Gemini-3-flash-preview is incorrectly outputting HTML tokens

Gemini API
, ,
1

Model: gemini-3-flash-preview

Environment: Production (via API)

Issue: Sporadic insertion of HTML tags at the end of assistant turns.

Description: We are developing an AI-powered voice agent using the Gemini 3 Flash Preview model. While the model generally follows our plain-text system instructions, it sporadically appends an HTML tag to the end of its responses.

The Impact: Because this output is fed directly into a Text-to-Speech (TTS) engine, the TTS attempts to phonetically pronounce the string, resulting in a distinct and unprofessional “g” or “glottal stop” sound at the end of the agent’s speech.

Observations:

  • Environment Discrepancy: The issue occurs in our Production environment but is much less common in our Local testing environment (could potentially be to high concurrent volume in production).

  • Sporadic Nature: It does not happen on every turn; it seems to occur more frequently in longer conversations or when the model is reviewing structured lists (e.g., medication schedules).

  • Example: Assistant: I see, you had some tests done but haven’t received the results yet; do you remember which facility did those tests so we can help track them down for you?

Configurations Used:

  • model: gemini-3-flash-preview

  • temperature: 1.0

Questions for the Community:

  1. Is this a known “Formatting Leakage” regression in the Gemini 3 preview models?

  2. Why would this manifest in Production but stay relatively clean in Local testing (possible server-side versioning or default parameter differences)?

  3. Besides post-processing Regex, are there specific safety_settings or stop_sequences that have successfully suppressed these web-markup hallucinations for other developers?

3

Hi @mith, welcome to the community!

To understand the issue better, could you please share the logs and a snippet of code so that we can analyze the issue?

Thank you!