Google’s Gemini API changelog for October 8, 2026 deprecates two Flash models without shutting them down: “API requests to gemini-3.7-flash are automatically routed to gemini-3.8-flash,” and requests to gemini-3.5-flash now go to gemini-3.6-flash. The same entry schedules the deep-research-pro-preview-12-2025 agent for shutdown no earlier than October 23. Gemini API changelog The Flash change is the unusual part. A retired model ID normally starts returning errors, which forces someone to notice. These IDs keep answering, from a different model, so output quality, latency and token counts can shift in an application that nobody has redeployed.
This follows our October 7 piece on the Veo and TTS preview shutdowns, which ran before Google announced these changes.
What changed on October 8
The deprecations page lists both old Flash models with a shutdown date of “N/A” and a note that the endpoint “is not shutting down”. Neither model can be selected in Google AI Studio any more. Gemini API deprecations Gemini 3.7 Flash model page
| Requested ID | Released | Now served by | Shutdown date (Gemini API) |
|---|---|---|---|
gemini-3.7-flash | Aug 13, 2026 | gemini-3.8-flash | N/A, endpoint stays up |
gemini-3.5-flash | May 19, 2026 | gemini-3.6-flash | N/A, endpoint stays up |
deep-research-pro-preview-12-2025 | Dec 11, 2025 | not routed | No earlier than Oct 23, 2026 |
Source: Gemini API deprecations, last updated October 12. Neither 3.6 Flash nor 3.8 Flash has an announced shutdown date.
The 3.7 Flash routing sits awkwardly next to Google’s own launch post for 3.8 Flash on September 2, which said 3.7 Flash “remains fully supported for efficiency-first workloads”. Google’s 3.8 Flash announcement Five weeks later, a team that stayed on 3.7 Flash for that reason is on 3.8 Flash anyway.
Per-token prices, and why they are not the whole bill
Google’s pricing page no longer lists 3.7 Flash or 3.5 Flash, so the old rates come from launch material. Gemini API pricing
| Per 1M tokens, standard paid tier | Input | Output |
|---|---|---|
| 3.7 Flash (introductory, through Dec 31, 2026) | $0.75 | $3.75 |
| 3.8 Flash, through Dec 31, 2026 | $0.75 | $3.75 |
| 3.8 Flash, from Jan 1, 2027 | $1.50 | $7.50 |
| 3.5 Flash (per Google’s 3.6 launch notes) | $1.50 | $9.00 |
| 3.6 Flash, through Dec 31, 2026 | $0.75 | $3.75 |
| 3.6 Flash, from Jan 1, 2027 | $1.50 | $7.50 |
The 3.7 and 3.8 rates are identical, per Google’s 3.8 launch post. Google’s 3.8 Flash announcement The 3.5 Flash figures are implied by the 3.6 notes, which give 3.6’s output price as “$7.50/1M versus $9.00/1M for 3.5 Flash” and say input stays at $1.50. What’s new in Gemini 3.6 Those notes, and the 3.6 launch post, quote $1.50 and $7.50 for 3.6 Flash; the current pricing page shows $0.75 and $3.75 until the end of the year. We treat the pricing page as current. On paper, then, 3.5 Flash traffic got cheaper per token and 3.7 Flash traffic stayed level, and both routed-to models step up on January 1, though 3.6 Flash’s January rates still sit at or below the 3.5 Flash figures above.
Matching per-token rates do not guarantee a matching bill. Google’s 3.8 announcement warns that “at times, the model might use more tokens to maximize performance, especially at higher effort levels.” Google’s 3.8 Flash announcement If 3.8 Flash spends more output tokens per request than 3.7 Flash did, the bill rises with no change in rate. Our 3.8 Flash cost piece walks through measuring cost per task rather than per token.
What to do with code still calling the old IDs
The decision is whether to keep relying on Google’s routing or to move the ID yourself. Moving it is the better choice in most cases, because the code then names the model that actually answers. Logs, cost dashboards and eval reports that still say “3.7 Flash” describe a model that is no longer serving the traffic.
A short checklist for the switch:
- Search code, config and environment variables for
gemini-3.7-flashandgemini-3.5-flash, including prompt-management tools and agent frameworks that hold their own model defaults. - Rerun your evals against
gemini-3.8-flashorgemini-3.6-flashdirectly. Any baseline recorded before October 8 under an old ID measured the old model. - Compare token usage per request, especially output tokens, against the weeks before October 8. That shows whether the routed model is spending more on the same work.
- Watch latency and structured-output parsing. A different model can format the same answer differently.
- Budget for the January 1 price step on both 3.6 and 3.8 Flash.
For 3.5 Flash users, the routing target is 3.6 Flash, which the pricing page lists as Google’s previous-generation Flash, not the newest one. Gemini API pricing If you are retesting anyway, run 3.8 Flash in the same pass, since the two currently share a price.
Deep Research Pro preview: a hard date
The old Deep Research agent gets no routing. It is scheduled to shut down no earlier than October 23, 2026 (still listed when we checked on October 14), and the deprecations page names deep-research-preview-04-2026 as its replacement. Gemini API deprecations Treat October 23 as the earliest date the agent could stop answering, and check the page again before relying on it.
Migrating means changing the agent parameter in interactions.create to one of two new IDs, which the changelog describes this way: Gemini API changelog
| Agent ID | Google’s description |
|---|---|
deep-research-preview-04-2026 | “Designed for speed and efficiency, ideal to be streamed back to a client UI.” |
deep-research-max-preview-04-2026 | “Maximum comprehensiveness for automated context gathering and synthesis.” |
The choice follows from where the output goes. A user waiting on screen points to the standard agent; a scheduled job that writes a report nobody reads until later can afford the Max agent. Both are still previews, and both run only through the Interactions API, not generate_content. The agents require background=True; streaming also needs stream=True, and without thinking_summaries set to "auto" in agent_config, the docs say “the stream may only provide the final results.” Deep Research docs A UI that showed progress from the old agent may go quiet until that setting is added.
What we could not confirm
Google’s pages do not say whether the Flash routing covers every path: Batch, Flex, Priority, cached content and tuned models are not addressed on the changelog, deprecations or model pages. We also did not find a primary source saying whether responses identify the model that actually served them, which would decide whether your logs can catch the switch on their own. This article covers the Gemini Developer API only; Google Cloud’s agent platform publishes its own lifecycle dates. And the Deep Research documentation we read did not show whether the new agents’ pricing or research behavior differs from the Pro preview’s, so run a few representative research tasks on each before October 23 instead of assuming the output will match.