Gemini Shutdown Dates This Fall: Omni Preview, the Original Nano Banana, and Gemini 2.5

The next Gemini shutdown dates arrive fast: gemini-omni-flash-preview turns off on September 30, 2026. The original Nano Banana model, gemini-2.5-flash-image, follows on October 2. Behind them sit Gemini 2.5 Pro, Flash and Flash-Lite, which Firebase lists as closing on October 20 on Google Cloud’s Agent Platform. The Gemini API’s own deprecations page lists no shutdown date for those three at all.

Update (2026-09-29): Google changed the Firebase AI Logic models page after we drafted this post. It no longer lists October 16 for Gemini 2.5 Pro, Flash and Flash-Lite on the Gemini Developer API. That column now reads "To be determined," with only previously active projects able to keep using the models. The October 20 date for the Agent Platform Gemini API (formerly Vertex AI) is unchanged, as are the Omni preview (September 30) and gemini-2.5-flash-image (October 2) dates. We have updated the dates below.

This piece covers the Gemini shutdown dates for each closing model ID, the replacement Google names, and where Google’s pages disagree. It ends with a way to find hardcoded model IDs. That covers WordPress and Drupal sites, chatbots, image pipelines and automations. One smaller item rides along: the managed agent antigravity-preview-05-2026 shuts down on October 5, replaced by antigravity-preview-09-2026.

The short version on the Gemini shutdown dates: move Omni and image work this week. For image generation, ignore the replacement printed on the deprecations page, because that model was shut down in June. For Gemini 2.5 text models, plan around October 20 if you call them through Google Cloud. On the Gemini Developer API, Google now lists no date, but new projects are already shut out, so move on the same schedule.

The Gemini shutdown dates at a glance

All dates below come from Google pages read on September 27, 2026. The Gemini API deprecations page was last updated September 24. The Firebase AI Logic models page carries the October dates for Gemini 2.5.

The deprecations page adds one qualifier worth knowing. It describes the Gemini shutdown dates it lists as the earliest possible dates, with exact dates to be communicated in advance. Treat every date in the table as the day a request can start failing.

September 30: the Gemini Omni Flash preview closes

Google released gemini-omni-1.1-flash as generally available on August 27, according to the Gemini API changelog. The same entry set September 30 as the end date for the preview endpoint. The deprecations page confirms both the date and the replacement.

The GA model adds features the preview did not have. The changelog lists video extension and interpolation between images. It also adds a resolution setting from 360p to 4K, with 720p as the default. The current Omni video guide uses only the new model ID in its code samples.

For most integrations the change is a string swap. Test it anyway. Video output is expensive to regenerate. Any default that changed between preview and GA will show up in every clip you produce. Check the resolution and aspect ratio your code sends. Then confirm large outputs still come back the way your code expects.

Our explainer on what Gemini Omni Flash is covered the preview. For anyone who built on it, this is the first maintenance task. Preview models carry exactly this risk. Firebase describes preview models as retired within a few weeks or months after their stable version ships.

One note for Google Cloud users. The Agent Platform documentation has its own Omni pages. At the time of writing, its Omni 1.1 Flash page still showed a Preview label. The September 30 date above is for the Gemini API. If you call Omni through Google Cloud, confirm the date for that platform in your console.

October 2: the original Nano Banana shuts down

gemini-2.5-flash-image is the model most people met as Nano Banana. Google released the stable version on October 2, 2025. Both the deprecations page and the Firebase page set its shutdown for October 2, 2026. That is a one-year lifespan for a stable model, which is worth remembering when you pick its successor.

Our explainer on what Nano Banana is covers the model family. The family now has four members on the Gemini API models page:

• Nano Banana: gemini-2.5-flash-image, the one shutting down.
• Nano Banana 2: gemini-3.1-flash-image, generally available since May 28, 2026.
• Nano Banana Pro: gemini-3-pro-image, also generally available since May 28.
• Nano Banana 2 Lite: gemini-3.1-flash-lite-image, generally available since June 30.

Image pipelines are the likeliest place for this ID to hide. Featured-image generators, product-shot tools and social card builders can set the model once and never touch it again.

The deprecations page names a replacement that is already gone

Here is the trap. The deprecations page lists gemini-3.1-flash-image-preview as the replacement for gemini-2.5-flash-image. The same page lists that preview as shut down on June 25, 2026, replaced by gemini-3.1-flash-image. Copy the replacement ID from the table, and you move from a model with days left to one that stopped working in June.

Google’s pages name three different successors:

• The deprecations page names gemini-3.1-flash-image-preview, which no longer runs.
• The Firebase AI Logic page tells readers to update to a newer model such as gemini-3.1-flash-image.
• The image generation guide calls Gemini 2.5 Flash Image the "legacy pioneer" of the series. It recommends a move to Nano Banana 2 Lite, gemini-3.1-flash-lite-image. The stated reasons are better quality, faster generation and lower API pricing.

Choosing between Nano Banana 2 and Nano Banana 2 Lite

The image generation guide draws the lines. Nano Banana 2 Lite is described as the fastest and cheapest Gemini image model. It outputs at 1K only and does not support Google Search grounding. Nano Banana 2 offers 512px, 1K, 2K and 4K output and supports search grounding. Nano Banana Pro is aimed at complex layouts and precise text rendering.

