Firebase AI Logic Gemini model shutdown: a 3.5 Flash plan before 19 May 2027

Summary. Google Cloud's Agent Platform lifecycle page lists gemini-3.5-flash, released on 19 May 2026, for retirement on 19 May 2027 or later. Firebase's own models page prints no date for it and names gemini-3.8-flash as the replacement. What the date means depends on your Gemini API provider. On the Agent Platform Gemini API (formerly Vertex AI), 3.5 Flash answers until retirement, then requests fail with a 404 error. On the Gemini Developer API nothing switches off: since 8 October 2026 Google has routed gemini-3.5-flash calls to gemini-3.6-flash, so your users already get a model you may not have tested. Changing the string is not enough. From 3.7 Flash onwards, requests that set the MINIMAL thinking level, a frequency or presence penalty or a candidate count fail with a 400 error, as do prefilled model turns. Two earlier dates shape the release: App Check enforcement began on 2 November 2026 and can be delayed only to 21 January 2027, and from April 2027 iOS uploads need the iOS 27 SDK. Ship one release before late January with App Check, Remote Config model names and 3.8-safe settings, then switch models remotely.

About this article. For engineering leads and release managers shipping Firebase AI Logic features on iOS, Android, Flutter, web or Unity. We checked every model date, SDK behaviour, price and deadline against Firebase documentation and release notes, Google's Gemini API and Agent Platform pages, and Apple and Google Play help pages on 7 November 2026. Google changes model schedules often, so re-check before you commit. The India note is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What Google lists for Gemini 3.5 Flash on 7 November 2026

Three newer Flash models shipped within four months of 3.5 Flash, and the Developer API now redirects older names. Firebase's models page, updated on 6 November, shows the routing and replacements but no date for 3.5 Flash, so the dates below come from Google's two provider pages.

Four rules from Firebase's FAQ explain how to read that table:

• Twelve months is the norm, not a promise. Google aims to keep stable models for at least 12 months, but Agent Platform short-term models can be retired as early as 45 days after a replacement ships. 3.6 Flash will have lived 121 days.

• Dates are floors. Listed dates are the earliest possible retirement, and Agent Platform says they can move later but not earlier.

• New projects are locked out first. Several weeks before a shutdown, projects that have never used the model are blocked from starting, so a fresh staging project can fail before production does.

• After shutdown, 404. A request naming a retired model fails.

Our AI model retirement policies comparison sets these rules beside other vendors'. One wrinkle: Firebase's 5 October 2026 release note, under its 3.5 Flash heading, gives gemini-3.6-flash a 19 May 2027 Agent Platform date. That reads as a typo, since 3.6 Flash retires on 19 November 2026.

Which backend is your app on? Two different failures

Apps choose a provider at initialisation: GenerativeBackend.googleAI() in Kotlin, .googleAI() in Swift or GoogleAIBackend on the web for the Gemini Developer API, or the Agent Platform backend, now agentPlatform or AgentPlatformBackend and formerly vertexAI or VertexAIBackend. That line decides what 19 May 2027 means for you.

Gemini Developer API: the change already happened. Google's deprecations page gives gemini-3.5-flash no shutdown date; the endpoint stays up and forwards requests to gemini-3.6-flash. Your hard-coded string keeps working, but answers come from another model, billed at its rates. Output length, JSON shape or refusal behaviour may have drifted with no code change. On 1 January 2027 Google doubles paid 3.6 Flash and 3.8 Flash rates, from $0.75 to $1.50 per million input tokens and from $3.75 to $7.50 per million output tokens. Our Gemini 3.8 Flash price guide works through the impact. Gemini 2.5 follows on 1 December 2026, when Developer API requests to gemini-2.5-pro route to gemini-3.8-flash and gemini-2.5-flash to gemini-3.5-flash-lite, which Firebase warns may break apps.

Agent Platform Gemini API: a hard stop. Here gemini-3.5-flash is served as itself until retirement, then every app version with the name compiled in fails. There is also a syntax risk. Firebase deprecated the legacy Vertex AI initialisation syntax on 6 August 2026 and removed it in October's breaking SDK releases: Apple v13.0.0, Android BoM v35.0.0 and JavaScript v13.0.0 on 7 October, Flutter v5.0.0 and Unity v14.0.0 on 15 October. Older builds keep running, but the legacy syntax defaults to us-central1, which most new Gemini models do not support, while the new syntax defaults to global. Firebase's 3.8 Flash note tells legacy-syntax users to set global, so an old build that receives gemini-3.8-flash through Remote Config but still points at us-central1 can fail, unless it also reads its location from Remote Config, as the agent_platform_location parameter in Firebase's solution guide does.

What replaces gemini-3.5-flash

Firebase's Remote Config sample uses gemini-3.8-flash as the default model_name, and its release notes name 3.8 Flash or Gemini 4 Argon as replacements. For most mobile features the real choice is between 3.8 Flash and 3.5 Flash-Lite.

