Gemini deprecations: 2.5, 3.5 Flash and 3.7 Flash dates

Google deprecated Gemini 3.5 Flash and Gemini 3.7 Flash on 8 October 2026, and their model strings now answer as other models: a request naming gemini-3.5-flash is served by gemini-3.6-flash, and one naming gemini-3.7-flash by gemini-3.8-flash. The Gemini 2.5 models are not deprecated and have no shutdown date announced; Google serves them until further notice announced; Google serves them until further notice, but since 18 September 2026 only people who used them before can call them. Nothing returns an error, so the change is easy to miss.

That is the unusual part. A shutdown breaks your app and you fix it the same day. A reroute keeps your app running on a model you did not test, with a different token count, a different default behaviour and a different price.

What each model string does today

Neither deprecated model can be picked in Google AI Studio any more. These dates are for the Gemini API.

The shutdowns still ahead

Google's deprecations page lists the earliest date each model might be turned off, and announces deprecations in the release notes. Plan for the earliest one.

The voice and speech previews are the urgent ones: under 4 weeks away, and a voice agent that loses its model fails in front of a caller. The Deep Research agent deep-research-pro-preview-12-2025 already shut down on 23 October; its replacements are deep-research-preview-04-2026 and deep-research-max-preview-04-2026.

An embedding change is the expensive kind. Google says the vectors of gemini-embedding-001 and gemini-embedding-2 are incompatible, so moving means embedding your whole collection again. You have until 2028, but the embedding cost calculator shows what the re-run costs today.

What changes when you are rerouted

From 3.7 Flash to 3.8 Flash. Google says 3.8 Flash "can use more tokens on longer running and complex tasks, by design": it takes smaller reasoning steps, calls tools iteratively and verifies its work along the way. For an agent that is often what you want. Not every workflow needs that much checking. Google's own advice is to lower the thinking level for everyday tasks, or to use 3.6 Flash. Both 3.7 and 3.8 take low, medium and high thinking levels, with medium the default, so your settings carry over.

From 3.5 Flash to 3.6 Flash. This one mostly helps. Google built 3.6 Flash to finish multi-step work in fewer reasoning steps, turns and tool calls than 3.5, and priced its output lower: $7.50 per million tokens against $9.00, with input the same at $1.50. Google notes that 3.6 runs diagnostic code before changing things more often, which can add steps on routine front-end work, and that people preferred older models' visual styling.

The price, this year and next. On the paid tier, 3.8 Flash and 3.6 Flash share one price, and both are on an introductory rate until 31 December 2026 (on the free tier both are free of charge):

So whatever your rerouted traffic costs in November, it doubles in January for the same tokens. If 3.8 Flash is also using more tokens than 3.7 did, the two add up. Price a month of your real traffic on both rates before the year ends; our cost calculator takes your token counts, and the 3.8 Flash price page has every tier.

If you are still on Gemini 2.5

The 2.5 models are the opposite case: no reroute and no date, just a closed door. If your project already calls gemini-2.5-flash, it keeps working. A user who has not called them before may be refused, so a new project or a fresh staging setup can fail where production works.

Google says to start new work on 3.5 Flash-Lite or 3.8 Flash. 3.5 Flash-Lite is the cheaper step from 2.5 Flash at $0.30 in and $2.50 out, the same as 2.5 Flash's price, and Google reports it scores higher on reasoning and chart benchmarks. 2.5 Flash-Lite, at $0.10 in and $0.40 out, has no like-for-like price in the 3 series; the nearest is 3.1 Flash-Lite at $0.25 in and $1.50 out, which itself shuts down from May 2027.

Moving off, in order

1. Find every model string. Search code, config, environment variables and saved prompts for gemini-3.5-flash, gemini-3.7-flash and gemini-2. names. Strings in a database or a feature flag are the ones that get missed.
2. Pin the model you mean. Change gemini-3.7-flash to gemini-3.8-flash and gemini-3.5-flash to gemini-3.6-flash (or 3.8), so your code says what actually answers.
3. Fix the parameters the 3 series dropped. Google's checklist: replace thinking_budget with thinking_level, and note 3.8 Flash has no minimal level; remove temperature, top_p and top_k, which are now ignored and will raise errors in later models; remove candidate_count and prefilled model turns.
4. Run your evals on the new model. Compare answers and token counts on the same inputs, not just pass rates. Our model retirement checklist covers what to measure, and the eval sample size calculator tells you whether a score change is real.
5. Watch the bill in the first week and again in January. Set a budget alert at your expected spend on the 2027 rate.

The pattern is not new to Google. Every maker retires models on its own schedule; what model deprecation means and how Anthropic and OpenAI handle it is in our concept page, and the Claude Sonnet 4.5 retirement is a worked example from the other side. If the move is a chance to rethink the model altogether, Gemini 4 Argon against Claude Opus 5.5 compares the two, and /prices has every current Gemini price.