Google Retired Two Gemini Flash Models by Rerouting Them. Your Evals May Be Testing the Wrong Model

If your code calls gemini-3.7-flash or gemini-3.5-flash, it probably still works. That is the problem worth thinking about.

On October 8, Google deprecated both models and began routing their traffic to newer ones. The call succeeds. The model answering it is not the one you named.

An error would have been easier to deal with. A quiet substitution can pass every health check you have.

What Google's release notes say

The Gemini API release notes for October 8, 2026 contain two relevant entries:

• gemini-3.7-flash is deprecated and replaced by gemini-3.8-flash. All requests to gemini-3.7-flash are automatically routed to gemini-3.8-flash.
• gemini-3.5-flash is deprecated and replaced by gemini-3.6-flash. All requests to gemini-3.5-flash are automatically routed to gemini-3.6-flash.

Both entries tell developers to update their model string, and no entry says calls will fail. The same release notes entry also deprecates a Deep Research agent, deep-research-pro-preview-12-2025, with a shutdown date of October 23, 2026. That one has a date and a named migration path.

The deprecations page lists both Flash models with "No shutdown date announced" and the routed replacement. It also says that listed shutdown dates are the earliest possible dates on which a model might be retired, and that Google will communicate exact dates with advance notice.

Why a working call is not enough

For uptime, the routing does its job. Requests keep working.

For quality, it says less than you need. A different model can change verbosity, formatting habits, refusal boundaries, tool-calling behavior, and latency. Those changes do not raise an exception.

Think about what in a typical system would notice:

• Health checks see a 200 response.
• Error alerts see nothing.
• Pinned evals keep running, but if they call the old ID, their results now describe the replacement model, not the one the report names.
• Cost dashboards may move, in either direction.

This is a description of how the failure would look, not a report of any specific breakage. I have not tested the rerouted models against a production workload.

What changed in price

Google's pricing page lists gemini-3.8-flash and gemini-3.6-flash at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, and $1.50 and $7.50 starting January 1, 2027, on the paid tier.

I could not find the deprecated models' earlier prices on Google's current pricing page, so I am not stating what a caller paid before. The independent analysis site Pick Right reports that the old 3.5 Flash rate was $1.50 and $9.00 and the old 3.7 Flash rate matched the new one. Treat that as a secondary claim.

The part I can state from Google's own page is the 2027 step. For the replacement models, every listed standard rate doubles on January 1, 2027. If you budget on these model families, use the 2027 numbers.

A simple hypothetical shows the size of it. A workload of 10 million input and 2 million output tokens per month costs $15.00 at the listed 2026 rates and $30.00 at the listed 2027 rates ($7.50 plus $7.50, then $15.00 plus $15.00). Your volumes will differ, and this ignores caching and batch discounts.

A short checklist

None of this requires new tooling.

1. Find every pinned ID. Search code, config, notebooks, eval harnesses, and prompts-as-config for gemini-3.5-flash and gemini-3.7-flash.

2. Decide on purpose. Either move to the replacement ID deliberately, after testing, or choose a different model. Do not leave the routing to decide for you.

3. Log what served the request. Record the model identifier your responses return, if your SDK exposes one, alongside the model you requested. A mismatch is the signal. I did not verify which response field Google uses for this, so check the current API reference for your SDK.

4. Re-run evals under the new ID. Keep the old results labeled with the date they were collected. Results collected after October 8 under an old ID may not describe the old model.

5. Watch the release notes. Subscribe to the changelog or check it on a schedule. The deprecations page lists shutdown dates, but in this case the table did not carry a date for when the routing took effect.

6. Put the January 1 price change in the budget.

7. Check the Deep Research agent. If you call deep-research-pro-preview-12-2025, the notes say it shuts down on October 23, 2026. The listed replacements are deep-research-preview-04-2026 and deep-research-max-preview-04-2026.

A note on sources that disagree

While researching this, I found third-party pages claiming other Gemini models were deprecated, including one saying Gemini 3.6 Flash was being retired on a specific November date for a particular app platform. I could not match that to Google's own release notes or deprecations page, so I left it out. When a vendor's changelog and a secondary site differ, the changelog wins, and the difference is worth a direct check.

What I would not conclude

Do not conclude that the rerouted models are worse. I have no test that says so. Google describes gemini-3.8-flash as its most intelligent Flash model, which is a vendor claim.

Do not conclude Google broke a promise. The documentation says deprecated models are eventually shut down and that routing is automatic in this case. The point is narrower: a successful call no longer proves which model ran.

Frequently asked questions

What did Google change on October 8?

It deprecated gemini-3.7-flash and gemini-3.5-flash and began routing their requests to gemini-3.8-flash and gemini-3.6-flash, according to the Gemini API release notes.

Will my code break?

Google says no immediate code changes are required to keep requests working. Behavior of the replacement model may differ from the old one.

Is there a shutdown date?

The deprecations page lists "No shutdown date announced" for both models, and says listed dates are earliest possible dates.

Does the price change?

For the replacement models, Google lists $0.75 input and $3.75 output per million tokens through December 31, 2026, doubling on January 1, 2027.

How do I know which model answered?

Log the model identifier in the response, if your SDK provides one, and compare it with the model you requested. Verify the exact field in current documentation.

What should I do about my evals?

Re-run them under a deliberately chosen model ID and label results by date. Treat results collected after October 8 under an old ID with caution.

Is anything else shutting down soon?

The same release notes entry gives October 23, 2026 as the shutdown date for the deep-research-pro-preview-12-2025 agent.

Sources

• Gemini API release notes: October 8 deprecations and routing, Deep Research agent shutdown date.
• Gemini deprecations: deprecation table, replacement models, and the earliest-possible-date note.
• Gemini API pricing: listed rates for the replacement models, including the January 1, 2027 step.
• Pick Right: independent analysis of the silent reroute, cited for the earlier price claim only.