Gemini-flash-latest on Vertex AI: Cannot determine actual model version + gemini-3-flash-preview quality regression

Gemini API
1

Context

We are running a production conversational AI agent on Vertex AI (location=global) using Google ADK (Agents Development Kit). Our agent uses tool calling (function calling) heavily — it orchestrates vehicle search, result presentation, and follow-up suggestions through multiple sub-agents.

We have been using gemini-flash-latest as our default model and are very happy with its quality — especially its tool-calling reliability, response formatting, and natural language output style.

Problem 1: Cannot resolve what gemini-flash-latest points to

There is no API to determine the actual model version behind the gemini-flash-latest alias on Vertex AI.

We tried:

  • models.get("gemini-flash-latest") returns name: publishers/google/models/gemini-flash-latest, version: default. No resolved model info.
  • Response metadata model_version field just echoes gemini-flash-latest back.
  • REST API GET /v1/publishers/google/models/gemini-flash-latest returns 404 on global endpoint.

We then ran empirical side-by-side tests (temperature=0, no thinking, identical prompts) comparing gemini-flash-latest against known model versions:

Test (7 deterministic prompts) Exact hash match with flash-latest
gemini-2.5-flash 5/7
gemini-3-flash-preview 4/7

Conclusion: gemini-flash-latest on Vertex AI (location=global) is not identical to either gemini-2.5-flash or gemini-3-flash-preview. It appears to be a distinct, unpinnable snapshot.

According to the changelog (Jan 21, 2026), gemini-flash-latest was switched to gemini-3-flash-preview on the Google AI API — but this does not appear to match Vertex AI behavior.

Problem 2: gemini-3-flash-preview quality regression

We ran the same agent pipeline (ADK, multi-agent with tool calling) across all three models:

Metric gemini-flash-latest gemini-2.5-flash gemini-3-flash-preview
Tool calling Correct Correct Correct
Output tokens 362 292 283
Response quality Natural, well-formatted Good, slightly different style Less polished
Bold formatting Appropriate emphasis Phrase-level bold Minimal formatting

In our broader testing, gemini-3-flash-preview shows less natural phrasing and more terse output that does not match our UX standards. We are not confident upgrading to it.

Requests

  1. Provide an API to resolve model aliases — models.get() or response metadata should return the actual underlying model version, not just echo the alias.
  2. Clarify Vertex AI vs Google AI alias mapping — The changelog documents alias switches for the Google AI API. Does the same schedule apply to Vertex AI?
  3. Provide a way to pin the current gemini-flash-latest snapshot — If the alias is about to change, we need the exact snapshot ID so we can pin before it rotates.

Related GitHub issue: `models.get()` does not resolve alias to underlying model version on Vertex AI · Issue #2271 · googleapis/python-genai · GitHub

Environment

  • Platform: Vertex AI
  • Location: global
  • SDK: google-genai (Python)
  • Framework: Google ADK (Agents Development Kit)
1 Like