, Lite can be the first candidate. The following reflects official capabilities as of October 4, 2026. It is a candidate-selection guide, not a ranking of image quality, speed, or success rates.
| Your requirement | First candidate to evaluate | What to check |
|---|---|---|
| Single-turn 1K illustrations with output cost as a priority | gemini-3.1-flash-lite-image |
Lite supports only 1K; do not migrate other 2K/4K requests in your application along with it |
| Repeated edits, or 0.5K, 2K, or 4K output | gemini-3.1-flash-image |
Replay the full editing sequence and check that later changes preserve previously approved details |
| Generation informed by Google Search | gemini-3.1-flash-image |
The model page lists Search grounding support; Lite does not support it |
See the image generation guide for resolution and multi-turn editing limits, and the Nano Banana 2 and Lite model pages for search capabilities. Google explicitly states that Lite is not optimized for multiple reference inputs or successive multi-turn edits. The shared Nano Banana name does not make these models interchangeable.
If your existing workflow depends on complex reference inputs, preserve the number of reference images, their order, and the prompt while testing the same task. Changing the model, assets, and prompt together makes failures harder to diagnose.
Validate a new request in a separate script
Confirm that billing, model access, and your existing API key work in a test project before sending a single request. Google currently recommends the Python google-genai SDK; installation instructions are in the official SDK documentation. Supply the key through an environment variable or your existing secret-management system. Never put it in the script, a frontend bundle, or logs; see the API key documentation.
Install the SDK in a separate virtual environment rather than upgrading production dependencies directly:
python -m venv .venv
# macOS / Linux; on Windows, use the activation command for your environment
source .venv/bin/activate
python -m pip install -U google-genai
The following minimal illustrative script uses the Interactions API and is adapted from the official image generation examples available on October 4, 2026. Save it as migration_check.py, make sure the process has access to GEMINI_API_KEY, and run python migration_check.py. We have not executed a paid API call for this example, and it comes with no measured output or compatibility guarantee. Running it incurs API charges.
import base64
import os
from pathlib import Path
from google import genai
client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])
result = client.interactions.create(
model="gemini-3.1-flash-image",
store=False,
input="Generate a blue ceramic mug on a white background, with no text, logos, or other objects.",
response_format={
"type": "image",
"mime_type": "image/png",
"aspect_ratio": "1:1",
"image_size": "1K",
},
)
image = result.output_image
if image is None or not image.data:
raise RuntimeError("No image was available to save; inspect the response status and diagnostics")
path = Path("migration-check.png")
path.write_bytes(base64.b64decode(image.data, validate=True))
print(f"Saved {path}; check that it opens and that its dimensions and content meet your requirements")
The example uses store=False to disable Interactions conversation-state storage. As of October 4, 2026, the official API overview states that requests are stored by default so a conversation can continue using previous_interaction_id. With storage disabled, you cannot rely on that server-side history. This setting does not disable all service logs or waive data-processing terms; check the applicable data policy before using sensitive material.
If your application uses GenerateContent, do not paste these request and response fields directly into its existing parser. Google still supports GenerateContent. Use the script as an independent connectivity check, then decide whether to validate the new model on your existing API or migrate APIs separately. Record and validate the model migration and API migration separately.
As of October 4, 2026, the official API limitations state that Interactions does not support the Batch API. Keep existing batch jobs on the GenerateContent-based Batch API path, and validate their request format and response parsing separately. Do not treat this single-request example as a batch replacement. Small-batch regression testing below means a group of application-level test tasks; it does not imply that the Batch API has been used or validated.
Once the minimal request works, change the output path to a test-only directory and try one real, non-sensitive business input. Saving a file only proves that this request produced an image. It does not validate dimensions, content, downstream uploads, or display behavior.
Replay existing tasks instead of judging one sample
Sample business assets that you are already authorized to use for testing, covering at least the entry points your application actually supports. If you have no historical outputs, write down acceptance criteria first rather than inventing an old-model baseline. The following is a reusable test record, not a set of measurements made by this site.
Task ID:
Model ID / SDK version / API name:
Prompt version / input asset IDs:
Aspect ratio / resolution / response format:
Request count / failure reason / retries:
Input and output usage / actual charges:
Text, objects, or brand elements that must be preserved:
Acceptance result / manual corrections:
| Existing application capability | Task to replay | If it fails |
|---|---|---|
| Text-to-image | A common prompt and an edge-case input | Check the model response, image decoding, and business rules separately |
| Reference-image editing | One authorized asset with a single clearly specified change | Compare regions that should remain unchanged; retain a manual editing path if changes are unacceptable |
| Successive edits | A real sequence, such as changing the background and then a color | Verify context handling; do not treat independent requests as a continuous conversation |
| Images containing text | A title, price, or label that the application actually needs | Check every character; move text to a precise layout tool if necessary |
| Batch processing | A small batch including a failed task | Check whether individual failures can be recovered without duplicate saves or deliveries |
An edge-case input means something your business permits but that may expose integration problems: an image near your application's upload limit, a long prompt, or a task missing optional fields. It does not mean a test designed to bypass safety restrictions.
Define your acceptance criteria before judging results. Product images may need to preserve packaging text, while an article illustration may only need the correct aspect ratio and subject. These tasks cannot share a single “looks good” standard. Do not automatically deliver unchecked text, branding, or client assets.
Budget for the complete task, not just the image price
As of October 4, 2026, Google's Gemini Developer API pricing page lists approximate image-output prices for the Standard paid tier: $0.067 per Nano Banana 2 1K image and $0.0336 per Lite 1K image, in US dollars. Neither API model has a free tier. These figures cover image output only; they exclude input, text/thinking output, retries, and applicable search charges.
For illustration, the output portion alone for 1,000 1K images would be about $67 or $33.60. That multiplication excludes failures, rework, and manual corrections, so it is not the total cost of delivering 1,000 accepted images.
For the same batch of tasks, record full API charges, the number of accepted images, and manual correction time. Compare costs only between candidates that pass your business acceptance checks. A lower output price cannot compensate for a missing required capability. Start with a small batch and application-side spending limits; do not resume the full backlog before understanding retry rates.
Decide whether an error needs a configuration fix or a retry
An unavailable model, incompatible parameters, missing billing setup, and temporary rate limits require different responses. The following sequence is based on the official troubleshooting guide as of October 4, 2026. Use the actual response to determine the cause.
- Model or API mismatch: check the stable model ID, service endpoint, API version, and method. Do not switch to another preview simply because its name looks similar
- Invalid parameters: return to the minimal request, then restore the aspect ratio, resolution, references, and other options one at a time to find what exceeds the candidate model's capabilities
- Access or billing issues: check the project, key permissions, and balance. Repeating the same request does not repair configuration
- 429 or temporary service errors: use capped exponential backoff when the error is retryable, and inspect the SDK's existing retry behavior to avoid adding unlimited application-level retries on top
- No image in the response: inspect the status, diagnostics, and parsed fields. Do not save text as an image or assume that every response without an image is a service failure
Before switching production traffic, confirm that:
- Model IDs in code, gateways, queues, and scheduled tasks have been checked
- The new model has been validated in the actual project and service endpoint you will use
- Existing editing, format, and failure-recovery paths have test records
- Billing scope, attempt counts, and business acceptance results have been recorded
- A failure fallback is in place: pause tasks, return a clear message, or use another validated candidate
If the old model is already unavailable, rolling back code will not restore the retired service. Keeping the previous adapter can help diagnose problems, but a production fallback must point to an available, validated path.