Requests to gemini-3.7-flash now get 3.8 Flash, with no shutdown date

Your config says gemini-3.7-flash. Since 8 October, Google's been answering it with gemini-3.8-flash, and nothing in your code had to change for that to happen. That's the whole notice: two lines in the Gemini API changelog, plus a routing note on the deprecations page. It's the same model Google told developers on 2 September to keep using if they cared about efficiency.

What the docs say, and what they leave out

The changelog entry is short. 3.7 Flash "is deprecated and has been replaced by gemini-3.8-flash", and the model page adds that "all requests to gemini-3.7-flash are automatically routed" to the newer one. The same treatment goes to 3.5 Flash, which now lands on 3.6 Flash. A third item in that entry has a real date: the deep-research-pro-preview-12-2025 agent shuts down on 23 October, with deep-research-preview-04-2026 as the replacement.

Then look at the deprecations table. The 3.7 Flash row reads "No shutdown date announced", followed by the routing note. Google's own definition on that page says a deprecation announces that a model "will be shut down in the near future". Here it's the other way round. The model stopped answering first, and the date never came. Strictly, nothing was shut down, since the endpoint still accepts your request. You just don't get the model you named.

That's a change of policy in practice, if not on paper. On 2 September Google's 3.8 Flash post told developers who care about compute efficiency to "continue to rely on Gemini 3.7 Flash, which remains fully supported for efficiency-first workloads." That sentence is still on the blog today, with no update note. It held for 36 days. Our write-up of the 3.8 launch repeated it as advice, and it's wrong now.

Same price per token, maybe not per task

On the 3.7 path, the rate card doesn't move. 3.7 and 3.8 Flash both sat at the introductory $0.75 per million input tokens and $3.75 output, and both jump to $1.50 and $7.50 on 1 January 2027. 3.7 Flash has simply vanished from the pricing page. So the risk isn't the price. It's the count.

Google said so itself when 3.8 shipped: "At times, the model might use more tokens to maximize performance, especially at higher effort levels." A developer on Google's forum opened a thread on 12 October asking to keep 3.7 Flash as a legacy endpoint, saying 3.8 uses 30 to 40 percent more tokens on simple routing and edits. That's one person's workload and we haven't measured it. Still, it points the same way Google's own sentence does. And from 1 January every one of those extra tokens costs double.

The 3.5 path is murkier. 3.5 Flash listed at $1.50 in and $9 out when 3.6 Flash shipped in July, while 3.6 Flash now sits at the $0.75 and $3.75 introductory rate. The pricing page no longer lists 3.5 Flash, and Google doesn't say which rate a routed call bills at. I'd expect the 3.6 rate, which would be a saving, but that's a guess until your billing report shows it. DeepSeek, for comparison, wrote it down when it did the same thing in September: calls to the retired deepseek-v4-flash are "billed at the Flash price".

One thing didn't break. thinking_level: minimal already returned an error on 3.7 Flash, and it still does on 3.8, so routing doesn't turn working requests into 400s. 3.6 Flash does accept minimal.

What we'd change this week

Change the string to gemini-3.8-flash yourself. Leaving the old name in place means your config, your logs and your evals all describe a model that isn't running. One public project's tracker issue spells it out: every scan since the switch ran on 3.8 while every saved record said 3.7. If you're keeping records, read the modelVersion field Google returns with each response ("the model version used to generate the response") and store that, not the name you sent.

Then set thinking_level explicitly. If you relied on 3.7 being the lean option, low is the closest you'll get on 3.8, and it's worth comparing usageMetadata on a few hundred of your own requests before and after. Put the token count per task next to the 1 January rate change, because that's when a 30 percent drift starts costing real money.

And don't read "No shutdown date announced" on any other Gemini row as a promise. 3.8 Flash carries the same words. We've updated the Gemini rows in our LLM API pricing tracker, and the wider list of dated retirements lives in the retirements calendar.

Frequently asked questions

Sources: Google, Gemini API changelog (8 October 2026 entry), Gemini deprecations, Gemini 3.7 Flash model page, Gemini API pricing, generateContent reference and the 3.8 Flash announcement of 2 September; independent: Google AI developers forum thread and a public GitHub issue on the redirect; DeepSeek, models and pricing. All read on 12 October 2026.