---
title: "Gemini Flash Deprecation: 3.7 and 3.5 Flash Auto-Routed on Oct 8"
description: "Gemini 3.7 Flash and 3.5 Flash were deprecated on Oct 8 and requests now auto-forward to newer models. Learn how to confirm the serving model and migrate."
url: https://cdnping.org/en/blog/gemini-flash-deprecation-auto-routing
language: en
published: 2026-10-18
category: "Service updates"
tags: ["Gemini","model deprecation","auto-routing","migration","regression testing"]
source: cdnping.org
license: Free to quote and summarize with a link to the url above.
---
# Gemini Flash Deprecation: 3.7 and 3.5 Flash Auto-Routed on Oct 8

If your code still calls `gemini-3.7-flash` or `gemini-3.5-flash`, a different model is now handling those requests. The Gemini API changelog recorded both Flash deprecations on October 8, 2026, and requests are automatically forwarded to the replacement models. So if your code didn't change but your output did, the model probably didn't get "dumber"; the backend was swapped. This post explains what happened and how to confirm and migrate. Information is current as of October 2026; the official docs are authoritative.

## What exactly happened on October 8?

According to the [Gemini API changelog](https://ai.google.dev/gemini-api/docs/changelog):

| Date | Change | Replacement |
|---|---|---|
| Oct 8 | `gemini-3.7-flash` deprecated; it can no longer be selected in AI Studio; API requests sent to it are automatically routed | `gemini-3.8-flash` |
| Oct 8 | `gemini-3.5-flash` deprecated; requests are automatically forwarded | `gemini-3.6-flash` (the changelog also suggests you can use `gemini-3.8-flash` directly) |
| Oct 8 | Deep Research agent `deep-research-pro-preview-12-2025` deprecated; shuts down Oct 23 | `deep-research-preview-04-2026` or `deep-research-max-preview-04-2026` |
| Oct 6 | `gemini-3.1-flash-image` deprecated; no shutdown date announced | `gemini-nano-banana-2.1` |

For Deep Research, change the `agent` parameter in your `interactions.create` request to the new agent ID. As the changelog describes it, the first replacement favors speed and suits streaming output to a UI, while the second is more thorough.

The [models page](https://ai.google.dev/gemini-api/docs/models) itself doesn't list detailed dates for these deprecations. Check the official deprecations page for the exact schedule.

## Why would output style, usage, and latency change?

Once a request is forwarded, a different model is answering. Differences you might see: tone and formatting habits, output length for the same prompt (which affects token usage and cost), time to first token and total latency, and how closely the model follows structured output and tool-calling requirements. These are reasonable places to look, not documented effects; there are no official numbers on how large the differences are, so measure with your own use cases.

Pricing deserves a separate check. See [Gemini API's two price increases, by date](/en/blog/gemini-api-price-increase-dec-2026) for a date-by-date breakdown.

## How do I confirm which model actually served a request?

1. Check the response body. The `generateContent` response includes a `modelVersion` field. Record it and write it to your logs. Whether it matches the model ID you requested is defined by the official API reference.
2. Log both the requested model ID and the response's `modelVersion`, and alert when they differ after rollout.
3. Compare similar requests from before and after October 8: output length, usage fields (`usageMetadata`), and latency distributions.
4. Cross-check against external timelines. cdnping's [/en/changes](/en/changes) records additions and removals in the official changelog line by line, and [/en/down/gemini](/en/down/gemini) has user-submitted "it got dumber" reports, so you can see whether they cluster after October 8. User reports are subjective signals; treat them as leads only.

## Regression checklist before and after migration

- Search globally for every hardcoded model ID, including config files, environment variables, scheduled jobs, and settings in third-party tools.
- Pick 20 to 50 representative real requests and compare archived old outputs (if you have them) against the new model's outputs one by one.
- Check format assertions: JSON structure, field names, code-block wrapping, language.
- Check length and cost: the change in average output tokens, multiplied by the current unit price, gives the difference.
- Check latency, especially time to first token. For a way to think about it, see [AI API latency breakdown](/en/blog/api-latency-connect-rtt-ttft).
- Check feature support: confirm that tool calling, structured output, multimodal input, and so on all work on the new model, per the official model guide.
- Pin to an explicit new model ID before shipping instead of relying on forwarding. How long automatic forwarding will last is up to the official documentation.

## How do I keep tracking changes like this?

Deprecation notices often appear in the changelog for only a day or two. [How to find out right away when an LLM API price changes](/en/blog/track-llm-price-and-model-changes) covers three channels: changelogs, model catalogs, and billing. You can also use [/en/models](/en/models) to see when models were listed, delisted, or repriced on OpenRouter. If the output change relates to quantization or routing on the provider side, [Did the model get dumber, or did the provider change?](/en/blog/llm-quantization-fp8-fp4-providers) explains how to tell.

Gemini also has a newer release: the official changelog records `gemini-4-argon` becoming generally available on October 13. For usage and pricing, see [Gemini 4 Argon is now GA](/en/blog/gemini-4-argon-ga-developers).

## Sources

- [Gemini API changelog](https://ai.google.dev/gemini-api/docs/changelog)
- [Gemini API models list](https://ai.google.dev/gemini-api/docs/models)