Google retired two Gemini Flash models by rerouting them, not by failing them

Every major model vendor’s deprecation policy promises the same mechanism. Anthropic: requests to a retired model fail. OpenAI: requests fail after the shutdown date. Mistral: retired models return a 404. Google’s own lifecycle documentation says the model “is turned off and its endpoint stops responding.”

That failure is not a defect in those policies. It is the product. A hard failure is what tells an engineering team that the model it pinned eighteen months ago is gone, on the day it goes, without anyone having subscribed to a changelog.

On 8 October 2026 Google retired two Gemini Flash models without it.

What the changelog says

Two lines in the Gemini API release notes for 8 October 2026:

All requests to gemini-3.7-flash are automatically routed to gemini-3.8-flash.

All requests to gemini-3.5-flash are automatically routed to gemini-3.6-flash.

Both models are marked deprecated. Neither has a shutdown date, because neither needs one — there is nothing to shut down. The endpoint keeps answering. It is simply a different model answering.

Checked on 10 October 2026, gemini-3.5-flash and gemini-3.7-flash are absent from Google’s models page and absent from its pricing page. You cannot look either model up, and you cannot buy either one.

Correction and update, 18:30 UTC 10 October. An earlier check for this article, at 11:00, found neither ID on Google’s deprecations page. A re-check the same evening found rows for both. We cannot tell from the outside whether Google added them during the day or the morning read hit a stale build of the docs, so both observations are recorded here rather than one being quietly replaced.

What the rows say is the part that still matters. The table’s columns are model, release date, shutdown date and recommended replacement:

There is no column for when the deprecation took effect, and the one date column that could have carried it says nothing will be shut down — correctly, because nothing will be. The page describes the current state of the catalogue; it does not record that the catalogue changed, or when. A monitoring job keyed to shutdown dates — which is what a deprecation page is normally for, and what we check on a rotation to maintain our own retirement calendar — reads these two rows and finds no action and no deadline. Contrast the third deprecation in the same changelog entry, deep-research-pro-preview-12-2025, which got a row and a 23 October 2026 shutdown date with deep-research-preview-04-2026 as the replacement. One retirement is legible as an event. Two are legible only as a changed row you had to have been diffing.

The bill is fine. That is not the issue

It would be easy, and wrong, to write this as a price story.

Per million tokens, input / output, verified against Google’s pricing page on 10 October 2026.

A 3.5 Flash caller just got a 50% input cut and a 58% output cut without asking. A 3.7 Flash caller’s bill does not move at all, because 3.8 Flash shared 3.7 Flash’s introductory rate and its 31 December 2026 expiry from the day it launched — something we flagged in September as buying no extra discounted runway, and which has now resolved in the most literal way available: the two models became one. Both substitutes double to $1.50 / $7.50 on 1 January 2027, which was already the published fate of the models they replaced.

Nobody is being overcharged. The thing being withdrawn is not money. It is the meaning of a model ID.

Fifty-six days

Gemini 3.7 Flash launched on 13 August 2026, billed by Google as its “most intelligent workhorse model yet for coding and agents,” with a 16-point jump on DeepSWE and an aggressive introductory price aimed squarely at the Flash-tier price war. We covered it as the best-value Flash for coding agents. It was deprecated 56 days later.

Gemini 3.5 Flash — the model Google’s own I/O 2026 keynote presented as surpassing Gemini 3.1 Pro — managed 142 days.

Set that against what the other vendors put in writing. OpenAI publishes a minimum of six months for generally available models, three months for specialised variants, and as little as two weeks for previews. Anthropic commits to at least 60 days’ notice before a publicly released model retires. Mistral publishes six months for GA models and one month for previews. Google publishes no minimum notice period at all, and states that the shutdown dates it does list are earliest-possible dates rather than commitments. On a 56-day-old GA model, it used zero days of notice, and the deprecations-page row it eventually carries states no shutdown date and no effective date.

