AI DevelopmentPlaybook4 min readPublished September 25, 2026

October 2 on the Gemini API · March 15, 2027 on Vertex · same model, two calendars

Gemini 2.5 Flash Image Retires October 2 on the API: Act Now

The Gemini API shuts down gemini-2.5-flash-image on October 2, 2026; Vertex says March 15, 2027. Which date applies, three replacements priced, and a test plan.

DA
Digital Applied Team
Research and practical guidance
PublishedSeptember 25, 2026
API shutdownOctober 2, 2026

If your code calls gemini-2.5-flash-image through the Gemini API, Google's deprecations page says it stops working on October 2, 2026, one week from today. If the same code calls it through Vertex AI, Google's model-versions page says March 15, 2027. Both pages were updated on September 24 and neither mentions the other. The replacement each page names is different, and the one on the API page was itself shut down in June.

We are not going to guess which page is right. This post sets out what each says as of September 25, tells you which one governs the surface you use, prices the three live replacements per image from Google's own pricing page, and gives a test plan that fits inside the week you have.

Key takeaways
  1. 01
    Gemini API callers have until October 2. Vertex callers have until March 15, 2027.Google's two lifecycle pages disagree by 164 days. The Vertex page states its dates are never brought forward; the Gemini API page states its dates are the earliest a model might be retired.
  2. 02
    The API page's named replacement no longer exists. Use the pricing page's pair instead.gemini-3.1-flash-image-preview shut down on June 25. The pricing page's warning names Gemini 3.1 Flash Image or Gemini 3.1 Flash Lite Image; the Vertex page names the Lite model.
  3. 03
    The cheapest replacement is cheaper than the model it replaces.A 1K image costs $0.0336 on 3.1 Flash Lite Image against $0.039 on 2.5 Flash Image, because the newer model bills 1,120 output tokens per image rather than 1,290 at the same $30 per million.
  4. 04
    What you lose on the Lite model is size, not price: 1K only.If anything downstream expects 2K or 4K, the choice is 3.1 Flash Image at $0.067 per 1K image or Nano Banana Pro at $0.134. The test plan below checks that before the prompts.

01 — The pagesFour Google pages, three different answers

All four rows below were read on September 25, 2026. They are the Gemini API deprecations page, the Gemini API pricing page, the Vertex AI model versions page and the Gemini API changelog.

Google's own pages, read September 25, 2026. The two lifecycle pages both carry a September 24 last-updated date.
PageDate givenReplacement namedNote
Gemini API deprecations pageOctober 2, 2026gemini-3.1-flash-image-previewThe named replacement was itself shut down on June 25, 2026, per the same page. Page last updated September 24.
Gemini API pricing page (warning box)October 2, 2026Gemini 3.1 Flash Image or Gemini 3.1 Flash Lite ImageThe one page of the four that names two live replacements.
Vertex AI model versions pageMarch 15, 2027gemini-3.1-flash-lite-imagePage last updated September 24. States that retirement dates may be extended but never brought forward.
Gemini API changelog, September 18"Until further notice"3.5 Flash-Lite or 3.8 Flash for new projectsSays access to "the 2.5 models" is limited to users who have actively used them, and that they are not deprecated. Does not name the image model.

The two lifecycle pages describe their own dates differently, and the difference matters. The Gemini API page says its shutdown dates are the earliest possible dates on which a model might be retired. The Vertex page says retirement timelines may be extended but will not be moved earlier. Read together, October 2 is a floor for API users and March 15, 2027 is a floor for Vertex users. On the API, October 2 is not a promise that the model runs until then: the September 18 changelog entry restricting 2.5 models to existing active users shows Google is already managing capacity on that generation.

The stale replacement on the API page is the detail to notice. A team that follows the page literally will point code at a model ID that returns an error. Our alias and retirement ledger tracks this class of problem across vendors: the retirement date is maintained, the replacement pointer is not.

02 — The decisionWhich date applies to you

You call ai.google.dev endpoints with an AI Studio API key
Plan for October 2, 2026. The deprecations page governs this surface. Migrate this week; the September 18 changelog also limits "the 2.5 models" to users who have actively used them, which may already cover this one.
Gemini API
You call the model through Vertex AI or the Enterprise Agent Platform
Plan for March 15, 2027, and treat it as a floor. The Vertex page names gemini-3.1-flash-lite-image as the replacement. There is no reason to wait: the replacement is live, cheaper per image and on the same surface.
Vertex AI
You reach it through a third-party model router
Follow the Gemini API date. Router catalogs carry their own expiry metadata; we observed one drop its March 15, 2027 expiry on this route on September 25 at 08:01 UTC. That is catalog housekeeping, not a Google decision.
Aggregator
You do not know which surface a pipeline uses
Check the request host and the credential. An AI Studio key against generativelanguage.googleapis.com is the Gemini API; a service account against a regional aiplatform endpoint is Vertex. Assume October 2 until proven otherwise.
Both

03 — The replacementsThe three live replacements, priced per image

Google prices image output in tokens, and each model consumes a fixed token count per image at a given size. The per-image figures below are the equivalents Google prints on its pricing page for a 1K image on the standard tier, read September 25. Batch tiers halve them (Flex, offered only on 2.5 Flash Image and Nano Banana Pro, matches Batch); the Priority tier on 2.5 Flash Image is $0.0702.

