Google Imagen 4 API Shutdown: Migration Checklist for Marketers

Google’s current Gemini API deprecations table lists three stable Imagen 4 model IDs for shutdown on August 17, 2026, and names gemini-nano-banana-2.1 as the replacement for each. The date has passed, but Google says its listed shutdown dates are the earliest possible dates and that it will communicate the exact date to users. The public table therefore does not confirm whether a particular project’s endpoint is already unavailable. If an integration still contains an Imagen 4 ID, treat it as urgent and check the notice for the affected project. Google’s deprecations table

The migration also changes the request and response contract. Google’s current image-generation guide shows a Gemini Developer API call through the Interactions API, with image output returned as base64 data. That is more than a model-name edit: update the API route, request fields and image parser, then compare representative outputs in staging. Current Gemini API image-generation guide

What Google scheduled for August 17

Google announced the Imagen 4 deprecation in its June 15 release notes. The current lifecycle table lists these three stable Gemini Developer API IDs, the same August 17 date and the same replacement. It lists separate dates for preview IDs, so do not mix those into this migration. Gemini API release notes · Gemini API deprecations

Google defines shutdown as an endpoint becoming completely unavailable. Its “earliest possible date” caveat means the table does not identify the precise time an individual project lost access. It also does not tell you whether a workflow’s fallback is still calling an old ID. Use the project’s own notice and logs to establish account status; do not infer it from the public date alone.

Why this matters to marketing and creative operations

A model ID can be hidden in a CMS image helper, a scheduled job, a workflow builder, an environment variable or a fallback. The person requesting a campaign asset may only see a generic failure, a missing image or a job that keeps retrying. Search every environment and saved workflow, including inactive projects and disaster-recovery paths, rather than relying on the main application’s settings screen.

A replacement model can also change how an existing prompt behaves. Text in an image, composition, aspect ratio, editing instructions, safety handling and output parsing all need review. A separate image-model review makes the same practical distinction: a model name does not prove that an accepted creative brief will produce an equivalent result. Keep your own approved outputs and human review decisions as the comparison baseline.

What Google recommends now—and what remains unclear

The current Gemini API deprecations table names gemini-nano-banana-2.1 for all three stable Imagen 4 IDs. Google announced Nano Banana 2.1 as generally available on October 6, describing it as an update to Gemini 3.1 Flash Image with changes to visual quality, prompt adherence, text rendering, multi-turn consistency and wide aspect ratios. Those are Google’s product claims, not an independent comparison with your Imagen 4 outputs. October 6 release note · Nano Banana 2.1 model page

The older replacement conflict in the Gemini Developer API material has changed: the current lifecycle table points to Nano Banana 2.1, while the pricing page now lists Gemini 2.5 Flash Image as a separate model with its own retirement notice. Do not carry forward the old instruction to choose Gemini 2.5 as the Imagen 4 replacement. Current Gemini API pricing and lifecycle notices

Check which Google surface your application actually calls before copying a migration example. The current Gemini Developer API image guide demonstrates the Interactions API. The Firebase AI Logic migration page says the Imagen deprecation and shutdown apply across Google for both the Gemini Developer API and Agent Platform Gemini API (formerly Vertex AI). Its provider-specific mapping still sends Imagen 4 Fast and Standard to Gemini 3.1 Flash Image with different thinking settings, and Ultra to Gemini 3 Pro Image. Use the guide and SDK for the surface your app actually calls; the code below covers only the Gemini Developer API. The reviewed API docs do not establish the same retirement status for Google Ads creative tools, the AI Studio interface or consumer Gemini. Firebase AI Logic Imagen migration guide · Gemini Developer API image guide

None of Google’s replacement labels promises output parity. The destination model, API request, image settings, safety behavior, limits and price need to be checked against the workflow that will use them.

A practical migration checklist

1. Inventory every place the model ID can live

Search application code, deployment variables, workflow tools, scheduled jobs, prompt presets, retry queues and fallback configuration for all three exact Imagen 4 IDs. Record the calling service, owner, project, environment and last successful request. Do not put API keys, project secrets or customer prompts in the inventory.

2. Preserve a representative output baseline

Choose accepted examples from the work the system actually produces: product scenes, social crops, blog images, text-heavy layouts, brand colors, people and less common aspect ratios. Keep the prompt, relevant settings, accepted output and reviewer decision together. Include editing and reference-image paths separately when the workflow uses them. Existing generated assets do not need to be regenerated just because an API model is retired.

3. Update the API call and image parser

The current Gemini Developer API image guide shows client.interactions.create(), the /v1beta/interactions REST route, a response_format object and image data under output_image.data. The following Python example follows that documented shape for an image-only 2K response. It uses Google’s current GenAI Python SDK (google-genai); provide GEMINI_API_KEY to a server-side process with access to the intended project, and keep the key out of browser or mobile code. Google lists no free tier for Nano Banana 2.1, so confirm paid-tier access and billing before using it. See the official SDK installation guide, API-key and security instructions and current model pricing. This is a documentation-based reference and has not been run against a Google project.

import base64
from pathlib import Path
from google import genai

client = genai.Client()
response = client.interactions.create(
    model="gemini-nano-banana-2.1",
    input="Create a clean studio background with an empty central display area, soft diffused lighting, neutral colors, no text or logos.",
    response_format={
        "type": "image",
        "mime_type": "image/png",
        "aspect_ratio": "16:9",
        "image_size": "2K",
    },
)

if response.output_image is None or response.output_image.data is None:
    raise RuntimeError("The interaction returned no image data")

Path("generated.png").write_bytes(
    base64.b64decode(response.output_image.data)
)

Add bounded error handling around the call and confirm that your installed SDK and selected API surface match the current guide. The sample does not test model availability, prompt quality or production compatibility.