Retirement dates are from Google Cloud's lifecycle page and billing notes from Firebase's models page.

Three cautions. First, hard-coding gemini-3.8-flash only moves the cliff to 2 September 2027, 106 days after 3.5 Flash's floor. Second, -latest aliases are hot-swapped with every release, with a two-week warning email only for breaking changes. Third, Gemini 4 Argon differs in Firebase AI Logic: output caps of 32,768 tokens non-streaming and 65,536 streaming, no explicit caching, and a 400 error for both MINIMAL and thinkingBudget. Paid 3.5 Flash-Lite is listed at $0.30 input and $2.50 output per million tokens with no dated change.

Why changing the model name alone can break old builds

Remote Config lets you change model_name without an app release, but every other request field is compiled into each build. Firebase's 2 September 2026 note for 3.8 Flash and its thinking guide list what fails or behaves differently.

Before you flip model_name, find out which app versions send any of these. If version 4.2 of your Android app sets candidateCount or MINIMAL, switching it to 3.8 Flash turns a working feature into a stream of 400s. Remote Config conditions can target app versions, so move compatible builds first and hold the rest on 3.5 Flash until they age out. Firebase lists MINIMAL as the default for 3.x Flash-Lite, so a build whose only problem is MINIMAL may suit 3.5 Flash-Lite, if your evaluations agree.

On the Developer API, gemini-3.5-flash builds are already served by 3.6 Flash, which still accepts MINIMAL and candidate count but, per Firebase, no longer supports prefilled model turns. A build that prefills may already be returning 400s, so test them against 3.6 Flash now.

Remote Config, server prompt templates, or both

Firebase recommends two ways to stop hard-coding, Remote Config and server prompt templates, and says released builds that use either start using a new model as soon as you change it.

Remote Config is the smaller change: define a model_name string parameter, ship in-app defaults, fetch and activate on launch, and add a real-time listener so updates reach running apps as soon as you publish. The Kotlin call from Firebase's guide is short:

            // Model name comes from Remote Config; the in-app default is the fallback
val model = Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(
    modelName = remoteConfig.getString("model_name")
)
          

Firebase's sample sets a minimum fetch interval of 3,600 seconds and suggests lower values only in development. Keep it that way, because since 1 September 2026 Remote Config is priced on fetch volume. Each project gets 100,000 fetches a day free. Above that, Blaze projects pay $0.000006 per request up to 10 million a day, and Spark projects are throttled from the 31st day after first going over. Projects that used Remote Config before September move to this pricing on 1 December 2026 on Spark, or on 1 February 2027 on Blaze or if they upgrade before 15 November 2026. Check your daily fetch count before an emergency switch depends on it.

Server prompt templates go further: the prompt, schema and configuration live on the server, and the app sends only a template ID and inputs, so model name, thinking level and generation settings leave the device. But the feature is in Preview, with no SLA or deprecation policy, and may change in backwards-incompatible ways.

We suggest five parameters as a minimum: model_name, thinking_level, agent_platform_location for Agent Platform apps, an ai_feature_enabled kill switch and a min_supported_version that triggers an upgrade prompt. Enforcing a minimum app version is one of Firebase's standard Remote Config use cases, and it is how you retire builds that can never be made 3.8-safe.

App Check: the deadline that arrives first

Unlike the model date, the App Check date has passed. Since 2 November 2026, Firebase enforces App Check for Firebase AI Logic automatically and blocks requests without a valid token, including released versions with no production attestation provider. You can delay auto-enforcement from the APIs tab of the App Check console, but not beyond 21 January 2027.

So any build without App Check is cut off by 21 January at the latest, whatever model it calls, and the release that adds App Check is the natural place to add Remote Config model switching and remove 400-prone parameters. After enforcement, the only builds that can reach the 19 May cliff are those with App Check that still hard-code the model.

What Firebase requires and recommends:

• Attestation providers. DeviceCheck, App Attest or reCAPTCHA Enterprise on Apple platforms; Play Integrity or reCAPTCHA Enterprise on Android; reCAPTCHA Enterprise on the web. Flutter supports all of them; Unity supports all but reCAPTCHA Enterprise for now. On mobile, reCAPTCHA Enterprise is still a Preview provider.

• Limited-use tokens. Default session tokens have a configurable time to live between 30 minutes and 7 days. Replay protection mints a fresh token per request, adds some latency and possibly cost, and needs useLimitedUseAppCheckTokens set to true at initialisation. Minimum SDKs are Apple v12.2.0, Android BoM v34.14.0, Web v12.14.0, Flutter v4.15.0 and Unity v13.12.0. Turn the flag on in this release even if you enforce replay protection later.

• Per-user limits. Firebase AI Logic sets a default of 100 requests per minute per user.

• Local development. Use the App Check debug provider so engineers can work with enforcement on.