A reasonable rule: if your pipeline makes 1K images from text prompts, Lite is the closer match on cost. If it needs higher resolutions or search grounding, use Nano Banana 2. Either way, run a sample of your real prompts through the new model first. Compare the output side by side before you switch production traffic. Check current prices on Google’s pricing page rather than trusting any figure you saw months ago.

Gemini 2.5 Pro, Flash and Flash-Lite: two official answers

Of all the Gemini shutdown dates, this one affects the most teams. Google has published two answers.

On September 18, the Gemini API changelog limited access to the 2.5 models. Only users who have actively used them in the past keep access. New projects are pointed to Gemini 3.8 Flash or 3.5 Flash-Lite instead. That is a quiet closure: an existing key keeps working, but a new project cannot adopt the models.

What the Gemini API deprecations page says

The deprecations page lists all three 2.5 text models with "No shutdown date announced." A note on the same page says these models are not deprecated and will continue to be served until further notice. It repeats the advice to use 3.5 Flash-Lite or 3.8 Flash for new projects.

What Firebase AI Logic says

The Firebase AI Logic models page tells a different story. A warning near the top says the Gemini 2.5 models will shut down in October 2026. It adds that new projects cannot start using them. As of September 29, its table splits the date by back end. It gives "To be determined" for the Gemini Developer API and October 20 for the Agent Platform Gemini API. Agent Platform is Google’s current name for Vertex AI.

Firebase AI Logic is a client SDK layer over the same two back ends, not a separate model host. Its dates describe the same models. If your app uses Firebase, our guide to what Firebase is explains where AI Logic sits.

Planning around conflicting Gemini shutdown dates

Plan for October 20. It is the only firm date on any Google page, and it applies to Agent Platform. Move Developer API projects on the same schedule. The September 18 access change also shows the direction of travel. If the models keep running past that date, you lose nothing by having moved. If they stop, you avoid an outage in a production chatbot or content tool. Google Cloud customers should also check the console and billing email for notices tied to their project.

One more source circulates. A Google Cloud customer notice, reposted in a public GitHub issue, describes a later staggered shutdown for Agent Platform running into 2027. We could not find those dates on any Google documentation page. Treat them as unconfirmed unless Google sends the same notice to your own account.

Moving from Gemini 2.5 to 3.8 Flash or 3.5 Flash-Lite

Google names two targets: gemini-3.8-flash, generally available since September 2, and gemini-3.5-flash-lite, generally available since July 21. Flash is the general replacement. Google’s changelog describes Flash-Lite as a low-latency, cost-effective option for high-volume automation. Our coverage of Gemini 3.7 Flash covers the introductory pricing on the previous Flash release.

A model string swap is rarely the whole job. Three documented changes can break a straight swap:

• Sampling parameters. The July 21 changelog entry says temperature, top_p and top_k are now deprecated. It does not say whether requests that send them are ignored or rejected. Test with and without them.
• Temperature defaults. The Gemini 3 developer guide recommends keeping temperature at its default of 1.0. It warns that lower values can cause looping or degraded performance. If your 2.5 code sets a low temperature for consistent output, revisit it.
• Thinking controls. The same guide says thinking_budget still works for backward compatibility but recommends thinking_level. A request cannot include both.

Output also changes in ways no parameter captures. Length, formatting and tone shift between model generations. Does your plugin parse model output, such as JSON for a form? Test the parser against real responses.

Before the Gemini shutdown dates: find hardcoded model IDs

A model ID set once in a settings screen is easy to forget. Before the Gemini shutdown dates arrive, search every place a model string can live.

Search your code. Run this from the root of each repository:

grep -rnE "gemini-2\.5-(pro|flash|flash-lite|flash-image)|gemini-omni-flash-preview|gemini-3\.1-flash-image-preview|antigravity-preview-05-2026" .

The pattern includes gemini-3.1-flash-image-preview on purpose. Anything still using it is already broken, and it is the ID a hurried migration might introduce.

Search configuration and the database. Model IDs often live outside code:

• Environment files and secrets managers.
• WordPress plugin settings, which are usually stored in the wp_options table. A database search for the strings above will find them.
• Drupal configuration. Run a configuration export and search the exported YAML.
• Serverless functions, scheduled jobs and CI variables.

Check vendor plugins and automations. Some AI plugins and no-code tools store the model as a dropdown on a settings page. Open each one and look. If the vendor hardcodes the model, ask them for their migration date in writing. For client sites, a quick inventory of which sites call which model is worth making now.

Check prototypes that became production. An app built in AI Studio keeps the model ID chosen on the day it was built. See our overview of Google AI Studio for where that setting sits. The guide to moving from AI Studio to Vertex AI covers the production side.

A migration order for the next three weeks

Work through the Gemini shutdown dates in order. Test each change before the day it matters:

• By September 30: move Omni video calls to gemini-omni-1.1-flash.
• By October 2: move image generation to gemini-3.1-flash-image or gemini-3.1-flash-lite-image. Do not use the ID printed on the deprecations page.
• By October 20: move Gemini 2.5 text calls to 3.8 Flash or 3.5 Flash-Lite. Run the parameter checks above.

The OpenAI Assistants API shutdown in August was the same kind of deadline on another platform. The work is the same: find the string, swap it, test it.

Frequently Asked Questions