4. Compare outputs in a non-production path

Run the preserved prompt set in staging and review more than one attractive result. Compare accepted-output rate, substantive edits, text accuracy, aspect ratios, latency, refusals, retries and image parsing. Separate Google’s documented features from what your own sample shows. If the output looks polished but misses the brief or brand, the technical request can succeed while the migration remains operationally incomplete.

5. Recheck pricing before moving volume

Google’s current pricing page lists these image-output equivalents for Nano Banana 2.1. Text, input and thinking tokens can add cost; the page lists no free tier for this model.

These are current Nano Banana 2.1 price-page equivalents, not an Imagen 4 before-and-after savings estimate. There is a narrow 4K documentation mismatch: the pricing page says a 4K output consumes 3,780 image tokens, while the image-generation guide’s resolution table lists 2,520 for 4K. The 1K and 2K token counts match. For planning, use the explicit per-image prices on the pricing page and confirm the billable usage before a large 4K batch. If you enable Google Search grounding, the pricing page meters the search requests separately. Nano Banana 2.1 pricing · Image size and token table

6. Handle errors by type

The Interactions API returns an error code and message. Use those fields to distinguish a retired or mistyped model from a malformed request, an access problem or temporary capacity pressure. Gemini API error reference

Google recommends bounded exponential backoff for transient errors such as 429 and 503, and says not to retry client errors such as 400, 402 or 403 unchanged. Its troubleshooting guide also explains that SDKs may retry some transient failures automatically; avoid adding a second unbounded retry loop around that behavior. Gemini API troubleshooting guidance · Rate limits

7. Recheck provenance, rights and rollback

Keep the prompt and source-asset record, verify that reference images are cleared for the intended use, and log which model produced each final asset. Review the current provider terms, watermark or provenance behavior and any channel disclosure that applies. The AI content provenance checklist covers broader governance; this API migration needs the same traceability at the asset level.

Move the model setting into a controlled configuration, update retries so a retired ID cannot loop indefinitely, and alert on model-not-found responses, empty image output and unexpected fallback use. Keep the last working configuration and accepted visual baseline for diagnosis. The production-controls guide explains the wider role of traces and bounded retries.

What this update does not establish

• It does not confirm the shutdown time or remaining access for a particular Google project. Check that project’s own notice and request logs. The cited pages cover Imagen API endpoints; they do not establish the same retirement status for Google Ads creative tools, the AI Studio interface or consumer Gemini.
• It does not prove that Nano Banana 2.1 will match the quality, latency, limits, safety behavior or output format of an accepted Imagen 4 workflow.
• It does not provide a current Imagen 4 price or a valid savings comparison. Use the current model’s pricing and your own workload.
• It does not say that image files already generated and stored by a customer stop working. The lifecycle notice concerns API models; asset use still depends on the owner’s rights and channel policies.
• The code above is documentation-based and unexecuted. No endpoint, account notice or migration test was checked for a customer project.

A migration action plan

1. Check the affected project’s deprecation notice and search every direct call, fallback and saved workflow for the three Imagen 4 IDs.
2. Adapt the request and response handling to the Gemini API surface your application uses.
3. Run the representative prompt set in staging and review output, error behavior, quotas and current prices.
4. Move production traffic only after the technical owner and creative owner accept the new outputs; then watch for retired IDs, empty responses and fallback loops.

If the old endpoint still responds, use the remaining time to finish the migration rather than treating availability as an extension. If it fails, verify the error and move to the smallest supported path you have tested.

Frequently asked questions

Which Imagen 4 model IDs are affected?

Google lists imagen-4.0-generate-001, imagen-4.0-ultra-generate-001 and imagen-4.0-fast-generate-001 in the Gemini API deprecations table, with August 17, 2026 as the earliest possible shutdown date shown.

Does the August 17 date prove every project has shut down?

No. Google says the table gives earliest possible retirement dates and that it communicates exact dates to users. The public table does not establish a particular project’s access state.

Which model does Google name as the replacement?

The current Gemini API lifecycle table names gemini-nano-banana-2.1 for all three stable Imagen 4 IDs. That label is migration guidance, not proof of output parity. Use a product-specific guide if your integration calls Firebase AI Logic or Agent Platform.

Can I change only the model string?

Do not assume so. The current Gemini Developer API image guide uses the Interactions API request and reads image bytes from output_image.data. Update the route, request fields and parser for your SDK, then confirm the result in staging.

How much does Nano Banana 2.1 cost?

The current pricing page lists image-output equivalents of $0.0336 at 1K, $0.0504 at 2K and $0.113 at 4K on Standard; Batch lists $0.0168, $0.0252 and $0.0567. The official image guide and pricing page disagree on the 4K token count, so verify actual usage before scaling.

Does this affect images already generated?

The shutdown notice is about API model endpoints. It does not say that image files already stored elsewhere become unavailable; their ongoing use depends on your rights and storage policies.

Should a marketer handle the migration alone?

The creative owner should define the accepted visual baseline and review outputs. The technical owner should update the API call, fallback, error handling and monitoring. Rights, provenance and channel disclosure may need a separate review.