Gemini 3 models: current endpoints, shutdowns & migration

Gemini 3 current models at a glance

Google's catalog changes quickly. The table is a dated routing guide, not a promise that every model appears in every Gemini app, Workspace, Search, Vertex AI or third-party surface. API developers should use exact endpoint IDs; app users should record the visible mode and account tier. Lifecycle labels apply to an exact ID: the Gemini Omni model card labels gemini-omni-1.1-flash stable and gemini-omni-flash-preview preview. Those names are not interchangeable.

What happened to the original Gemini 3 Pro?

Google's deprecation page lists gemini-3-pro-preview as released November 18, 2025, shut down March 9, 2026, and replaced by gemini-3.1-pro-preview. Code that still names the original endpoint needs migration. An old screenshot or launch article does not prove that the endpoint remains callable.

The broad name “Gemini 3” remains useful for the model generation. The exact suffix determines whether the route is stable or preview, optimized for general tasks, images, live audio, transcription or video, and subject to a published shutdown date.

How to choose a current Gemini 3 model

• Start with Gemini 3.8 Flash for Google's current stable general-purpose Flash route, especially complex coding and agentic workflows.
• Use Gemini 3.5 Flash-Lite when throughput and cost matter more than choosing the newest general-purpose Flash model.
• Evaluate Gemini 3.1 Pro when a preview endpoint is acceptable and the job needs harder reasoning, agentic work or coding.
• Use Nano Banana 2.1, Nano Banana 2, Nano Banana 2 Lite or Nano Banana Pro for image generation and editing. Nano Banana 2.1 is a separate current endpoint, gemini-nano-banana-2.1; the older IDs remain distinct models.
• Use Gemini Omni Flash for Google's current Gemini-branded video generation and editing route, and Veo for the separate Veo model family.
• Use Gemini 3.8 Live for interactive voice; consider gemini-3.8-live-extended-thinking when the live interaction needs more background reasoning. For a finished narration file, use 3.8 Flash TTS or Flash-Lite TTS instead of assuming a text or Live route has the same request format.

Before production use, confirm input types, output formats, context window, regional availability, rate limits, safety behavior and pricing in the current documentation. Google's pricing page bills different models and modalities differently, so “Gemini 3 pricing” has no single correct number.

How to migrate from a retired Gemini endpoint

• Inventory the exact endpoint IDs in source code, stored configurations, dashboards and evaluation logs.
• Map each retired ID to Google's current recommended replacement; do not replace every model with one global default.
• Replay a fixed set of representative prompts and media with the same acceptance criteria.
• Compare schema adherence, tool calls, safety responses, latency, token or media cost, and downstream correction work.
• Update parsing and retry logic only after checking the replacement's current response shape and limits.
• Record the test date and endpoint ID so later readers can reproduce the result.

Voice migration: update requests and audio handling

Google lists November 17, 2026 as the earliest shutdown date for gemini-3.1-flash-tts-preview in its deprecation schedule. Check for the final notice and migrate to 3.8 Flash TTS or Flash-Lite TTS before relying on the legacy preview.

The 3.8 TTS migration guide and 3.8 Live migration guide document changes beyond the model name:

• TTS transcript and directions. Keep spoken words in the input text; put delivery style and speaker labels in speech_metadata. Inline directions may otherwise be spoken. Each multi-speaker turn needs a speaker matching the configured voice.
• TTS audio bytes. Unary 3.8 TTS requests return WAV by default. Remove code that adds a second WAV header to the returned bytes; explicitly request audio/l16 if the pipeline needs headerless PCM.
• Live session configuration. When moving from 3.1 Live to gemini-3.8-live, remove the old thinking_level or thinking_config session settings. Google documents proactive audio as permanently enabled; sending proactive_audio=false produces an error.
• Live tool responses. Function calls default to NON_BLOCKING. If the application requires the model to wait for a tool result, configure BLOCKING deliberately and test interruptions and delayed responses.

Before rollout, replay a short narration and a representative live tool-call conversation. Inspect request validation, speaker assignment, audio headers and turn behavior. These are migration checks to perform, not tests this article claims to have run.

What Gemini 3 introduced at launch

Google's November 2025 launch collection introduced Gemini 3 Pro across the Gemini app, AI Mode in Search, AI Studio, Vertex AI and developer tools, with multimodal reasoning, coding, generative interfaces and agentic features. Those launch statements describe the products and access tiers available at that time.

The following three images are retained as launch-era evidence. They document the November 2025 article and should not be read as screenshots of the current model selector or current feature availability.

Launch-era Gemini 3 Pro image

Historical screenshot: Gemini 3 Pro artwork from the November 2025 launch article.

Launch-era Gemini Agent rollout

Historical screenshot: Gemini Agent was introduced as an early Ultra rollout. Check the current app and help documentation before promising access.

Launch-era Gemini Labs tools

Historical screenshot: Experimental Labs and generative-interface examples from the launch period. Experimental names and availability can change.

How to evaluate Gemini claims reliably

A useful comparison names the exact endpoint, access surface, date, region, account tier, prompt, inputs, settings and number of attempts. It separates Google's documented capabilities from observed test results and preserves failures as well as selected outputs.

• For text and coding, score factual correctness, citation quality, tool execution, schema adherence and correction effort.
• For images, score instruction adherence, reference preservation, text accuracy, edit locality and cost per accepted image.
• For voice or live interaction, measure latency, interruption handling, transcription accuracy and recovery from failed tool calls.
• For video, measure prompt adherence, temporal consistency, audio synchronization, duration, resolution, retries and editing effort.

For current image choices, use our AI image generator comparison and Nano Banana Pro review. If the deliverable is media rather than a general assistant answer, test a representative brief in Magic Hour's AI image generator or AI video generator.