This is the second time in a month that Google’s Gemini lifecycle has moved without a published date. On 4 October we covered the Gemini app’s free tier dropping to Flash-Lite with AI Plus losing the Pro model on a per-account date delivered by email rather than a published one. And Gemini 2.5 is still, formally, undeprecated while being closed to new projects — a state we had to write a migration guide to explain, because Google’s deprecation page lists no shutdown date for models you cannot adopt.

A pattern is now visible across all three: on the Gemini API, the published lifecycle record is not where the lifecycle actually happens.

Why silent substitution is the worse failure mode

Consider the two ways a vendor can take a model away.

Hard failure. Your requests 404. Your error rate spikes, your alerting fires, an engineer reads the message, pins the replacement, runs the evals, ships. The outage is loud, bounded and immediately attributable. Cost: one incident.

Silent substitution. Your requests succeed. Latency shifts slightly. Output distribution shifts — different verbosity, different formatting habits, different refusal boundaries, different tool-calling behaviour, a different knowledge cutoff. Nothing fires, because nothing failed. The drift surfaces weeks later as a quality regression in whatever sits downstream: a classifier’s accuracy sliding, an agent loop that now takes four tool calls where it took three, a summary format your parser half-handles, a RAG answer style your users notice before your dashboards do. Cost: unknown, unattributed, and paid by whoever eventually bisects it.

For an interactive chatbot, substitution is clearly the kinder default — users get a working, cheaper, better model instead of an error. That is a defensible product decision and it is probably why Google made it.

For anything with an eval suite, a reproducibility requirement, a regulated output or a parser downstream, substitution removes the one control the API offered. Model pinning exists so that the thing which answered yesterday answers today. On the public Gemini API, as of 8 October, pinning a Flash ID no longer does that, and the fact that it stopped doing it is recorded in a changelog rather than the deprecation list.

What to do about it

1. Diff the models endpoint, not the deprecation page. On this platform, disappearance from /models is the real retirement signal, it is unambiguous, and your own job dates it. A weekly diff of available model IDs would have caught both of these on 8 October. Reading the deprecations page would not: the rows arrived later, carry no effective date, and announce no shutdown — there is nothing on them for a date-driven alert to fire on. Diff the deprecations page too if you like, but diff it as a page, not as a calendar.
2. Add output canaries to your eval suite. Twenty fixed prompts, run on a schedule, outputs diffed. Silent substitution is only observable as drift, so you need something that watches for drift. This is cheap insurance that also catches in-place model updates behind an alias, which Google, like everyone else, does not announce.
3. Pin major.minor, and expect it to be a request rather than a guarantee. Mistral’s documentation is unusually blunt that an alias “may expose you to silent updates in model behavior and pricing.” That warning now applies to a pinned Gemini version too.
4. If a specific version is load-bearing, buy it somewhere with contractual lifecycle terms. Vertex AI is where Google sells enterprise lifecycle commitments; the public Gemini API is not that product, and the gap between them is the subject of this story. Anthropic’s note that Bedrock and Google Cloud set their own retirement schedules cuts the same way.
5. Re-price your Flash workload before 1 January 2027. Whichever Flash you are now on, it doubles to $1.50 / $7.50 that day. Run your real volumes through our API cost calculator, which carries both the current introductory rates and the January step-up.

The broader lesson is not that Google is careless with models — three Flash releases in under four months is the behaviour of a team shipping hard, and the substitution it chose left every affected customer on a cheaper, better model. The lesson is narrower and more useful: the strength of a vendor’s lifecycle promise is a purchasable feature, and right now Google’s published promise is the weakest of the six vendors we track. If that matters to your workload, it is a reason to run the Gemini API behind an adapter you can repoint in an afternoon, with a second vendor already qualified against your own evals.

The full Gemini lineup, plan arithmetic and what each tier actually includes is in the Gemini review. Every announced shutdown date we can verify, by vendor, with notice periods and post-retirement behaviour, is in the retirement calendar; the dated price history is in the pricing tracker.

Questions or corrections? Email Pick Right. Want the full list? See all news.