Release-train timing: working back from 19 May 2027

The worked example assumes a Flutter app on iOS and Android, initialised against Agent Platform with the legacy syntax, gemini-3.5-flash hard-coded, the thinking level set to MINIMAL, and App Check auto-enforcement delayed. Dates are sourced above; the durations are ours.

The arithmetic: 21 January to 19 May 2027 is 118 days, enough for two or three monthly releases, but the April iOS change compresses it. Apple's submitting page says iOS and iPadOS apps uploaded from April 2027 must be built with the iOS 27 SDK and target iOS 15 or later. Our iOS 27 SDK requirement guide explains why that makes the UIScene life cycle and Liquid Glass behaviour non-optional. Removing a hard-coded model name in late April means absorbing a toolchain migration in the same release, with App Review in the critical path.

Apple says it typically reviews at least half of submissions in under 24 hours and 90% in under 48, and it accepts expedited requests for critical bug fixes. Google Play warns of extended reviews of up to 7 days, or longer in exceptional cases. Neither is a promise, which is why a remotely switchable model name beats a fast review lane.

India and data handling notes

Firebase says the Gemini Developer API offers only global access, while the Agent Platform Gemini API offers a global location and specific locations that vary by model. A global request is handled by an available model anywhere in Google's global pool. Firebase AI Logic has also supported the us and eu locations since 31 August 2026. Moving a legacy build from us-central1 to global therefore changes where requests are processed. If prompts can carry personal data from Indian users, check the change against your privacy notices, your duties under the Digital Personal Data Protection Act and any customer contract that promised a region. Our DPDP Rules checklist covers the wider deadlines.

Migration checklist

1. List every live app version that calls Firebase AI Logic, with its provider, model name, thinking settings, SDK version and initialisation syntax.

1. Search code for gemini-3.5-flash, gemini-3.6-flash, gemini-3.7-flash, gemini-3.1-flash-lite and any -latest alias.

1. Remove frequency penalty, presence penalty and candidate count; replace MINIMAL and thinking budgets with thinking levels; end chat history on a user turn.

1. On Agent Platform, adopt the agentPlatform syntax or set the location to global.

1. Add Remote Config parameters for model name, thinking level, location, a kill switch and a minimum version, with real-time updates.

1. Implement a production App Check provider, enable limited-use tokens and enforce before 21 January 2027.

1. Evaluate real prompts, including long chats and image input, on gemini-3.8-flash and gemini-3.5-flash-lite.

1. Submit by mid-December 2026, roll out in stages and watch verified-request metrics.

1. Switch model_name by app-version condition, in stages, in February 2027.

1. Prompt remaining incompatible versions to upgrade before 19 May 2027, and check who receives Google's shutdown emails.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI mobile application development services team builds and maintains Firebase AI Logic features on iOS, Android, Flutter and web, including App Check rollout, Remote Config model control, model-swap evaluations and release planning around store reviews. Talk to us at /contact-us/.

Last updated: 7 November 2026.

References

1. Model versions and lifecycle, Gemini Enterprise Agent Platform — Google Cloud

1. Learn about supported models, Firebase AI Logic — Firebase

1. FAQ and troubleshooting, Firebase AI Logic — Firebase

1. Gemini API release notes, 8 October 2026 entry — Google AI for Developers

1. Firebase release notes, 2 September 2026 (Gemini 3.8 Flash) — Firebase

1. Submitting to the App Store — Apple Developer

1. Gemini deprecations — Google AI for Developers

1. Firebase release notes, 5 October 2026 (older Gemini 3.x Flash models) — Firebase

1. Gemini Developer API pricing — Google AI for Developers

1. Firebase release notes, 2 November 2026 (Gemini 2.5 routing) — Firebase

1. Firebase release notes, 6 August 2026 (Agent Platform initialisation syntax) — Firebase

1. Firebase release notes, 7 and 15 October 2026 (SDK releases and removal of the legacy syntax) — Firebase

1. Dynamically update your Firebase AI Logic app with Firebase Remote Config — Firebase

1. Remotely change the model name in your app — Firebase

1. Firebase release notes, 13 October 2026 (Gemini 4 Argon) — Firebase

1. Thinking, Firebase AI Logic — Firebase

1. Remote Config conditional expression reference — Firebase

1. Firebase release notes, 21 July 2026 (Gemini 3.6 Flash and 3.5 Flash-Lite) — Firebase

1. Pricing for Remote Config — Firebase

1. Get started with server prompt templates — Firebase

1. Explore use cases, Remote Config — Firebase

1. Firebase release notes, 27 October 2026 (delaying App Check enforcement) — Firebase

1. Prevent Gemini API abuse with Firebase App Check — Firebase

1. Firebase release notes, 14 August 2026 (reCAPTCHA Enterprise for mobile apps) — Firebase

1. App Review — Apple Developer

1. Publish your app — Play Console Help

1. Specify the location for accessing the model, Firebase AI Logic — Firebase

1. Firebase release notes, 31 August 2026 (us and eu locations) — Firebase