Standard-tier price for one 1K image, Gemini API

Google's Gemini API pricing page, per-image equivalents as printed, read September 25, 2026. Bars scaled to Nano Banana Pro at $0.134.
Gemini 3 Pro ImageNano Banana Pro, 1K or 2K
$0.134
Gemini 3.1 Flash ImageNano Banana 2, 1K
$0.067
Gemini 2.5 Flash Imageretiring, up to 1024 px
$0.039
Gemini 3.1 Flash Lite ImageNano Banana 2 Lite, 1K only
$0.0336
Source: Google's Gemini API pricing page and model cards, read September 25, 2026. Standard tier, US dollars.
ModelPer 1K imageSizes and tokensNotes
Gemini 2.5 Flash Image (retiring)$0.039Up to 1024 px, 1,290 output tokensInput $0.30 per M tokens. No thinking. Batch supported. Output limit 32,768 tokens.
Gemini 3.1 Flash Lite Image$0.03361K only; 2K and 4K unsupported. 1,120 output tokensInput $0.25 per M tokens (text, image, video, PDF). Thinking minimal or high. Batch at half price. Output limit 4,096 tokens.
Gemini 3.1 Flash Image (Nano Banana 2)$0.0670.5K $0.045 · 1K $0.067 · 2K $0.101 · 4K $0.151Input $0.50 per M tokens. Grounding with Google Search available. Batch at half price.
Gemini 3 Pro Image (Nano Banana Pro)$0.1341K and 2K $0.134 · 4K $0.24Input $2.00 per M tokens, about $0.0011 per input image. Priority tier at $216 per M image tokens.

For most pipelines that used 2.5 Flash Image for what it was good at, fast 1K images at volume, the Lite model is the direct successor: about 14% cheaper per image, with a lower input rate too. Our post on Nano Banana 2 Lite covers what Google said about it at launch. A pipeline that needs 2K or 4K output, or search grounding, moves up to Nano Banana 2 at roughly 1.7 times the old per-image price, and only a pipeline that needs the Pro model's quality justifies 3.4 times.

04 — The differencesWhat changes besides the price

Four things differ between the retiring model's card and the Lite model's card, and each is a place a migration can break quietly.

Output size2.5 Flash Image produced images up to 1024 px. The Lite card says 1K only and that 2K and 4K are unsupported. Same ceiling in practice; test how your client handles a 2K or 4K request against the new model before cutover.
1K only
Output tokens per image1,290 on the old model, 1,120 on the new. Any cost model or quota calculation keyed to the old figure over-estimates by 15%.
1,290 → 1,120Cheaper per image
ThinkingNot supported on 2.5 Flash Image. Supported on the Lite model at minimal or high, and thinking tokens are billed as text output at $1.50 per million. For prompt-parity tests, start at minimal.
New
Token limits and inputsBoth accept 65,536 input tokens. The Lite model's output limit is 4,096 against 32,768, and it accepts video and PDF input where the old model took image and text only.
4,096 out

The output-limit change is the one that surprises people. A workflow that asked 2.5 Flash Image for an image plus a long written description in the same call may now hit the 4,096-token ceiling on the text half. Split the call, or move the text to a text model. Our original guide to Gemini 2.5 Flash Image describes the prompting patterns that will need re-testing.

05 — The planA one-week test plan that fits before October 2

Day 1: inventory and size check
List every call site and the surface it uses. Record the requested output size on each. Anything above 1K goes to the Flash Image candidate list; everything else to Lite.
Engineering
Days 2 and 3: prompt parity
Run your 50 most common prompts through the old model and the candidate, thinking at minimal, same aspect ratio. Score side by side on the properties your product depends on, not on taste. Keep the prompts that regressed for a second pass with thinking at high.
Product
Day 4: safety filters and refusals
Replay the prompts that were near the line on the old model. A newer model's filters can refuse what the old one produced, and the reverse. Log every changed decision and decide whether your own filter needs adjusting.
Product
Day 5: cost and token accounting
Reconcile one day of candidate usage against the invoice. Confirm 1,120 tokens per image, confirm thinking tokens are what you expect, and move volume work to the Batch tier at half price if latency allows.
Finance and engineering
Days 6 and 7: cut over and keep the old path dark
Switch the model ID behind a flag. Do not delete the old path until October 3 on the API or until your Vertex date; a flag is cheaper than a rollback.
Engineering
What we could not resolve

We cannot tell you whether the Gemini API will keep serving the model past October 2, whether Vertex will extend it past March 15, 2027, or why the two pages disagree. Google's pages do not say, and we did not ask Google. The plan above assumes the earlier date on the surface it applies to, which is the only assumption that cannot be wrong in the direction that costs you.

The wider calendar of dated shutdowns across vendors is in our model deprecation calendar. If you would rather someone else ran the parity test, our AI transformation practice does this migration as a fixed-scope week.

06 — ConclusionThe date depends on the door you came in through

What to do

Migrate Gemini API callers to 3.1 Flash Lite Image this week, give Vertex callers the same test on a longer clock, and ignore the API page's replacement pointer

The replacement is cheaper and live on both surfaces. The only real work is checking output size, re-running your prompts with thinking at minimal, and reconciling one day's tokens against the bill.

Digital Applied

Retire a model before it retires you.

We run prompt-parity, safety-filter and cost-reconciliation tests for image and text model migrations, and cut over behind a flag before the vendor's date.

Migration test plansCost reconciliationFlagged cutovers
Your next project

A Gemini image migration done in a week

  • →Every call site and surface inventoried
  • →50-prompt parity scored
  • →One day of tokens reconciled to the invoice
Questions and answers

The questions we get about the Gemini 2.5 Flash Image retirement

October 2, 2026 according to the Gemini API deprecations page, and March 15, 2027 according to the Vertex AI model versions page, both read September 25, 2026. Plan for the date on the surface you use.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading